AMD 64 BIOS and Kenrel Dev's Guide

BIOS and Kernel Developer's Guide for AMD NPT Family 0Fh Processors
Publication # 32559 Issue Date: July 2007
Revision: 3.08
2005-2007 Advanced Micro Devices, Inc. All rights reserved. The contents of this document are provided in connection with Advanced Micro Devices, Inc. (AMD) products. AMD makes no representations or warranties with respect to the accuracy or completeness of the contents of this publication and reserves the right to make changes to specifications and product descriptions at any time without notice. The information contained herein may be of a preliminary or advance nature and is subject to change without notice. No license, whether express, implied, arising by estoppel or otherwise, to any intellectual property rights is granted by this publication. Except as set forth in AMDs Standard Terms and Conditions of Sale, AMD assumes no liability whatsoever, and disclaims any express or implied warranty, relating to its products including, but not limited to, the implied warranty of merchantability, fitness for a particular purpose, or infringement of any intellectual property right. AMDs products are not designed, intended, authorized or warranted for use as components in systems intended for surgical implant into the body, or in other applications intended to support or sustain life, or in any other application in which the failure of AMDs product could create a situation where personal injury, death, or severe property or environmental damage may occur. AMD reserves the right to discontinue or make changes to its products at any time without notice.
Trademarks AMD, the AMD Arrow logo, AMD Athlon, AMD Opteron, AMD Sempron, AMD Turion, and combinations thereof, 3DNow!, AMD PowerNow!, Cool'n'Quiet, AMD Virtualization, and AMD-V are trademarks of Advanced Micro Devices, Inc. HyperTransport is a licensed trademark of the HyperTransport Technology Consortium. MMX and Pentium are registered trademarks of Intel Corporation. Microsoft and Windows are registered trademarks of Microsoft Corporation. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies.
32559
Rev. 3.08
July 2007
BIOS and Kernel Developers Guide for AMD NPT Family 0Fh Processors
Contents
Revision History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .17 1 1.1 1.2 1.3 1.3.1 1.4 2 2.1 2.1.1 2.1.2 2.1.3 2.1.4 2.1.5 2.2 2.3 2.3.1 2.3.2 3 3.1 3.2 3.3 3.4 3.5 3.5.1 3.5.2 3.6 Introduction and Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19 About This Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19 Related Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19 Conventions and Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .20 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .20 Register Differences in Revisions of AMD NPT Family 0Fh Processors . . . . . . . . . . . .21 Processor Initialization and Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .25 Bootstrap Processor Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .25 Detecting AP and Initializing Routing Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .25 Initializing Noncoherent HyperTransport Technology Devices . . . . . . . . . . . . . .26 Initializing Link Width and Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .26 Initializing the Memory Controller on All Processor Nodes . . . . . . . . . . . . . . . . . .27 Initializing the Address Map Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .27 Application Processor Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .27 BIOS Requirement for 64-Bit Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .28 Sizing and Testing Memory above 4 Gbytes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .28 Using BIOS Callbacks in 64-Bit Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .28 Functional Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .29 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .29 Instruction Set Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30 Supported Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30 Address Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30 Internal Cache Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30 Level 1 Caches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30 Level 2 Cache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .30 Northbridge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .31 Contents 3
32559
Rev. 3.08 July 2007
3.6.1 3.6.2 3.6.3 3.6.4 3.6.5 3.6.6 3.7 3.7.1 3.7.2 3.7.3 4 4.1 4.1.1 4.1.2 4.2 4.3 4.3.1 4.3.2 4.3.3 4.3.4 4.3.5 4.3.6 4.3.7 4.3.8 4.3.9 4.3.10 4.3.11 4.3.12 4.3.13 4.3.14 4
HyperTransport Technology Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .31 DRAM Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .32 Main Memory Hardware Scrubbing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .34 Memory Interleaving Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .35 Memory Hoisting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .35 On-Line Spare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .37 Power Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .38 PWROK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .39 Thermal Sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .39 Thermal Diode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .40 Memory System Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .43 Configuration Space Accesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .43 Configuration Address Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .43 Configuration Data Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .44 Memory System Configuration Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45 Function 0HyperTransport Technology Configuration . . . . . . . . . . . . . . . . . . . . . .45 Device/Vendor ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .47 Status/Command Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .48 Class Code/Revision ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .48 Header Type Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .49 Capabilities Pointer Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .49 Routing Table Node i Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .50 Node ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .52 Unit ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .53 HyperTransport Transaction Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . .54 HyperTransport Initialization Control Register . . . . . . . . . . . . . . . . . . . . . . . . . .59 LDTi Capabilities Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .60 LDTi Link Control Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .62 LDTi Frequency/Revision Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .66 LDTi Feature Capability Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .67 Contents
32559
Rev. 3.08
July 2007
4.3.15 4.3.16 4.3.17 4.4 4.4.1 4.4.2 4.4.3 4.4.4 4.4.5 4.4.6 4.4.7 4.4.8 4.5 4.5.1 4.5.2 4.5.3 4.5.4 4.5.5 4.5.6 4.5.7 4.5.8 4.5.9 4.5.10 4.5.11 4.5.12 4.5.13 4.5.14 4.5.15 4.5.16 4.5.17
LDTi Buffer Count Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .68 LDTi Bus Number Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .70 LDTi Type Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .70 Function 1Address Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .72 Device/Vendor ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .74 Class Code/Revision ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .74 Header Type Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .75 DRAM Address Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .76 Memory-Mapped I/O Address Map Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . .79 PCI I/O Address Map Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .82 Configuration Map Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .84 DRAM Hole Address Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .85 Function 2DRAM Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .87 Device/Vendor ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .88 Class Code/Revision ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .89 Header Type Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .89 DRAM CS Base Address Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .90 DRAM CS Mask Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .95 DRAM Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .95 DRAM Initialization Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .96 DRAM Bank Address Mapping Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .98 Address to Node/Chip Select Translation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .102 DRAM Timing Low Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .104 DRAM Timing High Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .106 DRAM Configuration Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .109 DRAM Controller Additional Data Offset Register . . . . . . . . . . . . . . . . . . . . . . . .114 DRAM Controller Additional Data Port Register . . . . . . . . . . . . . . . . . . . . . . . . . .116 Output Driver Compensation Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . .116 Write Data Timing Control Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .119 Address Timing Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .121 Contents 5
32559
Rev. 3.08 July 2007
4.5.18 4.5.19 4.5.20 4.6 4.6.1 4.6.2 4.6.3 4.6.4 4.6.5 4.6.6 4.6.7 4.6.8 4.6.9 4.6.10 4.6.11 4.6.12 4.6.13 4.6.14 4.6.15 4.6.16 4.6.17 4.6.18 4.6.19 4.6.20 4.6.21 4.6.22 4.6.23 4.6.24 4.6.25 5 6
Read DQS Timing Control Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .123 DQS Receiver Enable Timing Control Registers . . . . . . . . . . . . . . . . . . . . . . . . . .125 DRAM Controller Miscellaneous Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .126 Function 3Miscellaneous Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .128 Device/Vendor ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .130 Class Code/Revision ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .131 Header Type Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .131 Machine Check Architecture (MCA) Registers . . . . . . . . . . . . . . . . . . . . . . . . . . .132 ECC and Chip Kill Error Checking and Correction . . . . . . . . . . . . . . . . . . . . . . . .148 Scrub Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .151 DRAM Scrub Address Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .152 HTC Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .154 Thermal Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .155 XBAR Flow Control Buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .156 XBAR-to-SRI Buffer Count Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .163 Display Refresh Flow Control Buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .165 Power Management Control Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .165 GART Aperture Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .168 GART Aperture Base Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .169 GART Table Base Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .170 GART Cache Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .171 Power Control Miscellaneous Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .171 On-Line Spare Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .172 Clock Power/Timing Low Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .174 Clock Power/Timing High Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .176 HyperTransport FIFO Read Pointer Optimization Register . . . . . . . . . . . . . . . .177 Thermtrip Status Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .179 Northbridge Capabilities Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .182 CPUID Family Model Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .183 DRAM Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .185 Contents
32559
Rev. 3.08
July 2007
5.1 5.1.1 5.1.2 5.1.3 5.1.4 5.1.5 5.2 6 6.1 6.2 6.2.1 6.3 6.3.1 6.3.2 6.4 6.4.1 6.4.2 6.4.3 6.4.4 6.4.5 6.5 6.6 6.6.1 7 7.1 7.2 7.3 7.3.1 7.4 7.5
Programming Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .185 SPD ROM-Based Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .185 Non-SPD ROM-Based Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .189 Mismatched DIMM Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .194 DRAM Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .194 DRAM Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .195 DDR2 DRAM Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .198 Machine Check Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .203 Determining Machine Check Support . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .203 Machine Check Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .203 Sources of Machine Check Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .204 Machine Check Architecture Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .206 Global Machine Check Model-Specific Registers (MSRs) . . . . . . . . . . . . . . . . . .207 Error Reporting Bank Machine Check MSRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . .209 Error Reporting Banks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .213 Data Cache (DC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .214 Instruction Cache (IC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .219 Bus Unit (BU) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .222 Load Store Unit (LS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .227 Northbridge (NB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .228 Initializing the Machine Check Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .231 Using Machine Check Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .231 Handling Machine Check Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .232 Advanced Programmable Interrupt Controller (APIC) . . . . . . . . . . . . . . . . . . . . . . . . .235 Interrupt Delivery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .236 Vectored Interrupt Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .236 Spurious Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .237 Spurious Interrupts Caused by Timer Tick Interrupt . . . . . . . . . . . . . . . . . . . . . . .237 Lowest-Priority Arbitration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .238 Inter-Processor Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .239 Contents 7
32559
Rev. 3.08 July 2007
7.6 7.7 7.8 7.8.1 7.8.2 7.8.3 7.8.4 7.8.5 7.8.6 7.8.7 7.8.8 7.8.9 7.8.10 7.8.11 7.8.12 7.8.13 7.8.14 7.8.15 7.8.16 7.8.17 7.8.18 7.8.19 7.8.20 7.8.21 7.8.22 7.8.23 7.8.24 7.8.25 7.8.26 7.8.27 8
APIC Timer Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .239 State at Reset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .239 Register Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .240 APIC ID Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .241 APIC Version Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .241 Task Priority Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .242 Arbitration Priority Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .242 Processor Priority Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .243 End of Interrupt Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .243 Remote Read Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .243 Logical Destination Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .244 Destination Format Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .244 Spurious Interrupt Vector Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .245 In-Service Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .245 Trigger Mode Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .246 Interrupt Request Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .247 Error Status Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .248 Interrupt Command Register Low . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .249 Interrupt Command Register High . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .251 Timer Local Vector Table Entry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .251 Thermal Local Vector Table Entry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .252 Performance Counter Local Vector Table Entry . . . . . . . . . . . . . . . . . . . . . . . . . .253 Local Interrupt 0 (Legacy INTR) Local Vector Table Entry Register . . . . . . . . . .253 Local Interrupt 1 (Legacy NMI) Local Vector Table Entry . . . . . . . . . . . . . . . . . .254 Error Local Vector Table Entry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .255 Timer Initial Count Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .255 Timer Current Count Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .256 Timer Divide Configuration Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .256 Extended APIC Feature Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .257 Threshold Count Interrupt 0 Local Vector Table Entry . . . . . . . . . . . . . . . . . . . . .257 Contents
32559
Rev. 3.08
July 2007
8 8.1 8.2 8.3 8.4 8.5 8.6 8.7 8.8 8.8.1 8.8.2 8.9 8.10 8.11 8.11.1 8.11.2 8.11.3 8.11.4 8.11.5 8.11.6 8.12 8.13 9 9.1 9.2 9.3 9.4 9.5 9.5.1 10
System Management Mode (SMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .259 SMM Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .259 Operating Mode and Default Register Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .259 SMM State Save Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .261 SMM Initial State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .264 SMM-Revision Identifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .265 SMM Base Address . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .266 Auto Halt Restart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .266 SMM I/O Trap and I/O Restart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .267 SMM I/O Trap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .268 SMM I/O Restart Byte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .268 Exceptions and Interrupts in SMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .269 NMI Mask . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .269 Protected SMM and ASeg/TSeg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .270 SMM_MASK Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .270 SMM_ADDR Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .271 SMM ASeg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .272 SMM TSeg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .272 Closing SMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .273 Locking SMM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .273 SMM Special Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .274 MP-SMM Spring Boarding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .274 HyperTransport Technology Configuration and Enumeration . . . . . . . . . . . . . . . . .277 Initial Configuration Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .277 One-Node Coherent HyperTransport Technology Initialization . . . . . . . . . . . . . . . .278 Two-Node Coherent HyperTransport Technology Initialization . . . . . . . . . . . . . . . .278 Generic HyperTransport Technology Configuration . . . . . . . . . . . . . . . . . . . . . . . . .278 Rules for Valid Routing Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .279 2-Hop Cycle Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .279 Power and Thermal Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .281 Contents 9
32559
Rev. 3.08 July 2007
10.1 10.2 10.2.1 10.2.2 10.2.3 10.2.4 10.3 10.3.1 10.4 10.5 10.5.1 10.5.2 10.5.3 10.5.4 10.5.5 10.5.6 10.5.7 10.6 10.6.1 10.6.2 10.6.3 10.6.4 10.6.5 10.7 10.8 10.9 10.9.1 10.9.2 11 10
Stop Grant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .282 C-States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .284 C1 Halt State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .284 C2 and C3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .284 C3 and AltVID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .285 C1 Enhanced (C1E) Halt State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .285 Throttling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .286 BIOS Requirements for PROCHOT_L Throttling . . . . . . . . . . . . . . . . . . . . . . . . .286 Processor ACPI Thermal Zone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .286 Processor Performance States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .287 BIOS Requirements for P-State Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .287 BIOS Requirements for P-State Transitions in a Multiprocessor System . . . . . . .289 BIOS Requirements for P-State Transitions in Systems Using Dual Core Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .290 BIOS-Initiated P-State Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .290 BIOS Support for Operating System/CPU Driver-Initiated P-State Transitions . .290 Processor Driver Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .291 P-State Transition Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .291 ACPI 2.0 Processor P-State Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .303 _PCT (Performance Control) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .304 _PSS (Performance-Supported States) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .304 _PPC (Performance Present Capabilities) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .310 PSTATE_CNT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .312 CST_CNT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .312 ACPI 3.0 Processor P-State Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .312 BIOS Support for AMD PowerNow! Software with Legacy Operating Systems . . .313 System Configuration for Power Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .315 Chipset Configuration for Power Management . . . . . . . . . . . . . . . . . . . . . . . . . . .315 Processor Configuration for Power Management . . . . . . . . . . . . . . . . . . . . . . . . . .315 Performance Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .321 Contents
32559
Rev. 3.08
July 2007
11.1 11.2 11.2.1 12 13 13.1 13.2 13.3 13.4 13.5 13.6 13.7 13.8 13.9 13.9.1 13.10 13.10.1 13.11 13.12 13.13 13.13.1 13.13.2 13.14 13.14.1 13.14.2 13.15 13.15.1 13.15.2 Chapter
Performance Counters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .321 Performance Event-Select Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .322 Performance Monitor Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .324 CPUID Function Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .345 BIOS Checklist . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .353 CPUID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .353 CPU Speed Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .353 HyperTransport Link Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .353 HyperTransport Link Frequency Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .354 Multiprocessing Capability Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .354 Setup for Dual Core Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .355 APIC ID Assignment Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .356 Processors With Different Revisions or Maximum Frequencies in Multiprocessor Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .357 Model-Specific Registers (MSRs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .358 Software Considerations for Accessing Northbridge MSRs in Dual Core Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .358 Machine Check Architecture (MCA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .359 GART Table Walk Error Reporting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .359 GART Register Restoration After S3 Resume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .360 Register Settings in UMA Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .360 Memory Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .361 I/O and Memory Type and Range Registers (IORRs, MTRRs) . . . . . . . . . . . . . . .361 Memory Map Registers (MMRs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .361 Access Routing And Type Determination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .361 Memory Access to the Physical Address Space . . . . . . . . . . . . . . . . . . . . . . . . . . .361 Northbridge Processing of Accesses To Physical Address Space . . . . . . . . . . . . .363 Error Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .363 Error Handling Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .363 Errors and Recommended Error Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .364 Contents 11
32559
Rev. 3.08 July 2007
13.16 13.17 13.18 13.19 13.20 13.21 13.22 14 14.1 14.1.1 14.1.2 14.1.3 14.1.4 14.1.5 14.2 14.2.1 14.2.2 14.2.3 14.2.4 14.2.5 14.2.6 14.2.7 14.2.8 14.2.9 14.2.10 14.2.11 14.2.12 14.2.13
Cache Initialization For General Storage During Boot . . . . . . . . . . . . . . . . . . . . . . . . .366 Cache Testing and Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .367 Memory System Configuration Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .367 Processor ID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .368 XSDT Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .368 Detect Target Operating Mode Callback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .368 SMM Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .369 Processor Configuration Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .371 General Model-Specific Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .371 System Software Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .373 Memory Typing Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .375 APIC Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .392 Software Debug Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .393 Performance Monitoring Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .394 AMD NPT Family 0Fh Processor Model-Specific Registers . . . . . . . . . . . . . . . . . . . . .394 Feature Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .396 Identification Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .403 HTC Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .403 Thermal Control Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .404 Memory Typing Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .405 I/O Range Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .406 System Call Extension Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .407 Segmentation Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .409 Power Management Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .411 I/O and Configuration Space Trapping to SMI . . . . . . . . . . . . . . . . . . . . . . . . . . . .416 Interrupt Pending Message Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .419 VM_CR Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .421 VM_HSAVE_PA Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .422 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .423 Index of Register Names . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .429
12
Contents
Chapter
32559
Rev. 3.08
July 2007
List of Figures
Figure 1. Figure 2. Figure 3. Figure 4. Figure 5. Figure 6. Figure 7. Figure 8. Figure 9. Figure 10. Figure 11. Figure 12. Figure 13. Figure 14. Figure 15. Processor Block Diagram .........................................................................................29 TCASE Max and TCONTROL Max Relationship...................................................42 Interleave Example (IntlvEn Relation to IntlvSel) ...................................................79 8P Ladder Configuration ........................................................................................159 8P Twisted Ladder Configuration...........................................................................160 Default SMM Memory Map ...................................................................................261 A Four-Node Configuration....................................................................................280 High-Level P-state Transition Flow .......................................................................295 Example P-State Transition Timing Diagram ........................................................296 Phase 1: Core Voltage Transition Flow ..................................................................297 TargetVID Calculation ...........................................................................................298 Phase 2: Core Frequency Transition Flow..............................................................299 Transition to a Higher Core Frequency ..................................................................300 Transition to a Lower Core Frequency ...................................................................301 Phase 3: Core Voltage Transition Flow ..................................................................302
List of Figures
13
32559
Rev. 3.08 July 2007
14
List of Figures
32559
Rev. 3.08
July 2007
List of Tables
Table 1. Table 2. Table 3. Table 4. Table 5. Table 6. Table 7. Table 8. Table 9. Table 10: Table 11: Table 12. Table 13. Table 14. Table 15. Table 16. Table 17. Table 18. Table 19. Table 20. Table 21. Table 22. Table 23. Table 24. Table 25. Table 26. Table 27. Table 28. Table 29. Table 30. Table 31. Table 32. Table 33. Table 34. Table 35. Table 36. Table 37. Table 38. Table 39. Related Documents .........................................................................................................20 Packages..........................................................................................................................30 DIMM Support Per Package ...........................................................................................32 DRAM Interface Speed vs. CPU Core Clock Multiplier................................................33 Function 0 Configuration Registers ................................................................................45 Function 1 Configuration Registers ................................................................................72 Function 2 Configuration Registers ................................................................................87 DRAM CS Base Address and DRAM CS Mask Registers.............................................91 DRAM CS Base Address and DRAM CS Mask Registers Mismatched DIMM Support............................................................................................................................92 DRAM address mapping 64-bit interface. ......................................................................99 DRAM address mapping 128-bit interface. ....................................................................99 Swapped physical address lines for interleaving with 64-bit interface.........................100 Swapped physical address lines for interleaving with 128-bit interface.......................101 DctOffset Field Encodings............................................................................................115 Function 3 Configuration Registers ..............................................................................129 Error Code Field Formats..............................................................................................140 Transaction Type Bits (TT)...........................................................................................140 Cache Level Bits (LL)...................................................................................................140 Memory Transaction Type Bits (RRRR) ......................................................................140 Participation Processor Bits (PP) ..................................................................................141 Time-Out Bit (T) ...........................................................................................................141 Memory or I/O Bits (II).................................................................................................141 Northbridge Error Codes...............................................................................................141 Northbridge Error Status Bit Settings ...........................................................................144 ECC Syndromes ............................................................................................................149 Chip Kill ECC Syndromes ............................................................................................150 Scrub Rate Control Values............................................................................................152 XBAR Input Buffers .....................................................................................................156 Default XBAR Command Buffer Allocation................................................................157 Default Virtual Channel Command Buffer Allocation .................................................158 An Example of a Non Default Virtual Channel Command Buffer Allocation .............158 2P Dual Coherent Link XBAR Command Buffer Allocation ......................................159 8P Outer Node Virtual Channel XBAR Command Buffer Allocation .........................160 8P Inner Node Virtual Channel XBAR Command Buffer Allocation..........................160 Recommended Flow Control Buffer Allocations in UMA systems .............................161 GART PTE Organization..............................................................................................170 RDIMM and Unbuffered DIMM ODT Settings ...........................................................188 SO-DIMM ODT Settings..............................................................................................189 Four Bank Activate Window Values ............................................................................190 List of Tables 15
32559
Rev. 3.08 July 2007
Table 40. Table 41. Table 42. Table 43. Table 44. Table 45. Table 46. Table 47. Table 48. Table 49. Table 50. Table 51. Table 52. Table 53. Table 54. Table 55. Table 56. Table 57. Table 58. Table 59. Table 60. Table 61. Table 62. Table 63. Table 64. Table 65. Table 66. Table 67. Table 68. Table 69. Table 70. Table 71. Table 72. Table 73. Table 74. Table 75. Table 76. Table 77. Table 78. Table 79. Table 80.
RdPadRcvFifoDly Values.............................................................................................190 Unbuffered DIMM Address Timings and Drive Strengths for AM2 Package .............191 SO-DIMM Address Timings and Drive Strengths for S1g1 Package ..........................192 Registered DIMM Address Timings and Drive Strengths for F(1207) Package 4-DIMM System ...........................................................................................................192 Registered DIMM Address Timings and Drive Strengths for F(1207) Package 8-DIMM System ...........................................................................................................193 Sources of Machine Check Errors.................................................................................205 DC Error Codes.............................................................................................................216 DC Error Status Bit Settings .........................................................................................218 Valid MC0_ADDR Bits................................................................................................218 IC Error Codes ..............................................................................................................221 IC Error Status Bit Settings...........................................................................................221 Valid MC1_ADDR Bits................................................................................................222 BU Error Codes.............................................................................................................225 BU Error Status Bit Settings .........................................................................................226 Valid MC2_ADDR Bits................................................................................................226 LS Error Codes..............................................................................................................228 LS Error Status Bit Settings ..........................................................................................228 APIC Register Summary...............................................................................................240 Valid Combinations of ICR Fields................................................................................251 SMM Save State (Offset FE00FFFFh) .......................................................................261 SMM Entry State...........................................................................................................265 SMM ASeg-Enabled Memory Types............................................................................272 SMM TSeg-Enabled Memory Types ............................................................................273 Power Management Categories.....................................................................................281 ACPI-State Support by System Class ...........................................................................282 Required SMAF Code to Stop Grant Mapping.............................................................283 Low FID Frequency Table (< 1600 MHz)....................................................................292 High FID Frequency Table (>= 1600 MHz).................................................................292 Sample VST Values ......................................................................................................307 MVS Values ..................................................................................................................308 RVO Values ..................................................................................................................308 VID Code Voltages .......................................................................................................309 IRT Values ....................................................................................................................309 _PSS Status Field ..........................................................................................................310 Performance State Block Structure ...............................................................................313 Configuration Register Settings for Power Management .............................................316 FADT Table Entries......................................................................................................319 Northbridge MSRs ........................................................................................................358 General MSRs ...............................................................................................................371 AMD NPT Family 0Fh Processor MSRs......................................................................395 FID Code Translations ..................................................................................................412
16
List of Tables
32559
Rev. 3.08
July 2007
Revision History
Date July 2007 Rev. 3.08 Description Updated Table 4 DRAM Interface Speed vs. CPU Core Clock Multiplier. Updated 3.7.3 Thermal Diode. Added 3.7.3.1 TCASE max and TCONTROL max for Socket F (1207) and Socket AM2 Processors. Added 3.7.3.2 TDIE max and TCONTROL for Socket S1g1 Processors. Updated 4.5.11 DRAM Timing High Register. Updated 4.5.17 Address Timing Control Register. Corrected 4.6.4.2 MCA NB Configuration Register. Updated 4.6.4.7 Watchdog Timer Errors. Updated 4.6.10 XBAR Flow Control Buffers. Clarified 5.1.1.19 DIMM ODT. Updated Table 37. Updated Table 38 SO-DIMM ODT Settings. Updated Table 39 Four Bank Activate Window Values. Updated Table 42 SO-DIMM Address Timings and Drive Strengths for S1g1 Package. Corrected Table 44 Registered DIMM Address Timings and Drive Strengths for F(1207) Package 8-DIMM System. Corrected 5.1.5.1 DQS Receiver Enable Training. Corrected Table 64 ACPI-State Support by System Class. Updated Table 66 Low FID Frequency Table (< 1600 MHz). Clarified 10.3 Throttling. Clarified 10.5.7.2.2, 10.6.2.1.3 and 10.8 for PLL lock time. Updated Table 75 Configuration Register Settings for Power Management. Updated 13.16 Cache Initialization For General Storage During Boot. Corrected 14.2.1.3 HWCR Register. Updated 14.2.9.1 FIDVID_CTL Register. Clarified 14.2.11 Interrupt Pending Message Register. Added 14.2.12 VM_CR Register. Added 14.2.13 VM_HSAVE_PA Register.
Revision History
17
32559
Rev. 3.08 July 2007
Date
Rev.
Description Corrected 4.5.20 DRAM Controller Miscellaneous Register. Clarified 4.6.22 HyperTransport FIFO Read Pointer Optimization Register. Updated 4.6.23 Thermtrip Status Register. Updated Table 42 SO-DIMM Address Timings and Drive Strengths for S1g1 Package. Updated Table 66 Low FID Frequency Table (< 1600 MHz) for 100 MHz frequency steps. Updated Table 67 High FID Frequency Table (>= 1600 MHz) for 100 MHz frequency steps. Updated Figure 10. Added Figure 11. Updated Figure 12. Added Figure 13. Added Figure 14. Added 10.7 ACPI 3.0 Processor P-State Objects. Updated Table 75 Configuration Register Settings for Power Management. Updated 12 CPUID Function Registers Fn8000_0001_ECX, Fn8000_0007_EDX, and Fn8000_000A_EDX. Updated Table 80 FID Code Translations with odd FID codes.
December 2006 3.04
May 2006
3.00
Initial Public Release
18
Revision History
32559
Rev. 3.08
July 2007
Introduction and Overview
The BIOS and Kernel Developers Guide for AMD NPT Family 0Fh Processors is intended for programmers involved in the development of low-level BIOS (basic input/output system) functions, drivers, and operating system kernel modules. It assumes previous experience in microprocessor programming, as well as fundamental knowledge of legacy x86 and AMD64 microprocessor architecture. The reader should also have previous experience in BIOS or OS kernel design issues, as related to microprocessor systems, and a familiarity with various platform technologies, such as DDR, HyperTransport technology, and JTAG.
1.1
About This Guide
This guide covers the implementation-specific features of AMD NPT Family 0Fh Processors, as opposed to architectural features. AMD64 architectural features include the AMD64 technology, general-purpose, multimedia, and x87 floating-point registers, as well as other programmer-visible features defined to be constant across all processors. A subset of implementation-specific features are not defined by the processor architectural specifications. These implementation-specific features may differ in various details from one implementation to the next. Note: The term processor in this document refers to AMD NPT processor architecture. NPT, or New Platform Technology, refers to the technology developed to support AMD processors such as AMD F (1207), AMD AM2, and AMD S1g1 processors. The implementation-specific features covered in the following chapters include: Model-specific registers Processor initialization Integrated memory system configuration HyperTransport technology fabric initialization Performance monitoring and special debug features DRAM configuration Machine check error codes Thermal and power management
1.2
Related Documents
The references listed in Table 1 may prove invaluable towards a complete understanding of the subject matter in this volume. Chapter 1 Introduction and Overview 19
32559
Rev. 3.08 July 2007
Table 1.
Title
Related Documents
Order# 24592 24593 24594 (three-volume kit)
AMD64 Architecture Programmers Manual, Volume 1, Application Programming AMD64 Architecture Programmers Manual, Volume 2, System Programming AMD64 Architecture Programmers Manual, Volume 3, General-Purpose and System Instructions AMD64 Architecture Programmers Manual, Volume 4, 128-Bit Media Instructions AMD64 Architecture Programmers Manual, Volume 5, 64-Bit Media and x87 Floating-Point Instructions AMD Socket F (1207) Processor Functional Data Sheet AMD Socket AM2 Processor Functional Data Sheet AMD Socket S1g1 Processor Functional Data Sheet AMD NPT Processor Electrical Data Sheet AMD NPT Family 0Fh Server and Workstation Processor Power and Thermal Data Sheet AMD NPT Family 0Fh Desktop Processor Power and Thermal Data Sheet AMD NPT Family 0Fh Mobile Processor Power and Thermal Data Sheet AMD Athlon 64 and AMD Opteron Processors Thermal Design Guide
31118 31117 31731 31119 33953 33954 33952 26633
AMD Athlon 64 and AMD Opteron Processors Thermal Test Kit Instruc- 26956 tion Manual Thermal Design and Testing Guide for AMD NPT Family 0Fh Processors in Mobile Systems CPUID Guide for AMD Athlon 64 and AMD Opteron Processors
TM TM
32769 25481 33610 http://www.hypertransport.org
Revision Guide for the AMD NPT Family 0F Processors HyperTransport I/O Link Specification, Rev. 1.03
1.3
Conventions and Definitions
Some of the following definitions assume a knowledge of the legacy x86 architecture. See Table 1 for documents that include information on the legacy x86 architecture. Additional definitions are provided in the glossary at the end of this book, beginning on page 423. That glossary includes the most important terminology of the AMD64 architecture.
1.3.1
Notation
1011b. A binary value. In this example, a 4-bit value is shown. F0EAh. A hexadecimal value. In this example, a 2-byte value is shown. 20 Introduction and Overview Chapter 1
32559
Rev. 3.08
July 2007
[1,2]. A range that includes the left-most value (in this case, 1) but excludes the right-most value (in this case, 2). 74. A bit range, from bit 7 to 4, inclusive. The high-order bit is shown first. #GP(0). Notation indicating a general-protection exception (#GP) with error code of 0. ^ . Logical exclusive-OR operator; sometimes used as raised to the power of as well, as indicated by the context in which it is used. CR0CR4. A register range, from register CR0 through CR4, inclusive, with the low-order register first. CR0.PE = 1. Notation indicating that the PE bit of the CR0 register has a value of 1. DS:rSI. The contents of a memory location whose segment is DS and whose byte address is located in the rSI register. EFER.LME = 0. Notation indicating that the LME bit of the EFER register has a value of 0. FF /0. Notation indicating that FF is the first byte of an opcode and a sub-opcode field in the MODRM byte has a value of 0. DramEn. Notation indicating that the bit is an enable (En or EN is part of the name) and that a 1 specifies that the function is enabled. MemClkDis. Notation indicating that the bit is a disable (Dis or DIS is part of the name) and that a 1 specifies that the function is disabled.
1.4
Register Differences in Revisions of AMD NPT Family 0Fh Processors
Some changes to the register set are introduced with different silicon revisions. Refer to the Revision Guide for the AMD NPT Family 0F Processors, order# 33610 for information about how to identify different processor revisions. The following summarizes register changes after the initial revision. Revision F (differences from Revision E DDR1 Family 0Fh Processors) Function 3, Offset E4h, ThermtpSense (bit 3) renamed ThermtpSense0. Function 3, Offset E4h, TjOffset (bits 28-24), CurTmp (bits 23-16), ThermSenseSel (bit 6), ThermtpSense1 (bit 4) and ThermSenseCoreSel (bit2) added. Function 3, Offset B0h, On-Line Spare Control Register added. Function 2 Offset 40h, 44h, 48h, 4Ch, 50h, 54h, 58h, 5Ch, DRAM CS Base Registers modified. Function 2 Offset 60h, 64h, 68h, 6Ch, DRAM CS Mask Registers modified. Function 2 Offset 70h, 74h, 78h, 7Ch, DRAM CS Mask Registers removed. Function 2 Offset 80h, DRAM Bank Address Mapping Register modified. Introduction and Overview 21
Chapter 1
32559
Rev. 3.08 July 2007
Function 2 Offset 88h, DRAM Timing Low Register modified. Function 2 Offset 8Ch, DRAM Timing High Register modified. Function 2 Offset 90h, DRAM Configuration Low Register modified. Function 2 Offset 8Ch, DRAM Timing High Register modified. Function 2 Offset 94h, DRAM Configuration High Register modified. Function 2 Offset78h, DRAM Control Register added. Function 2 Offset 7Ch, DRAM Initialization Register added. Function 2 Offset 98h, DRAM Controller Additional Data Offset Register added. Function 2 Offset 9Ch, DRAM Controller Additional Data Register added. Function 2 Offset A0h, DRAM Controller Miscellaneous Register added. Function 2 Offset 98h, DRAM Delay Line Register removed. Function 2 Offset 9Ch, Scratch Register removed. MSR C001_0114 (VM_CR) added. MSR C001_0117 (VM_HSAVE_PA) added. MSR C001_0041h (FIDVID_CTL), NewVID (bit13) added. MSR C001_0042h (FIDVID_STATUS), AltVidOffset (bits 60-57), PstateStep (bit 56), MaxVID (bit 53), StartVID (bit 45), CurrVID (bit 37), and MaxRampVID (bit 28) added. MSR C001_0042h (FIDVID_STATUS) MinVID (61-56) removed. APIC 30h (APIC_VER), ExtApicSpace (bit 31) added. APIC 400h (EXT_APIC_FEAT) added. APIC 500h (THRCNT_INT0_LVT) added. MSR C001_0060h (BIST Results) added. Function 0 Offset 6Ch, BiosRstDet (bits 9-8) added. MSR C001_0010h (SYSCFG), Tom2ForceMemTypeWB (bit 22) added. Function 3 Offset A0h, Power Control Miscellaneous Register added. Function 3 Offset FCh, CPUID Family Model Register added. MSR 0413h (MC4_MISCDRAM Errors Threshold) added. Function 0, Offset 68h, InstallStateS (bit 23) added.
22
Chapter 1
32559
Rev. 3.08
July 2007
Function 3, Offset 44h, SyncOnDramAdrParErrEn (bit 30), DisMstAbtCpuErrRsp (bit 29) and DisTgtAbtCpuErrRsp (bit 28) added. Function 3, Offset D8h, GfxMode (bit 27) and AltVidTriEn (bit 26) added. Function 3 Offset 40h, DramParEn (bit 18) added. MSR C001_0055h (Interrupt Pending), C1eOnCmpHalt (bit 28) and SmiOnCmpHalt (bit 27) added. CPUID Fn[8000_0001]_EBX BrandId (bits 15-12) added. CPUID Fn[0000_0001]_ECX CMPXCHG16B (bit 13) added. CPUID Fn[8000_0001]_ECX LockMovCr0 (bit 4), ExtApicSpace (bit 3), and SVM (Bit 2) added. CPUID Fn[8000_0001]_EDX RDTSCP (bit 27) added. CPUID Fn[8000_000A] added.
Revision G CPUID Fn[8000_0001]_ECX 3DNowPrefetch (bit 8) added. CPUID Fn[8000_0007]_EDX 100MhzSteps (bit 6) added. CPUID Fn[8000_000A]_EDX LBR (bit 1) added. Function 2 Offset 9Ch, Index 04h, AtcDllMaxPhases (bit 28) added. Function 3 Offset E4h, CurTmp (bits 15-14) added.
Chapter 1
23
32559
Rev. 3.08 July 2007
24
Chapter 1
32559
Rev. 3.08
July 2007
Processor Initialization and Configuration
Each AMD NPT Family 0Fh Processor based system has one processor with its HyperTransport link connected to a HyperTransport I/O hub. When a reset signal is applied, this processor is initialized as the bootstrap processor (BSP). In a multiple processor system, any other processors are initialized as application processors (APs). The BSP begins executing code from the reset vector (0xFFFFFFF0), while each of the APs waits for its Request Disable bit to clear to 0. An AP does not fetch code until its Request Disable bit is cleared. Both BSP and AP operate in 16-bit Real mode after reset. The BSP has the Boot Strap Processor bit set in its APIC_BASE (001Bh) MSR, and each AP has this bit cleared. The processor node is addressed by its Node ID on the HyperTransport link and can be accessed with a device number in the PCI configuration space on Bus 0. The Node ID 0 is mapped to Device 24, the Node ID 1 is mapped to Device 25, and so on. The BSP is initialized with Node ID 0 after reset, and all APs are initialized with Node ID 7. The BSP is ready to access its Northbridge and memory controller after its routing table is enabled. The APs are not accessible from the BSP until the links and the routing table are configured. The initial processor states are described in the Processor Initialization State section of the AMD64 Architecture Programmer's Manual: Volume 2, System Programming.
2.1
Bootstrap Processor Initialization
The BSP is responsible for the execution of the BIOS Power-On Self Test (POST) and initialization of APs. The BSP must perform the following tasks: AP detection and routing table initialization Noncoherent HyperTransport device initialization Link width and frequency initialization Memory controller initialization on all processor nodes Address map table initialization
2.1.1
Detecting AP and Initializing Routing Table
The AMD Opteron processor has three HyperTransport links; therefore, an AMD Opteron system can have from one to eight processors. The BSP must detect all APs in the system, starting from its adjacent processor. Since the APs are initialized at Node 7, the BSP must set entry 7 of its Routing Table before the AP can be accessed. The adjacent AP will respond to a PCI configuration Chapter 2 Processor Initialization and Configuration 25
32559
Rev. 3.08 July 2007
read if it is present. A new Node ID can be assigned to the AP after the routing information is updated in the routing table. Chapter 9, HyperTransport Technology Configuration and Enumeration explains the steps for enumerating and initializing routing tables in one-processor (UP), twoprocessor (DP), and n processor systems. The AMD Athlon 64 processor has one HyperTransport link. Because one link must be connected to the HyperTransport I/O hub, an AMD Athlon 64 system can only be used in a uniprocessor system.
2.1.2
Initializing Noncoherent HyperTransport Technology Devices
Once all APs have been detected and the routing tables and link control registers have been initialized, the BSP must initialize noncoherent HyperTransport technology devices. A noncoherent HyperTransport technology device has either one or two links. A tunnel device has two links and a terminal device has one link. The HyperTransport I/O hub is a typical example of a terminal noncoherent HyperTransport device. The noncoherent HyperTransport device is identified by its Unit ID in a noncoherent HyperTransport link and can be accessed with a device number in the PCI configuration space. The device number in PCI space is the same as the Unit ID assigned to the noncoherent HyperTransport device. After reset, the Unit ID of each noncoherent HyperTransport technology device is initialized with a value of 0. When a PCI configuration read is performed from the BSP through a noncoherent HyperTransport link, the noncoherent HyperTransport device connected to the port responds. A new non-zero Unit ID value must be assigned to this device. The BIOS continues this process until no device responds at Device 0. The bus number range must be set to the noncoherent HyperTransport link on the BSP and in the address map table prior to the detection. The link control and status registers of a noncoherent device are implemented in the capability register block. The noncoherent HyperTransport technology devices connected to a port on the AP node can be detected in the same way. Again, the bus number range must be set to the noncoherent HyperTransport link on this AP and in the address map table. The noncoherent device can be detected with the starting bus number, Device 0 on the PCI configuration space. A noncoherent HyperTransport technology device may use more than one Unit ID. The new ID assigned to a device is its starting ID, the Base Unit ID (BaseUID). The next logic device in this noncoherent HyperTransport device can be identified with BaseUID + 1, and so on. The Unit ID Count field in the Capabilities register indicates how many Unit IDs this device uses.
2.1.3
Initializing Link Width and Frequency
The HyperTransport link width and frequency are initialized between the adjacent coherent and/or noncoherent HyperTransport technology devices during the reset sequence. After AP and noncoherent HyperTransport device detection, the link width and frequency can be changed based on the capability of the adjacent devices or the implementation of the system. See HyperTransport Link Frequency Selection on page 354 for more information. To make the new link width and 26 Processor Initialization and Configuration Chapter 2
32559
Rev. 3.08
July 2007
frequency take effect, an LDTSTOP_L needs to be asserted or a warm reset must be performed. The BIOS must ensure that LDTSTOP_L assertion for link width and frequency changes does not occur within 200 s of reset. LDTSTOP_L must be asserted for a minimum of 2 s for link width and frequency changes.
2.1.4
Initializing the Memory Controller on All Processor Nodes
The BSP is responsible for configuring the memory controller on all processor nodes and for initializing the DRAM map table. Chapter 5, DRAM Configuration, describes the functionality of the memory controller and the steps required to initialize memory. After memory configuration, the memory address MSRs must be set accordingly. The Top of Memory MSR is used for the top of DRAM below 4 Gbytes, and TOP_MEM2 is set to the top of DRAM above 4 Gbytes.
2.1.5
Initializing the Address Map Table
Each processor node has an address map table. This table contains the address map for the DRAM area, the address map for PCI memory spaces (MMIO), the address map for PCI I/O spaces, and the bus number range for each noncoherent HyperTransport link. The address map table must be duplicated on all processors in a system.
2.2
Application Processor Initialization
The application processor (AP) waits until its request disable bit is cleared to 0. The BIOS may clear this bit for an AP as soon as the AP node detection is completed and the routing tables are initialized, or just before the AP register initialization. When the request disable bit is cleared, the corresponding AP starts to fetch code from the reset vector (0xFFFFFFF0). The BSP bit in AP APIC_BASE register is cleared at reset, so it can be used to terminate the AP execution after the initial APIC initialization. The AP register initialization is the same as that of other AMD x86 processor families, such as the AMD Athlon processors.
Chapter 2
27
32559
Rev. 3.08 July 2007
2.3
BIOS Requirement for 64-Bit Operation
In general, the BIOS need not be aware of 64-bit mode in POST. There are no additional requirements to support 64-bit mode. The operating system determines whether to enable 64-bit mode after the BIOS invokes the operating system loader in legacy mode. There are two areas that need additional explanation: Sizing and testing memory above 4 Gbytes Using BIOS callback when in 64-bit mode
2.3.1
Sizing and Testing Memory above 4 Gbytes
AMD NPT Family 0Fh Processors have extended physical address extension (PAE) to support a 40bit address space. Thus, the BIOS can set up a 32-bit page table that allows it to size and test all physical memory in the system.
2.3.2
Using BIOS Callbacks in 64-Bit Mode
A BIOS callback is a mechanism that allows an operating system to call a service routine in the BIOS. BIOS callbacks on an AMD NPT Family 0Fh Processor platform in 64-bit mode are not supported by AMD64 operating systems. An operating system loader must be invoked in legacy mode with paging disabled, so that the loader can use 16/32-bit BIOS callbacks. ACPI code and operating system drivers are not affected by AMD64 operating systems that do not use a BIOS callback in 64-bit mode. ACPI is coded in ASL, which is not affected by the operating system mode. AMD64 operating system drivers are not allowed to use ROM callbacks.
28
Chapter 2
32559
Rev. 3.08
July 2007
3
3.1
Functional Description
Overview
The processor is designed to support performance desktop and workstation applications. It provides single or dual core capability, up to three high-performance HyperTransport links, as well as a single 128-bit high-performance DDR2 SDRAM memory controller. A block diagram of the processor is shown in Figure 1.
64-Kbyte 64-Kbyte L1 I-Cache L1 D-Cache
L2 Cache
512K/1M Option
CPU Core
64-Kbyte 64-Kbyte L1 I-Cache L1 D-Cache
L2 Cache
512K/1M Option
CPU Core
HyperTransport Interface
Northbridge
DDR2 SDRAM Interface
3 x 16/16 4002000 MT/s
Control PLLs and Clocks JTAG and Debug
Figure 1. Chapter 3
Processor Block Diagram Functional Description 29
128-bits DDR SDRAM 200/266/333/400 MHz 16-bits ECC
32559
Rev. 3.08 July 2007
3.2
Instruction Set Support
The processor supports the standard x86-instruction set defined in the AMD64 Architecture Programmers Manual, volumes 35, order# 24594. The processor also supports the following extensions to the standard x86 instruction set, which are described in the same volume set: AMD64 instructions MMX and 3DNow! technology instructions SSE, and SSE2, and SSE3 instructions
3.3
Package S1g1 AM2 F(1207)
Supported Packages
Memory I/Fs Unbuffered DIMMs; two channels Unbuffered DIMMs; two channels Registered DIMMs; two channels Links One 16-bit One 16-bit Notes Targets the mobile market. Targets the desktop market.
The processor is implemented in the following packages.
Three 16-bit Targets the server and workstation market.
Table 2.
Packages.
3.4
Address Space
The processor supports 48 address bits of virtual memory space (256 terabyte) as indicated by CPUID Fn8000_0008_EAX and 40 address bits of physical memory space (1 terabyte).
3.5
3.5.1
Internal Cache Structures

Level 1 Caches
The processor implements internal caching structures as described in the following sections.
The L1 data cache (L1 D-Cache) contains 64 Kbytes of storage organized as two-way set associative. The L1 data cache is protected with ECC. Two simultaneous 64-bit operations (load, store or combination) are supported. The L1 instruction cache (L1 I-Cache) contains 64 Kbytes of storage organized as 2-way associative. The L1 Instruction Cache is protected with parity.
3.5.2
Level 2 Cache
The L2 cache contains both instruction and data stream information. It is organized as 16-way setassociative. The L2 cache data and tag store is protected with ECC. When a given cache line in the L2 cache contains instruction stream information, the ECC bits associated with the given line are used to store predecode and branch prediction information. 30 Functional Description Chapter 3
32559
Rev. 3.08
July 2007
3.6
Northbridge
The Northbridge logic in the processor refers to the HyperTransport technology interface, the memory controller, and their respective interfaces to the CPU core. These interfaces are described in more detail in the following sections.
3.6.1
HyperTransport Technology Overview
The processor includes up to three 16-bit HyperTransport technology interfaces capable of operating up to 2000 mega-transfers per second (MT/s) with a resulting bandwidth of up to 8.0 Gbytes/s (4.0 Gbytes/s in each direction). Refer to Table 2 on page 30 for product-specific details on HyperTransport interfaces. The processor supports HyperTransport technology synchronous clocking mode. Refer to the HyperTransport I/O Link Specification (www.hypertransport.org) for details of link operation. Note: The processors HyperTransport links operate in Synchronous (Sync) mode as described in the HyperTransport I/O Link Specification. Sync mode operation requires the transmit and receive clocks for all links to be derived from the same time base. If different clock generators are used to drive the reference clock to the devices on both sides of the link, then the following requirements must be satisfied to ensure proper link operation: 1. All clock generators must use the same reference clock source. 2. Spread spectrum clocking must not be enabled in any of the clock generators unless the frequency deviations are synchronized between the outputs of all clock generators. 3.6.1.1 Link Initialization
The HyperTransport I/O Link Specification details the negotiation that occurs at power-on to determine the widths and rates used with the link. See Chapter 9, HyperTransport Technology Configuration and Enumeration on page 277 for information about link initialization and setup of routing tables. 3.6.1.2 HyperTransport Technology Transfer Speeds
The HyperTransport link of the processor is capable of operating at 400, 800, 1200, 1600, and 2000 MT/s. The link transfer rate is determined during the software configuration of the system, as specified in the HyperTransport I/O Link Specification. Processors with multiple HyperTransport links are capable of operating the links at different frequencies. There are three supported link frequency groups: 800 MHz, 400 MHz, 200 MHz 1000 MHz 600 MHz Functional Description 31
Chapter 3
32559
Rev. 3.08 July 2007
Processors can be configured with link frequencies from up to two link frequency groups. 3.6.1.3 System Address Map
System software must not map memory in the reserved HyperTransport address regions. The HyperTransport I/O Link Specification details the address map available to system hosts and devices. Downstream host accesses to reserved HyperTransport address regions result in a page fault. Upstream system device accesses to reserved HyperTransport address regions result in undefined operation.
3.6.2
DRAM Controller
The processors DRAM controller provides a programmable interface to a variety of standard DDR2 SDRAM DIMM configurations. The following features are supported: Self-Refresh mode Unbuffered and registered DIMMs. The package type determines the number of logical DIMMs supported and what type of DRAM is supported.
DIMM Support Per Package
Number of DIMMs per Channel Registered DIMMs 4 0 0 4 Rank Registered Unbuffered DIMMs DIMMs 2 0 0 0 2 0
Table 3.
Package F(1207) AM2 S1g1
SO-DIMMs 0 1 1
The controller provides programmable control of DRAM timing parameters to support the following memory speeds: 200-MHz (DDR2-400) PC2-3200 DIMMs 266-MHz (DDR2-533) PC2-4200 DIMMs 333-MHz (DDR2-667) PC2-5300 DIMMs 400-MHz (DDR2-800) PC2-6400 DIMMs DRAM devices that are 4, 8 and 16 bits wide. DIMM sizes from 128 Mbytes to 16 Gbyte. Interleaving memory within DIMMs. Stacked registered DIMMs. ECC checking with double-bit detect with single-bit correct in AM2 and F(1207) packages. Functional Description Chapter 3
32
32559
Rev. 3.08
July 2007
May be configured for 32-byte or 64-byte burst length. Programmable page-policy: Supports up to sixteen open pages total across all chip-selects. Statically idle open-page time. Optional dynamic precharge control based on page-hit/miss history.
For programming information and specific details of these features, see 4.5 DRAM Controller page 87. 3.6.2.1 DRAM Operation
At power on reset, the M[B, A]_CLK_H/L and M[B, A]_RESET_L (RDIMMs only) pins are driven low while the processor PLLs are ramping. Clocks are driven on the M[B, A]_CLK_H/L pins only after BIOS programs the appropriate clock ratio value in the memory controller configuration registers. The actual DRAM frequency may vary for some speeds based on the CPU clock multiplier and other configuration options, as shown in Table 4 on page 33 (the memory controller automatically adjusts refresh counters at all speeds as required to meet the device refresh specifications).
Table 4.
DRAM Interface Speed vs. CPU Core Clock Multiplier

Core Frequency 800 MHz 900 MHz 1000 MHz 1100 MHz 1200 MHz 1300 MHz 1400 MHz 1500 MHz 1600 MHz 1700 MHz 1800 MHz 1900 MHz DRAM Frequency 200 MHz 160.00 180.00 200.00 183.33 200.00 185.71 200.00 187.50 200.00 188.89 200.00 190.00 266 MHz 160.00 180.00 200.00 220.00 240.00 260.00 233.33 250.00 266.67 242.86 257.14 237.50 333 MHz 160.00 180.00 200.00 220.00 240.00 260.00 280.00 300.00 320.00 283.33 300.00 316.67 400 MHz 160.00 180.00 200.00 220.00 240.00 260.00 280.00 300.00 320.00 340.00 360.00 380.00
Multiplier 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5
Chapter 3
33
32559
Rev. 3.08 July 2007
Table 4.
DRAM Interface Speed vs. CPU Core Clock Multiplier

Core Frequency 2000 MHz 2100 MHz 2200 MHz 2300 MHz 2400 MHz 2500 MHz 2600 MHz 2700 MHz 2800 MHz 2900 MHz 3000 MHz 3100 MHz DRAM Frequency 200 MHz 200.00 190.9 200.00 191.67 200.00 192.3 200.00 192.86 200.00 193.33 200.00 193.75 266 MHz 250 262.50 244.44 255.56 266.67 250.00 260.00 245.46 254.55 263.64 250.00 258.33 333 MHz 333.33 300.00 314.29 328.57 300.00 312.50 325.00 300.00 311.11 322.22 333.33 310.00 400 MHz 400.00 350.00 366.67 383.33 400.00 357.14 371.43 385.71 400.00 362.50 375.00 387.50
Multiplier 10 10.5 11 11.5 12 12.5 13 13.5 14 14.5 15 15.5
Note: Fractional (half-step)multipliers are supported in revision G and later revisions.
The use of 2T timing allows support of many DIMM combinations at the higher DDR speeds. The 2T timing feature causes commands and addresses to be driven for two clock cycles and qualified with an associated chip select on the second clock cycle, allowing an extra clock of setup to accommodate heavy DIMM loading (such as double-rank DIMMs). The use of 2T timing is only supported for unbuffered DIMMs. The controller supports programmable timing and refresh. Auto-refresh is supported and is staggered by tRFC across chip-selects to reduce system noise. Unpopulated DIMM slots are not refreshed. 3.6.2.2 DRAM Power Management
The memory controller supports self-refresh mode to accommodate various power management states such as ACPI C3, S1, and S3 states. The M[B,A]_RESET_L pin is provided for resetting the registers on registered DDR2 SDRAM DIMMs as required for the S3 (Suspend to RAM) power management state.
3.6.3
Main Memory Hardware Scrubbing
The memory controller scrubs the main memory arrays to prevent the build up of soft errors. Any correctable or non-correctable errors are logged to the machine check logs and can be programmed to
34
Chapter 3
32559
Rev. 3.08
July 2007
invoke the machine check interrupt. The scrubbing function can be used in three modes as described in the following sections. 3.6.3.1 Sequential Scrubbing
In this mode, the scrubber sequentially proceeds through main memory, performing a read-write cycle or a read-modify-write cycle if a correctable error is found. The scrubber scrubs one cache line on each scrub interval that is programmable from 40 ns to 84 ms. 3.6.3.2 Source Correction Scrubbing
In this mode, the scrubber is directed to scrub any cache line that is the source of any corrected error during normal accesses. During normal operation when source correction scrubbing is disabled, single-bit errors are corrected on the fly and the corrected data is passed without updating the source memory location. When source scrubbing is enabled the scrubber also corrects the source memory location. 3.6.3.3 Sequential Plus Source Correction Scrubbing
When both sequential and source correction scrubbing are enabled, the scrubber sequentially proceeds through main memory. If a correctable error is detected during normal operation, the scrubber is redirected to the location of the error, and after it corrects that location in main memory it resumes sequential scrubbing at the previous location.
3.6.4
Memory Interleaving Modes
Interleaving is defined as the spreading contiguous physical address space over multiple DIMM banks, as opposed to each DIMM owning a single contiguous address space. This is accomplished by using lower-order address bits to select between DIMMs. The processor supports two different types of interleaving modes: CS: interleaving between the DIMM banks of a channel. This is controlled though DRAM CS Base Address Registers (Function 2: Offsets 40h-5Ch). Node: interleaving between DIMMs of different processor nodes. This is controlled though DRAM Base i Registers (Function 1: Offsets 40h, 48h, 50h, 58h, 60h, 68h, 70h, 78h) and DRAM Limit i Registers (Function 1: Offsets 44h, 4Ch, 54h, 5Ch, 64h, 6Ch, 74h, 7Ch). Any combination of these interleaving modes may be enabled concurrently.
3.6.5
Memory Hoisting
Typically, memory mapped I/O space is mapped from MMIOLowAddr to 4GB-1, and DRAM memory is mapped from address 0 to TOP_MEM, an address that is less or equal to MMIOLowAddr-1. DRAM memory can also be mapped from 4GB to TOM2, an address greater than
Chapter 3
35
32559
Rev. 3.08 July 2007
4GB. Memory hoisting is defined as reclaiming the DRAM space that would naturally reside in the MMIO hole just below the 4G address level. Memory that resides in the MMIO is repositioned above the 4G level by programming the DRAM Hole Address Register and the DramLimit properly. The memory hoisting offset field, DramHoleOffset, is programmed as shown in the following equation: DramHoleOffset[31:24] = ( (100h - DramHoleBase[31:24]) + DramBase[31:24]); The DramLimit field must be programmed to include the size of the DRAM hole as shown in the following equation: DramLimit = 4GB + Hole Size; 3.6.5.1 Non-Node-Interleave Mode
For a system that is configured in the non-node-interleave mode, the DRAM Hole Address Register is enabled only on the node that has its DRAM space overlapped with the MMIO hole if the calculated DramHoleOffset is not equal to 4GB. The DramLimit of the current node must be adjusted by adding the size of the MMIO hole. The DramBase and DramLimit registers for the following nodes must also adjusted by adding the size of the MMIO hole. Example. A two processor system with 4 GB DRAM on each processor and a 1 GB MMIO hole. Processor 0 DramBase[39:24] = 0000h Processor 0 DramLimit[39:24] = 013Fh (5GB -1; DramLimit + Hole Size) Processor 0 DramHoleBase[31:24] = C0h (3GB; Starting address of MMIO hole) Processor 0 DramHoleOffset [31:24] = 40h (1GB) Processor 0 DramHoleValid = 1(Memory hoisting enabled) Processor 1 DramBase[39:24] = 0140h (5GB) Processor 1 DramLimit[39:24] = 023Fh (9GB -1) Processor 1 DramHoleBase[31:24] = 00h (Memory hoisting is not enabled on node 1) Processor 1 DramHoleOffset [31:24] = 00b (Memory hoisting is not enabled on node 1) Processor 1 DramHoleValid = 0 (Memory hoisting disabled) If the calculated DramHoleOffset is equal to 4GB, meaning the DramBase of the current node is aligned to the starting address of the MMIO hole, the DRAM Hole Address Register must not be enabled. Both the DramBase and DramLimit of current node, and for the following nodes, must be adjusted by adding the size of the MMIO hole. Example. A two processor system with 2GB memory on each processor and a 2GB MMIO hole. Processor 0 DramBase[39:24] = 0000h Processor 0 DramLimit[39:24] = 007Fh (2GB - 1) 36 Functional Description Chapter 3
32559
Rev. 3.08
July 2007
Processor 0 DramHoleBase[31:24] = 00h Processor 0 DramHoleOffset[31:24] = 00h (calculated DramHoleOffset = 4GB) Processor 0 DramHoleValid = 0 (Memory hoisting disabled) Processor 1 DramBase[39:24] = 0100h (4GB; DramBase + hole size) Processor 1 DramLimit[39:24] = 017Fh (6GB - 1; DramLimit + hole size) Processor 1 DramHoleBase[31:24] = 00h Processor 1 DramHoleOffset[31:24] = 00h Processor 1 DramHoleValid = 0 (Memory hoisting disabled) 3.6.5.2 Node-Interleave Mode
For a system with the node-interleaving enabled, the DRAM Hole Address Register is enabled on every node with the same setting in its DramHoleBase and DramHoleOffset fields. The DramLimit of every node in the DRAM address map must be adjusted by adding the size of the MMIO hole. Example. A two processor system with 2GB memory on each processor and a 2GB MMIO hole with the node-interleave mode enabled. Processor 0 DramBase[39:24] = 0000h Processor 0 DramLimit[39:24] = 017Fh (6GB - 1; DramLimit + hole size) Processor 0 DramHoleBase[31:24] = 80h (2G; Starting address of MMIO hole) Processor 0 DramHoleOffset[31:24] = 80h(Calculated DramHoleOffset = 2GB) Processor 0 DramHoleValid = 1 Processor 1 DramBase[39:24] = 0000h (Same as processor 0) Processor 1 DramLimit[39:24] = 017Fh (Same as processor 0) Processor 1 DramHoleBase[31:24] = 80h (Same as processor 0) Processor 1 DramHoleOffset[31:24] = 80h (Same as processor 0) Processor 1 DramHoleValid = 1
3.6.6
On-Line Spare
On-line spare is a RAS mechanism supported in F(1207) and AM2 packages that allows the system to reserve one rank of one logical DIMM to be used as a spare rank. System software reserves a spare rank by setting the spare bit (Function 2 Offsets 40h, 44h, 48h, 4Ch, 50h, 54h, 58h, 5Ch bit 1) in one of the CS Base address registers. The spare rank must be greater than or equal to the size of all other ranks in the system. On-line spare should never be used in a UMA system. The system can switch to the spare rank when system software determines that one of the ranks in the system is no longer functioning properly and needs to be replaced. The system software initiates the
Chapter 3
37
32559
Rev. 3.08 July 2007
swap to the spare rank by writing the chip select number of the bad rank to BadDramCS (Function 3 Offset B0h bits 6:4) and setting the swap enable SwapEn (Function 3 Offset B0h bit 0). 3.6.6.1 On-Line Spare and CS Interleaving
The on-line spare feature can only be used with 2 way and 4 way CS interleaving under the following conditions. All ranks of each DIMM present must be of the same size and configuration. Only the following populations are supported: 2 DIMMs per channel (2 way CS interleaving) One single rank DIMM and one dual rank DIMM. Any rank can be used as the spare rank. Both DIMMs are dual rank. Any rank can be used as the spare rank. One rank must be marked as bad since only two ranks can be active. 3 DIMMs per channel (4 way CS interleaving) Two dual rank DIMMs and one single rank DIMM. Any rank can be used as the spare rank. All DIMMs are dual rank. Any rank can be used as the spare rank. One rank must be marked as bad since only four ranks can be active. 4 DIMMs per channel (4 way CS interleaving) One dual rank DIMMs and three single rank DIMMs. Any rank can be used as the spare rank. Two dual rank DIMMs and two single rank DIMMs. Any rank can be used as the spare rank. One rank must be marked as bad since only four ranks can be active.
3.7
Power Management
The processor provides the following power management features designed to be compliant with the Advanced Configuration and Power Interface (ACPI) Specification and HyperTransport technology: For more details see Chapter 10, Power and Thermal Management on page 281 38 Halt state with associated programmable power savings STPCLK/Stop Grant protocol capable of supporting eight distinct versions of Stop Grant LDTSTOP_L signal support Memory controller and host bridge power management Processor Performance state (P-state) transition support Voltage plane isolation based upon PWROK signal Low-power state while RESET_L signal is asserted Hardware Thermal Control Functional Description Chapter 3
32559
Rev. 3.08
July 2007
Software Thermal Control
3.7.1

PWROK
When PWROK is deasserted, the processor performs the following steps: Isolates its VDDIO- and VTT-powered logic from all other internal logic to prevent leakage current paths between power planes. Tristates all DDR SDRAM I/O pins except for the M[B,A]_CKE and M[B,A]_RESET_L outputs, which are driven Low. Drives its VID[5:0] outputs to the value that selects the startup core voltage level. RESET_L and M[B,A]_RESET_L
3.7.1.1
When RESET_L is asserted, the processor performs the following steps: The processor core is held in a low-power state. The M[B,A]_CKE and M[B,A]_RESET_L outputs are forced low.
After RESET_L is deasserted, BIOS must program the appropriate clock divisor in the memory controller configuration registers, causing the M[B,A]_CLK_H/L clocks to be driven.
3.7.2
Thermal Sensor
Each processor core contains two thermal sensors. The temperature calculated by the thermal sensor is used by HTC, STC, THERMTRIP_L, and the PROCHOT_L signal. 3.7.2.1 Software Thermal Control
Software Thermal Control (STC) is a software mechanism used to reduce power consumption. When the die temperature becomes greater than the high temperature threshold, or when the die temperature becomes less than the low temperature threshold after the high temperature threshold was exceeded, an interrupt can be generated. There are two ways to generate an interrupt: (1) via APIC Thermal Local Vector Table Register that allows software to define the interrupt type, or (2) via a special bus cycle, and the I/O hub. Inside the interrupt handler software may take an action that will result in power consumption reduction. For example, it can turn the fan on/off, enable normal StpClk throttling, or do a FID/VID change. 3.7.2.2 Hardware Thermal Control
Hardware Thermal Control (HTC) is a hardware mechanism used to reduce processor power consumption. The processor enters HTC active state when the die temperature is greater than the HTC temperature limit or when PROCHOT_L is asserted. While it is in HTC active state, processor core clock grid (GCLK) is periodically ramped down for a specified period of time, and consequently processor performance is reduced. Deterministic power savings are accomplished by unconditional Chapter 3 Functional Description 39
32559
Rev. 3.08 July 2007
GCLK ramping down. After the temperature drops below the HTC temperature limit (plus a specified number of degrees of hysteresis), the processor automatically exits the HTC-active state if PROCHOT_L is not asserted. The processor can enter the HTC-active state only when PWROK and RESET_L are both high. The processor can not enter the HTC-active state while in the C3 ACPI state. Special bus cycles can be generated when HTC active state is entered or exited. The Northbridge clock grid (BCLK) is not ramped down when HTC is active. 3.7.2.3 PROCHOT_L
The PROCHOT_L pin is asserted as an input to force the processor into the HTC-active state or becomes asserted as an output to indicate when the die temperature is greater than the HTC temperature limit. As in input to the processor, PROCHOT is ignored if THERMTRIP_L, PWROK or RESET_L is low or if HtcEn (Function 3 Offset 64h) is zero. As an output, it is low to reflect the HTC-active state. PROCHOT_L will be recognized when it's state has been stable high or low for greater than 10ns, however to get any effective power reduction from PROCHOT_L assertion, the assertion time must allow for several throttling periods to elapse. Each throttling period is approximately 6us. See BIOS Requirements for PROCHOT_L Throttling on page 286 for how to enable PROCHOT_L based throttling. 3.7.2.4 THERMTRIP_L
The processor provides a hardware-enforced thermal protection mechanism. When the processors die temperature exceeds a specified temperature, the processor is designed to stop its internal clocks and assert the THERMTRIP_L output. THERMTRIP_L assertion is only valid when PWROK is asserted and RESET_L is deasserted. THERMTRIP_L assertion indicates the processor die temperature has exceeded normal operating parameters. PWROK must be deasserted in response to a THERMTRIP_L assertion to enable proper processor operation. Once asserted THERMTRIP_L remains asserted until RESET_L is asserted. If the processors die temperature still exceeds the thermal trip point when RESET_L is deasserted, THERMTRIP_L will immediately be reasserted and the processors internal clocks stop.
3.7.3
Thermal Diode
An on-die thermal diode is provided as a tool for thermal management. An external sensor is necessary to measure the temperature of the thermal diode. Hardware thermal solutions for lidded processors should be designed and validated against the case temperature specification per the methodology specified in AMD Athlon 64 and AMD Opteron Processors Thermal Design Guide, order# 26633. Hardware thermal solutions for lidless processors
40
Chapter 3
32559
Rev. 3.08
July 2007
should be designed against the die temperature specification in Thermal Design and Testing Guide for AMD NPT Family 0Fh Processors in Mobile Systems, order# 32769. System thermal management (e.g. fan control) for lidded processors using the thermal diode should be designed against the case temperature specification. See TCASE max and TCONTROL max for Socket F (1207) and Socket AM2 Processors on page 41 for more information. 3.7.3.1 TCASE max and TCONTROL max for Socket F (1207) and Socket AM2 Processors
TCASE max is the maximum case temperature specification for a lidded processor. Thermal solutions should be designed to this specification. The thermal test procedure is defined in AMD Athlon 64 and AMD Opteron Processors Thermal Test Kit Instruction Manual, order# 26956. TCASE max is a physical temperature specification in degrees Celsius that can be measured at the center of the lid with a thermocouple. The correct method for measuring case temperature is discussed in the AMD Athlon 64 and AMD Opteron Processors Thermal Design Guide, order# 26633. The case temperature specification is provided in the power and thermal data sheet (AMD NPT Family 0Fh Server and Workstation Processor Power and Thermal Data Sheet, order# 33953 and AMD NPT Family 0Fh Desktop Processor Power and Thermal Data Sheet, order# 33954). TCONTROL max represents the maximum allowed TCONTROL value for the processor to be within its functional temperature specification. The value for TCONTROL max is provided in the power and thermal data sheet. TCONTROL (processor control temperature) is a non physical temperature on an arbitrary scale that can be used for system thermal management policies. TCONTROL does not equal the physical temperature of the die, lid, or any other temperature. Under low power conditions, TCONTROL can be less than the ambient temperature. Negative TCONTROL values are undefined. TCONTROL= dual sourcing current temperature measurement + TOFFSET TOFFSET (thermal diode temperature offset) is supplied as a read only field in the Thermtrip Status Register. System thermal management should be designed to prevent the case temperature from being exceeded even in transient situations. For example if the processor is in an ACPI C1 Halt state with a low fan speed and a high power application is started, the fan speed policy must ensure that the processor TCONTROL never exceeds the TCONTROL max limit. This requires increasing the fan speed before reaching TCONTROL max. Additionally, the accuracy of the thermal diode monitor, typically 1 C to 5 C, must be considered when setting thermal trip points.
Chapter 3
41
32559
Rev. 3.08 July 2007
T CONTROL ma x T CASE max T CASE
Temperature
T CONTROL
when
T CASE max reached T CONTROL reads T CONTROL max.
Power
Figure 2. 3.7.3.2 TCASE Max and TCONTROL Max Relationship TDIE max and TCONTROL for Socket S1g1 Processors
TDIE max is the maximum die temperature specification for a lidless processor. Thermal solutions should be designed to this specification. TDIE max is a physical temperature specification in degrees Celsius that can be measured at the die with a thermocouple. The correct method for measuring die temperature is discussed in the Thermal Design and Testing Guide for AMD NPT Family 0Fh Processors in Mobile Systems, order# 32769. TCONTROL = dual sourcing current temperature measurement + TOFFSET TOFFSET (thermal diode temperature offset) is supplied as a read only field in the Thermtrip Status Register. For lidless processors, this is the ideality factor which calibrates the monitor reading to the actual die temperature. System thermal management should be designed to prevent the processor control temperature from exceeding the maximum die temperature even in transient situations. For example if the processor is in an ACPI C1 Halt state with a low fan speed and a high power application is started, the fan speed policy must ensure that the processor TCONTROL never exceeds the TDIE max limit. This requires increasing the fan speed before reaching TDIE max. Additionally, the accuracy of the thermal diode monitor, typically 1 C to 5 C, must be considered when setting thermal trip points.
42
Chapter 3
32559
Rev. 3.08
July 2007
4
4.1
Memory System Configuration

Configuration Space Accesses
The AMD NPT Family 0Fh Processors implement configuration space as defined in the PCI Local Bus Specification, Rev. 2.2, and the HyperTransport I/O Link Specification, Rev. 1.03. Both coherent HyperTransport links and the compatibility bus (the bus to the HyperTransport I/O hub) are accessed through Bus 0. Configuration accesses to Bus 0, devices 0 to 23, are transmitted to the compatibility noncoherent HyperTransport bus. Coherent HyperTransport device configuration space is accessed by using Bus 0, devices 24 to 31, where Device 24 corresponds to Node 0 and Device 31 corresponds to Node 7. All configuration cycles are type 1 on coherent HyperTransport links. When transmitted down a noncoherent HyperTransport chain by the host bridge, configuration cycles are translated to type 0 if they target the noncoherent HyperTransport bus immediately behind the host bridge. If they target a subordinate bus, they remain as type 1. Accesses to memory system configuration registers are controlled through the Configuration Address register (see Configuration Address Register on page 43) and the Configuration Data register (see Configuration Data Register on page 44), which can be accessed through I/O reads/writes to addresses 0CF8h and 0CFCh, respectively.
4.1.1
Configuration Address Register
Writes to the Configuration Address register specify the target of a configuration access in terms of the bus, device, function, and register. Access to the Configuration Address register must be full doubleword reads or writes. To access one of the memory system configuration registers defined in this chapter, the bus number should be 0, the device number should be the target processor node number plus 24, and the function and register number should be set based on the register function and offset as specified in this chapter. When the Enable bit of the Configuration Address register is set, reads and writes to the Configuration Data register access the register specified in the Configuration Address register. Configuration Address Register
31 EnReg reserved 24 23 BusNum 16 15 DevNum 11 10 8
0CF8h (doubleword)
7 RegNum 2 1 reserved 0
FuncNum
Chapter 4
43
32559
Rev. 3.08 July 2007
Bit 31 3024 2316 1511 108 72 10
Mnemonic EnReg reserved BusNum DevNum FuncNum RegNum reserved
Function Enable Register Bus Number Device Number Function Number Register Number
R/W R/W R R/W R/W R/W R/W R
Field Descriptions Register Number (RegNum)Bits 72. Specifies the doubleword offset of the configuration address. Function Number (FuncNum)Bits 108. Specifies the function number of the configuration address. Device Number (DevNum)Bits 1511. Specifies the device number of the configuration address. Bus Number (BusNum)Bits 2316. Specifies the bus number of the configuration address. Enable (EnReg)Bit 31. When this bit is set, aligned doubleword accesses to the Configuration Data Register result in configuration-space transactions. 0 = Configuration transactions disabled. 1 = Configuration transactions enabled.
4.1.2
Configuration Data Register
Accesses to the Configuration Data Register are translated to configuration-space transactions at the address specified by the Configuration Address Register. Configuration Data Register
31 CfgData
0CFCh (Doubleword)
0
Bits 310
Mnemonic CfgData
Function Configuration Data
R/W R/W
Field Descriptions Configuration Data (CfgData)Bits 310.
44
Chapter 4
32559
Rev. 3.08
July 2007
4.2
Memory System Configuration Registers
The AMD NPT Family 0Fh Processor implement memory system configuration registers in the Northbridge block of the processor. These registers are mapped into the coherent HyperTransport technology configuration space (Bus 0, device numbers 2431). The number of devices implemented is equal to the number of processor nodes in the system. Configuration space Device 24 corresponds to Node 0 and is normally the bootstrap processor (BSP). Configuration space device numbers 25 through 31 correspond to nodes 1 through 7, which may or may not be implemented. The processor implements configuration registers in PCI configuration space using the following four headers: Function 0: HyperTransport technology configuration (see page 45) Function 1: Address map configuration (see page 72) Function 2: DRAM configuration (see page 87) Function 3: Miscellaneous configuration (see page 128)
Each of these functions are detailed in the following sections. Note: Contents of some fields in the headers and HyperTransport technology capability blocks are maintained through a warm reset. If not specified otherwise, a field is initialized by a warm reset.
4.3
Table 5.
Offset 00h 04h 08h
Function 0HyperTransport Technology Configuration

Function 0 Configuration Registers
Register Name Device ID Status Base Class Code BIST Subclass Code Header Type Vendor ID (AMD) Command Program Interface Latency Timer
1
Reset 1100_1022h 0010_0000h Revision ID Cache Line Size 0600_0000h
Access RO RO RO
Description page 47 page 48 page 48
0Ch 10h
0080_0000h 0000_0000h
RO RO
page 49
Base Address 0
Notes: 1. The unimplemented registers in the standard PCI configuration space are implemented as read-only and return 0 if read. 2. Reads and writes to unimplemented registers in the extended PCI configuration space will result in unpredictable behavior.
Chapter 4
45
32559
Rev. 3.08 July 2007
Table 5.
Offset 14h 18h 1Ch 20h 24h 28h 2Ch 30h 34h 38h 3Ch 40h 44h 48h 4Ch 50h 54h 58h 5Ch 60h 64h 68h 6Ch 80h 84h 88h 8Ch 90h 94h 98h A0h A4h
Function 0 Configuration Registers (Continued)

Register Name Base Address 1 Base Address 2 Base Address 3 Base Address 4 Base Address 5 Card Bus CIS Pointer Sub-System ID ROM Base Address Capabilities Pointer reserved Max Latency Min GNT Int Pin Int Line Reset 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h Sub-System Vendor ID 0000_0000h 0000_0000h page 49 0000_0000h 0000_0000h 0001_0101h 0001_0101h 0001_0101h 0001_0101h 0001_0101h 0001_0101h 0001_0101h 0001_0101h 0000_0000h 0000_00E4h 0F00_0000h page 59 page 60 0011_0000h page 66 0000_0002h page 68 0000_0000h page 70 page 60 0011_0000h Access RO RO RO RO RO RO RO RO RO RO RO RW RW RW RW RW RW RW RW RW RW RW RW RO RW RW RO RW RW RO RO RW page 50 page 50 page 50 page 50 page 50 page 50 page 50 page 50 page 52 page 53 page 54 page 59 page 60 page 62 page 66 page 67 page 68 page 70 page 70 page 60 page 62 page 49 Description
Routing Table Node 0 Routing Table Node 1 Routing Table Node 2 Routing Table Node 3 Routing Table Node 4 Routing Table Node 5 Routing Table Node 6 Routing Table Node 7 Node ID Unit ID HyperTransport Transaction Control HyperTransport Initialization Control LDT0 Capabilities LDT0 Link Control LDT0 Frequency/Revision LDT0 Feature Capability LDT0 Buffer Count LDT0 Bus Number LDT0 Type LDT1 Capabilities LDT1 Link Control
46
Chapter 4
32559
Rev. 3.08
July 2007
Table 5.
Offset A8h ACh B0h B4h B8h C0h C4h C8h CCh D0h D4h D8h

Register Name LDT1 Frequency/Revision LDT1 Feature Capability LDT1 Buffer Count LDT1 Bus Number LDT1 Type LDT2 Capabilities LDT2 Link Control LDT2 Frequency/Revision LDT2 Feature Capability LDT2 Buffer Count LDT2 Bus Number LDT2 Type Reset page 66 0000_0002h page 68 0000_0000h page 70 page 60 0011_0000h page 66 0000_0002h page 68 0000_0000h page 70 Access RW RO RW RW RO RO RW RW RO RW RW RO Description page 66 page 67 page 68 page 70 page 70 page 60 page 62 page 66 page 67 page 68 page 70 page 70
4.3.1
Device/Vendor ID Register
This register specifies the device and vendor IDs for the Function 0 registers and is part of the standard PCI configuration header. Device/Vendor ID Register
31 DevID 16 15 VenID
Function 0: Offset 00h

0
Bits 3116 150
Mnemonic DevID VenID
Function Device ID Vendor ID
R/W R R
Reset 1100h 1022h
Field Descriptions Vendor ID (VenID)Bits 150. This read-only value is defined as 1022h for AMD. Device ID (DevID)Bits 3116. This read-only value is defined as 1100h for the HyperTransport technology configuration function.
Chapter 4
47
32559
Rev. 3.08 July 2007
4.3.2
Status/Command Register
This register contains status and command information for the Function 0 registers and is part of the standard PCI configuration header. Status/Command Register
31 Status 16 15 Command

0
Bits 3116 150
Mnemonic
Function Status Command
R/W R R
Reset 0010h 0000h
Field Descriptions CommandBits 150. This read-only value is defined as 0000h. StatusBits 3116. This read-only value is defined as 0010h to indicate that the processor has a capabilities list containing configuration information specific to HyperTransport technology.
4.3.3
Class Code/Revision ID Register
This register specifies the class code and revision for the Function 0 registers and is part of the standard PCI configuration header. Class Code/Revision ID Register
31 BCC 24 23 SCC 16 15 PI

8 7 RevID 0
Bits 3124 2316 158 70
Mnemonic BCC SCC PI RevID
Function Base Class Code Subclass Code Programming Interface Revision ID
R/W R R R R
Reset 06h 00h 00h 00h
Field Descriptions Revision ID (RevID)Bits 70. Programming Interface (PI)Bits 158. This read-only value is defined as 00h. Sub Class Code (SCC)Bits 2316. This read-only value is defined as 00h. 48 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
Base Class Code (BCC)Bits 3124. This read-only value is defined as 06h for a host bridge device.
4.3.4
Header Type Register
This register specifies the header type for the Function 0 registers and is part of the standard PCI configuration header. Header Type Register
31 BIST 24 23 HType 16 15 LatTimer
Function 0: Offset 0Ch

8 7 CLS 0
Bits 3124 2316 158 70
Mnemonic BIST HType LatTimer CLS
Function BIST Header Type Latency Timer Cache Line Size
R/W R R R R
Field Descriptions CacheLineSize (CLS)Bits 70. This read-only value is defined as 00h. LatencyTimer (LatTimer)Bits 158. This read-only value is defined as 00h. HeaderType (HType)Bits 2316. This read-only value is defined as 80h to indicate that multiple functions are present in the configuration header and that the header layout corresponds to a device header as opposed to a bridge header. BISTBits 3124. This read-only value is defined as 00h.
4.3.5
Capabilities Pointer Register
This register contains a capabilities pointer to a linked capabilities list of registers specific to HyperTransport technology, with one set for each HyperTransport link. It is part of the standard PCI configuration header. Capabilities Pointer Register
31 reserved

8 7 CapPtr 0
Chapter 4
49
32559
Rev. 3.08 July 2007
Bits 318 70
Mnemonic reserved CapPtr
Function Capability Pointer
R/W R R
Reset 0
Field Descriptions Capability Pointer (CapPtr)Bits 70. Points to the first available HyperTransport technology capabilities block. Depending on the specific HyperTransport link configuration physically present on the processor, this will be either 80h (LDT0), A0h (LDT1) or C0h (LDT2).
4.3.6
Routing Table Node i Registers
The routing table contains three entriesone for requests RQRoute[7:0], one for responses RPRoute[7:0], and one for broadcasts BCRoute[7:0]. A maximum of eight nodes with three links per node is supported in an AMD Opteron system. These table entries can be read and written, and its contents are not maintained through a warm reset. At reset, all table entries are initialized to the value 01h, indicating that packets should be accepted by this node. Care must be exercised when changing table entries after reset to ensure that connectivity is not lost during the process. 4.3.6.1 Request Routing Table
Each node contains a routing table for use with directed requests, with one entry for each of the eight possible destination Node IDs. The value in each entry indicates which outgoing link is used for request packets directed to that particular destination node (DstNode). A 1 in a given bit position indicates that the request is routed through the corresponding output link. Bit 0, when set to 1, indicates that the request must be accepted by this node. 4.3.6.2 Response Routing Table
Each node contains a routing table for use with responses, with one entry for each of the eight possible destination node IDs. The value in each entry indicates which outgoing link is used for response packets directed to that particular destination node (DstNode). A 1 in a given bit position indicates that the response is routed through the corresponding output link. Bit 0, when set to 1, indicates that the response must be accepted by this node. 4.3.6.3 Broadcast Routing Table
Each node contains a routing table for use with broadcast and probe requests with one entry for each of the eight possible source node IDs. Each entry contains a single bit for each of the outbound links from the node. The packet is forwarded on all links with their corresponding links set to a 1. Bit 0 when set to 1 indicates that the broadcast must be accepted by this node. The algorithm to generate the routing table follows the backbone first rule. 50 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
Routing Table Node 07 Registers

31 reserved 20 19 BCRte 16 15
Function 0: Offset 40h, 44h, 48h, 4Ch, 50h, 54h, 58h, 5Ch
12 11 reserved RPRte 8 7 reserved 4 3 RQRte 0
Bits 3120 1916 1512 118 74 30
Mnemonic reserved BCRte reserved RPRte reserved RQRte
Function Broadcast Route Response Route Request Route
R/W R R/W R R/W R R/W
Reset 000h 1h 0h 1h 0h 1h
Field Descriptions Request Route (RQRte)Bits 30. The Request Route field defines the node or link to which a request packet is forwarded. Request packets are routed to only one destination and the table index is based on the destination Node ID. Bit [0] = Route to this node Bit [1] = Route to Link 0 Bit [2] = Route to Link 1 Bit [3] = Route to Link 2 Response Route (RPRte)Bits 118. The Response Route field defines the node or link to which a response packet is forwarded. Response packets are routed to only one destination and the table index is based on the destination Node ID. Bit [0] = Route to this node Bit [1] = Route to Link 0 Bit [2] = Route to Link 1 Bit [3] = Route to Link 2 Broadcast Route (BCRte)Bits 1916. The Broadcast Route field defines the node or link(s) to which a broadcast packet is forwarded. Broadcasts may be routed to more than one destination. The Node ID that indexes into the table is the source Node ID. Bit [0] = Route to this node Bit [1] = Route to Link 0 Bit [2] = Route to Link 1 Bit [3] = Route to Link 2
Chapter 4
51
32559
Rev. 3.08 July 2007
4.3.7
Node ID Register
The Node ID register contains node ID, node count and CPU core count information (a node can have more than one CPU core). It is expected that system configuration software will program the Node ID as part of coherent HyperTransport link configuration. Correct system operation depends on an assignment of distinct node ID values not exceeding NodeCnt[2:0]. For example, the Node IDs in a 4-node system must be 0, 1, 2, and 3; an example of an incorrect Node ID assignment in this system is 0, 1, 3, and 4. The Node ID will also be used as the initial APIC ID for the local APIC. Node ID Register
31 reserved 20 19 CpuCnt 16 15 14 reserved 12 11 10 reserved 8

7 reserved 6 4 3 reserved 2 0
LkNode
SbNode
NodeCnt
NodeId
Bits 3120 1916 15 1412 11 108 7 64 3 20
Mnemonic reserved CPUCnt reserved LkNode reserved SbNode reserved NodeCnt reserved NodeID
Function CPU Count Lock Controller Node ID HyperTransport I/O Hub Node ID Node Count This Node ID
R/W R R/W R R/W R R/W R R/W R R/W
Reset 0 0 0 0 0 0 0 0 0
Field Descriptions Node ID (NodeID)Bits 20. Defines the Node ID of this node. It resets to 0 for the boot strap processor (BSP) and to 7h for all other nodes. Node Count (NodeCnt)Bits 64. Specifies the number of coherent nodes in the system. Note that the hardware allows only values to be programmed into this field that are consistent with the multiprocessor capabilities of the device, as specified in Northbridge Capabilities Register on page 182. This field will not be updated if there is an attempt to write a value that is inconsistent with the specified multiprocessor capabilities. 000b = 1 node 001b = 2 nodes 010b = 3 nodes 011b = 4 nodes 101b = 6 nodes 52 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
111b
= 8 nodes
HyperTransport I/O Hub Node ID (SbNode)Bits 108. Defines the Node ID of the node that owns the HyperTransport link to the HyperTransport I/O hub. Lock Controller Node ID (LkNode)Bits 1412. Defines the Node ID of the node that contains the Lock Controller. The lock controller node (LkNode) must be the same as the HyperTransport I/O hub node (SbNode). CPU Count (CPUCnt)Bits 1916. Defines the number of CPU cores in the system. 0000b = 1 CPU ... 1111b = 16 CPUs
4.3.8
Unit ID Register
The Unit ID register specifies the Unit ID of each functional unit on the processor. Under normal operation, there is no need for software to change the default Unit ID assignments. This register also contains a pointer to the HyperTransport I/O hub link. Unit ID Register
31 reserved 10 9 SbLink

8 7 HbUnit 6 5 McUnit 4 3 C1Unit 2 1 C0Unit 0
Bits 3110 98 76 54 32 10
Mnemonic reserved SbLink HbUnit McUnit C1Unit C0Unit
Function HyperTransport I/O Hub Link ID Host Bridge Unit ID Memory Controller Unit ID CPU1 Unit ID CPU0 Unit ID
R/W R R/W R/W R/W R/W R/W
Reset 0 00b 11b 10b 01b 00b
Field Descriptions CPU0 Unit ID (C0Unit)Bits 10. Defines the Unit ID of CPU0 core. CPU1 Unit ID (C1Unit)Bits 32. Defines the Unit ID of CPU1 core, if it is implemented. Memory Controller Unit ID (McUnit)Bits 54. Defines the Unit ID of the memory controller. Host Bridge Unit ID (HbUnit)Bits 76. Defines the Unit ID of the host bridge. The host bridge is used to bridge between coherent and noncoherent HyperTransport technology domains.
Chapter 4
53
32559
Rev. 3.08 July 2007
HyperTransport I/O Hub Link ID (SbLink)Bits 98. Defines the link to which the HyperTransport I/O hub is connected. It is only used by the node which owns the HyperTransport I/O hub, as indicated in the Node ID register. 00b = LDT0 (default) 01b = LDT1 10b = LDT2 11b = Undefined
4.3.9
HyperTransport Transaction Control Register
The HyperTransport Transaction Control register configures the high-level HyperTransport protocols. It can be used to optimize system performance or change HyperTransport technology transaction protocols. HyperTransport Transaction Control Register
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 SeqIdSrcNodeEn DisIsocWrMedPri DisRmtPMemC DisIsocWrHiPri EnCpuRdHiPri ApicExtBrdCst MedPriBypCnt ChgIsocToOrd DsNpReqLmt RspPassPW InstallStateS ApicExtSpur HiPriBypCnt LimitCldtCfg DisWrLoPri 8 DisPMemC

7 CPURdRspPassPW 6 CPUReqPassPW 5 4 3 2 1 0
DisRdDwP
DisWrDwP
ApicExtId
BufRelPri
Bits 31 30 29 28 2726 2524 23 2221 20 19 18 17 16 15 1413 12 11 10
Mnemonic DisIsocWrMedPri DisIsocWrHiPri DisWrLoPri EnCpuRdHiPri HiPriBypCnt MedPriBypCnt InstallStateS DsNpReqLmt SeqIdSrcNodeEn ApicExtSpur ApicExtId ApicExtBrdCst LintEn LimitCldtCfg BufRelPri ChgIsocToOrd RspPassPW DisFillP
Function Disable medium priority isochronous writes Disable high priority isochronous writes Disable low priority writes Enable high priority CPU reads High-priority bypass count Medium-priority bypass count Read Block Cache Install State Shared Downstream non-posted request limit Sequence ID source node enable APIC extended spurious vector enable APIC extended ID enable APIC extended broadcast enable Local interrupt conversion enable Limit coherent HyperTransport configuration space range Buffer release priority select Change ISOC to Ordered Response PassPW Disable fill probe
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 11b 11b 0 00b 0 0 0 0 0 0 00b 0 0 0
54
Chapter 4
DisRdBP
DisWrBP
Cpu1En
DisMTS
DisFillP
LintEn
32559
Rev. 3.08
July 2007
Bits 9 8 7 6 5 4 3 2 1 0
Mnemonic DisRmtPMemC DisPMemC CPURdRspPassPW CPUReqPassPW Cpu1En DisMTS DisWrDwP DisWrBP DisRdDwP DisRdBP
Function Disable remote probe memory cancel Disable probe memory cancel CPU Read response PassPW CPU request PassPW CPU1 enable Disable memory controller target start Disable write doubleword probes Disable write byte probes Disable read doubleword probe Disable read byte probe
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0
Field Descriptions Disable Read Byte Probe (DisRdBP)Bit 0. This bit determines if probes are issued for CPUgenerated RdSized byte (must be 0 for multi-CPU core systems; recommended to be 1 for uniprocessor, single CPU core systems) 0 = Probes issued 1 = Probes not issued Disable Read Doubleword Probe (DisRdDwP)Bit 1. This bit determines if probes are issued for CPU-generated RdSized Doubleword (must be 0 for multi-CPU core systems; recommended to be 1 for uniprocessor, single CPU core systems) 0 = Probes issued 1 = Probes not issued Disable Write Byte Probes (DisWrBP)Bit 2. This bit determines if probes are issued for CPUgenerated WrSized byte (must be 0 for multi-CPU core systems; recommended to be 1 for uniprocessor, single CPU core systems) 0 = Probes issued 1 = Probes not issued Disable Write Doubleword Probes (DisWrDwP)Bit 3. This bit determines if probes are issued for CPU-generated WrSized Doubleword (must be 0 for multi-CPU core systems; recommended to be 1 for uniprocessor, single CPU core systems) 0 = Probes issued 1 = Probes not issued Disable Memory Controller Target Start (DisMTS)Bit 4. Disables use of TgtStart. TgtStart is used to improve scheduling of back-to-back ordered transactions by indicating when the first transaction is received and ordered at the memory controller. 0 = TgtStart packets are generated 1 = TgtStart packets are not generated
Chapter 4
55
32559
Rev. 3.08 July 2007
CPU1 Enable (Cpu1En)Bit 5. Enables a second CPU core, if it is implemented on the node. When this bit is set the second CPU core begins fetching code. This bit should only be set for nodes which have two CPU cores. 0 = Second CPU core disabled or not present 1 = Second CPU core enabled CPU Request PassPW (CPUReqPassPW)Bit 6. Allows CPU-generated requests to pass posted writes. 0 = CPU requests do not pass posted writes 1 = CPU requests pass posted writes CPU Read Response PassPW (CPURdRspPassPW)Bit 7. Allows RdResponses to CPUgenerated reads to pass posted writes. 0 = CPU responses do not pass posted writes 1 = CPU responses pass posted writes Disable Probe Memory Cancel (DisPMemC)Bit 8. Disables generation of the MemCancel command when a probe hits a dirty cache block. MemCancels are used to attempt to save DRAM and/or HyperTransport technology bandwidth associated with the transfer of stale DRAM data. 0 = Probes may generate MemCancels 1 = Probes may not generate MemCancels Disable Remote Probe Memory Cancel (DisRmtPMemC)Bit 9. Disables generation of the MemCancel command when a probe hits a dirty cache block unless the probed cache is on the same node as the memory controller. MemCancels are used to attempt to save DRAM and/or HyperTransport technology bandwidth associated with the transfer of stale DRAM data. 0 = Probes hitting dirty blocks generate memory cancel packets, regardless of the location of the probed cache 1 = Only probed caches on the same node as the target memory controller generate memory cancel packets Disable Fill Probe (DisFillP)Bit 10. Disables probes for CPU-generated fills (must be 0 for multiprocessor system; recommended to be 1 for uniprocessor system). 0 = Probes issued for cache fills 1 = Probes not issued for cache fills Response PassPW (RspPassPW)Bit 11. Causes the host bridge to set the PassPW bit in all downstream responses. This technically breaks the PCI ordering rules but it is not expected to be an issue in the downstream direction. Setting this bit improves the latency of upstream requests by allowing the downstream responses to pass posted writes. 0 = Downstream response PassPW based on original request 1 = Downstream response PassPW set to 1 56 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
Change ISOC to Ordered (ChgIsocToOrd)Bit 12. Causes bit 1 of RdSz/WrSz commands to be treated as an ordered bit in coherent HyperTransport technology devices, instead of as an isochronous bit. Setting this bit will disable prioritization of isochronous requests, but will enable tracking of ordered commands across coherent HyperTransport links. This bit should only be set if the performance of non-ordered peer-to-peer traffic across coherent HyperTransport links is optimized, or to disable isochronous prioritization. 0 = Bit 1 of coherent HyperTransport technology sized command used for isochronous prioritization 1 = Bit 1 of coherent HyperTransport technology sized command used for ordering Buffer Release Priority Select (BufRelPri)Bits 1413. Selects the number of HyperTransport technology doublewords sent with a buffer release pending before the buffer release is inserted into the command/data stream of a busy link. 00b = 64 (default) 01b = 16 10b = 8 11b = 2 This should be set to a value of 10 (8 HyperTransport packets) to maximize HyperTransport technology bandwidth. Limit Coherent HyperTransport Configuration Space Range (LimitCldtCfg)Bit 15. Limits the extent of the coherent HyperTransport technology configuration space based on the number of nodes in the system. When this bit is set, configuration accesses that normally map to coherent HyperTransport space will be sent to noncoherent HyperTransport links instead, if these accesses attempt to access a non-existent node, as specified through the node count in the Node ID register. This bit should be set by BIOS once coherent HyperTransport fabric initialization is complete. Failure to do so will result in PCI configuration accesses to nonexistent nodes being sent into the coherent HyperTransport routing fabric, causing the system to hang. 0 = No coherent HyperTransport configuration space restrictions 1 = Limit coherent HyperTransport configuration space based on number of nodes Local Interrupt Conversion Enable (LintEn)Bit 16. Enables the conversion of broadcast ExtInt/NMI HyperTransport technology interrupts to LINT0/1 before delivering to the local APIC. This conversion only takes place if the local APIC is hardware enabled. 0 = ExtInt/NMI interrupts unaffected 1 = ExtInt/NMI broadcast interrupts converted to LINT0/1 APIC Extended Broadcast Enable (ApicExtBrdCst)Bit 17. Enables extended APIC broadcast functionality. 0 = APIC broadcast is 0Fh 1 = APIC broadcast is FFh APIC Extended ID Enable (ApicExtId)Bit 18. Enables extended APIC ID functionality. Chapter 4 Memory System Configuration 57
32559
Rev. 3.08 July 2007
0 1
= APIC ID is 4 bits = APIC ID is 8 bits
APIC Extended Spurious Vector Enable (ApicExtSpur)Bit 19. Enables extended APIC spurious vector functionality. 0 = Lower 4 bits of spurious vector are read-only 1111b 1 = Lower 4 bits of spurious vector are writable Sequence ID Source Node Enable (SeqIdSrcNodeEn)Bit 20. Causes the source node of requests to be used in the SeqID field of downstream noncoherent HyperTransport technology request packets. This may be desirable for some debug HyperTransport packet tracing applications, to match downstream packets with their originating CPU. For normal operation, this bit should be cleared, since this use of SeqID for source node determination may cause problems with packet ordering. Downstream non-posted request limit (DsNpReqLmt)Bits 2221. Sets the number of downstream non-posted requests issued by CPU(s) which may be outstanding on the noncoherent HyperTransport links attached to this node. Non-posted requests from CPU(s) will be throttled such that they do not exceed the programmed value in this field. 00b = No limit 01b = Limited to 1 10b = Limited to 4 11b = Limited to 8 Read Block Cache Install State Shared (InstallStateS)Bit 23. This bit controls the state that a cache line is installed at when the cache performs a read block. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. 0 = Install state is dependent on probe results. 1 = Install state shared. Medium-Priority Bypass Count (MedPriBypCnt)Bits 2524. The maximum number of times a medium priority access can pass a low priority access before medium priority mode is disabled for one access. High-Priority Bypass Count (HiPriBypCnt)Bits 2726. The maximum number of times a high priority access can pass a medium or low priority access before high priority mode is disabled for one access. Enable High Priority CPU Reads (EnCpuRdHiPri)Bit 28. Enables CPU reads to be treated as high priority. CPU reads are treated as medium priority by default. Setting EnCpuRdHiPri changes their priority to high. Disable Low Priority Writes (DisWrLoPri)Bit 29. Disables non-isochronous writes from being treated as low priority. Non-isochronous writes are treated as low priority by default. Setting DisWrLoPri changes their priority to medium.
58
Chapter 4
32559
Rev. 3.08
July 2007
Disable High Priority Isochronous Writes (DisIsocWrHiPri)Bit 30. Disables isochronous writes from being treated as high priority. Isochronous writes are treated as high priority by default. Setting DisIsocWrHiPri changes their priority to medium (if DisIsocWrMedPri is clear) or low (if DisIsocWrMedPri is set). Disable Medium Priority Isochronous Writes (DisIsocWrMedPri)Bit 31. Disables isochronous writes from being treated as medium priority. Isochronous writes are treated as high priority by default. Setting DisIsocWrMedPri along with DisIsocWrHiPri changes their priority to low.
4.3.10
HyperTransport Initialization Control Register
The HyperTransport Initialization Control Register controls the routing and request generation that follows device initialization. This register also contains a set of scratchpad read/write bits that are cleared by different initialization conditions and that may be used to distinguish between various types of initialization events. Note that the Request Disable and Routing Table Disable bits must not be cleared until the routing tables have been initialized. HyperTransport Initialization Control Register
31 11 10 9 BiosRstDet[2:1]

8 reserved 7 6 5 BiosRstDet0 4 ColdRstDet 3 DefLnk 2 1 ReqDis 0 RouteTblDis
reserved
Bits 3111 109 87 6 5 4 32 1 0
Mnemonic reserved BiosRstDet[2:1] reserved InitDet BiosRstDet0 ColdRstDet DefLnk ReqDis RouteTblDis
Function BIOS Reset Detect[2:1]
R/W R R
InitDet R
Reset 0 0 0 0 0
INIT Detect BIOS Reset Detect[0] Cold Reset Detect Default Link Request Disable Routing Table Disable
R/W R/W R/W R/W R/W
Field Descriptions Routing Table Disable (RouteTblDis)Bit 0. This bit determines whether the routing tables are used or whether default configuration access routing is used. It resets to 1, thereby routing requests to the Configuration Special Register (CSR) block and routing responses based on DefLnk. Once the routing tables have been set up this bit should be cleared. 0 = Packets are routed according to the routing tables Chapter 4 Memory System Configuration 59
32559
Rev. 3.08 July 2007
= Packets are routed according to the default link field (DefLnk)
Request Disable (ReqDis)Bit 1. This bit determines if the node is allowed to generate request packets. It resets to 0 for the BSP and to 1 for all other processors. This bit should be cleared by system initialization firmware once the system has been initialized from the BSP. This bit is set by hardware and cleared by software. 0 = Request packets may be generated 1 = Request packets may not be generated Default Link (DefLnk)Bits 32. This field is used after a reset and before the routing tables are initialized. It is used by hardware to determine which link to route responses to and may be used by software as part of system connectivity discovery. The default link field is loaded each time an incoming request is received, with the link ID of the link on which the packet arrived. It is read-only from software. It is only used to route packets during initialization, when the Routing Table Disable bit is set to 1, and only one outstanding request is active in the system at a time. During this interval, the value in the Default Link field is used to route responses, so that responses are always sent out on the link from which the last request was received. The register is updated as soon as a request is received, so reads of this register from the fabric return the link ID of the link on which the read command arrived. Reads from the CPU on the local node cause this field to get loaded with 11b. 00b = LDT0 01b = LDT1 10b = LDT2 11b = CPU on same node Cold Reset Detect (ColdRstDet)Bit 4. This bit may be used to distinguish between a cold versus a warm reset event by setting the bit to 1 before an initialization event is generated. This bit is cleared by a cold reset but not by a warm reset. BIOS Reset Detect (BiosRstDet)Bits 109,5. These bits may be used to distinguish between a reset event generated by the BIOS versus a reset event generated for any other reason by setting one or more of the bits to 1 before initiating a BIOS-generated reset event. These bits are cleared by a cold reset but not by a warm reset. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. INIT Detect (InitDet)Bit 6. This bit may be used to distinguish between an INIT and a warm/cold reset by setting the bit to 1 before an initialization event is generated. This bit is cleared by a warm or cold reset but not by an INIT.
4.3.11
LDTi Capabilities Register
These registers define the capabilities of the LDT links. They also contain a capabilities pointer that points to the next item in the linked capabilities list.
60
Chapter 4
32559
Rev. 3.08
July 2007
LDT0, LDT1, LDT2 Capabilities Registers

31 29 28 27 26 25 24 23 22 DropOnUnInit InbndEocErr ActAsSlave ChainSide HostHide reserved 18 17 16 15 WarmReset DblEnded
Function 0: Offset 80h, A0h, C0h

8 7 0
CapType
DevNum
CapPtr
CapID
Bits 3129 28 27 26 25 24 23 2218 17 16 158 70
Mnemonic CapType DropOnUnInit InbndEocErr ActAsSlave reserved HostHide ChainSide DevNum DblEnded WarmReset CapPtr CapID
Function Capability Type Drop on Uninitialized Link Inbound End-of-Chain Error Act As Slave Host Hide Chain Side Device Number Double Ended Warm Reset Capability Pointer Capability ID (Always Reads 08h)
R/W R R R R R R R R R R R R
Reset 001b 0 0 0 0 1 0 0 0 1 08h
Field Descriptions Capability ID (CapID)Bits 70. A capability ID of 08h indicates HyperTransport technology capability. Capability Pointer (CapPtr)Bits 158. Contains a pointer to the next capability in the list. Depending on the specific HyperTransport link configuration physically present on the processor, this will be either A0h (LDT1), C0h (LDT2), or 00h, if it is the last one. Warm Reset (WarmReset)Bit 16. Allows a reset sequence initiated by the HyperTransport technology control register to be either warm or cold. Since resets initiated through the HyperTransport control register are not supported, this field is read-only 1. Double Ended (DblEnded)Bit 17. Indicates that there is another bridge at the far end of the noncoherent HyperTransport chain. Since double-hosted chains are not supported, this field is read-only 0. Device Number (DevNum)Bits 2218. The device number of configuration accesses which Northbridge responds to when accessed from the noncoherent HyperTransport technology chain. Since double-hosted chains are not supported, this field is read-only 0. Chain Side (ChainSide)Bit 23. This bit indicates which side of the host bridge is being accessed. Since double-hosted chains are not supported, this bit is read-only 0.
Chapter 4
61
32559
Rev. 3.08 July 2007
Host Hide (HostHide)Bit 24. This bit causes the memory system configuration space to be inaccessible from the noncoherent HyperTransport technology chain. Since double-hosted chains are not supported, this bit is read-only 1. Act As Slave (ActAsSlave)Bit 26. Since this function is not currently implemented, this bit is read-only 0. Inbound End-of-Chain Error (InbndEocErr)Bit 27. Since this function is not currently implemented, this bit is read-only 0. Drop on Uninitialized Link (DropOnUnInit)Bit 28. Since this function is not currently implemented, this bit is read-only 0. Capability Type (CapType)Bits 3129. The code 001b designates a host interface HyperTransport technology capability.
4.3.12
LDTi Link Control Registers
These registers control LDT link operation and log link errors. LDT0, LDT1, LDT2 Link Control Registers
31 30 DwFcOutEn WidthOut 28 27 26 DwFcInEn 24 23 22 MaxWidthOut DwFcOut 20 19 18 MaxWidthIn LdtStopTriClkOvr

8 7 TransOff 6 5 InitComplete 4 3 CrcForceErr 2 CrcStartTest 1 CrcFloodEn 0 reserved
16 15 14 13 12 11 10 9 LdtStopRcvDis LdtStopTriEn reserved
Bits 31 3028 27 2624 23 2220 19 1816 15 14 13 12 11 10 98 7
Mnemonic DwFcOutEn WidthOut DwFcInEn WidthIn DwFcOut MaxWidthOut DwFcIn MaxWidthIn LdtStopTriClkOvr ExtCTL LdtStopTriEn IsocEn LdtStopRcvDis reserved CrcErr TransOff
Function Doubleword Flow Control Out Enable Link Width Out Doubleword Flow Control In Enable Link Width In Doubleword Flow Control Out Max Link Width Out Doubleword Flow Control In Max Link Width In HyperTransport Stop Tristate Clock Override Extended CTL Time HyperTransport Stop Tristate Enable Isochronous Enable HyperTransport Stop Receiver Disable CRC_Error Transmitter Off
R/W R R/W R R/W R R R R R/W R/W R/W R R/W R R/W R/W
Reset 0 0 0 0 0 001b 0 001b 0 0 0 0 0 0 0 0
62
LinkFail
DwFcIn
WidthIn
ExtCTL
IsocEn
RcvOff
CrcErr
Chapter 4
32559
Rev. 3.08
July 2007
Bits 6 5 4 3 2 1 0
Mnemonic RcvOff InitComplete LinkFail CrcForceErr CrcStartTest CrcFloodEn reserved
Function Receiver Off Initialization Complete Link Failure CRC Force Error CRC Start Test CRC Flood Enable
R/W R/W R R/W R/W R R/W R
Reset 0 0 0 0 0 0 0
Field Descriptions CRC Flood Enable (CrcFloodEn)Bit 1. If cleared to 0, this bit prevents CRC errors both from generating sync packets and causing the system to come down, and from setting the LinkFail bit. However, CRC checking logic still runs on all lanes enabled by LinkWidthIn, and detected errors still set the CRC Error bits to 1. Its reset value is 0. 0 = Do not generate sync packets on CRC error (default) 1 = Generate sync packets on CRC error CRC Start Test (CrcStartTest)Bit 2. Since this function is not currently implemented, this bit is read-only 0. CRC Force Error (CrcForceErr)Bit 3. When this bit is set, a bad CRC is generated on all transmitting lanes as enabled by LinkWidthOut. The covered data is not affected. This bit is readable and writable by software and resets to 0. 0 = Do not generate a bad CRC (default) 1 = Generate a bad CRC Link Failure (LinkFail)Bit 4. The LinkFail bit is set to 1 when a CRC error or a sync packet is detected after link initialization or if the link is not connected. See Machine Check Architecture (MCA) Registers on page 132 for more details. This bit is maintained through a warm reset and is cleared to 0 on a cold reset. LinkFail is set to 1 by hardware in the event of a link error that results in a sync flood. If this bit is set after a warm reset, BIOS must attempt to clear the bit by writing a 1 to determine if the link is not connected or if a sync flood occurred. It can be cleared by software by writing a 1. This bit becomes valid after Function 0, Offset 98h, B8h, or D8h, LinkConPend bit is cleared by hardware. 0 = No link failure detected 1 = Link failure detected Initialization Complete (InitComplete)Bit 5. This read-only bit is reset to 0, and set to 1 by hardware when low-level link initialization is successfully complete. If there is no device on the other end of the link, or if that device is unable to properly perform the low-level link initialization protocol, the bit is not set to 1. Software must not attempt to generate any packets across the link until this bit is set to 1. CRC checking in the hardware does not begin until initialization is complete. 0 = Initialization not complete Chapter 4 Memory System Configuration 63
32559
Rev. 3.08 July 2007
= Initialization is complete
Receiver Off (RcvOff)Bit 6. This bit disables the receiver from accepting any further packets. This bit resets to 0, and is set by software writing a 1 to the bit. This bit cannot be cleared by softwarea write of 0 to this bit position has no effect. If the HyperTransport link is not connected, then this bit is set by hardware. 0 = Receiver on 1 = Receiver off Transmitter Off (TransOff)Bit 7. This bit provides a mechanism to shut off a link transmitter for power savings or EMI reduction. When set to 1, no output signals toggle on the link. This bit resets to 0, and is set by software writing a 1 to the bit. This bit cannot be cleared by softwarea write of 0 to this bit position has no effect. If the HyperTransport link is not connected, then this bit is set by hardware. 0 = Transmitter on 1 = Transmitter off CRC_Error (CrcErr)Bits 98. These bits are set to 1 by hardware when a CRC error is detected on an incoming link. Errors are detected and reported on a per byte lane basis where bit 8 corresponds to the least significant byte lane. Two bits are required to cover the maximum HyperTransport link width of 16 bits. Error bits for unimplemented (as specified by MaxWidthIn) or unused byte lanes return 0 when read. These bits are maintained through a warm reset and is cleared to 0 on a cold reset. They are readable from software and are individually cleared to 0 by software by writing a 1. 00b = No error Bit [0] =1 Error on Byte Lane 0 Bit [1] =1 Error on Byte Lane 1 HyperTransport Stop Receiver Disable (LdtStopRcvDis)Bit 11. This bit enables power savings by controlling the state of the HyperTransport receiver. When this bit is set for any HyperTransport link, the HyperTransport receiver is disabled when LDTSTOP_L is asserted. Isochronous Enable (IsocEn)Bit 12. Since this function is not currently implemented, this bit is read-only 0. HyperTransport Stop Tristate Enable (LdtStopTriEn)Bit 13. This bit controls whether the transmitter tristates the link during the disconnected state of an LDTSTOP_L sequence. When the bit is set the transmitter tristates the link. When the bit is clear, it continues to drive the link. When this bit is set and LdtStopTriClkOvr is clear, the delay specified in the ReConDel (Function 3, Offset D4h) field is applied before restarting the link. The bit is cleared by a cold reset and its value is maintained through a warm reset. 0 = Driven during disconnect of an LDTSTOP_L 1 = Tristated during disconnect of an LDTSTOP_L Extended CTL Time (ExtCTL)Bit 14. When this bit is set, CTL will be asserted for 50s instead of 16 bit times during the link initialization sequence. This bit should be set if extended CTL 64 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
is required by the device at the other end of the HyperTransport link, as indicated by the Extended CTL Required bit in the Feature Capability register. The bit is cleared by cold reset and its value is maintained through warm reset. HyperTransport Stop Tristate Clock Override (LdtStopTriClkOvr)Bit 15. This bit prevents the HyperTransport clock from being tristated when LDTSTOP_L is asserted and LdtStopTriEn is set. Max Link Width In (MaxWidthIn)Bits 1816. This field contains three bits that indicate the physical width of the incoming side of the HyperTransport link implemented by this device. For AMD NPT Family 0Fh Processors, this field is set to 000 or 001 to indicate an 8-bit or 16bit link. Doubleword Flow Control In (DwFcIn)Bit 19. This bit is read-only 0 to indicate that this link does not support doubleword flow control. Max Link Width Out (MaxWidthOut)Bits 2220. This field contains three bits that indicate the physical width of the outgoing side of the HyperTransport link implemented by this device. It is read-only. For AMD NPT Family 0Fh Processors, this field is set to 000 or 001 to indicate an 8-bit or 16-bit link. Doubleword Flow Control Out (DwFcOut)Bit 23. This bit is read-only 0 to indicate that this link does not support doubleword flow control. Link Width In (WidthIn)Bits 2624. This field is similar to the WidthOut field, except that it controls the utilized width of the incoming side of the links implemented by this device. The hardware will only allow values to be programmed into this field which are consistent with the width capabilities of the link as specified by MaxWidthIn. Attempts to write values inconsistent with the capabilities will result in this field not being updated. Doubleword Flow Control In Enable (DwFcInEn)Bit 27. Since this function is not currently implemented, this bit is read-only 0. Link Width Out (WidthOut)Bits 3028. This field controls the utilized width (which may not exceed the physical width) of the outgoing side of the HyperTransport link. Software can read/write this field, and its value is maintained through a warm reset. After cold reset, this field is initialized by hardware based on the results of the link width negotiation sequence described in the HyperTransport technology specification. Based on sizing the devices at both ends of the link, software may then write a different value into the register. The chain must pass through warm reset or a LDTSTOP_L disconnect sequence for the new width values to be reflected on the link. The WidthOut CSR in the link transmitter must match the WidthIn CSR in the link receiver of the device on the other side of the link. The LinkWidthIn and LinkWidthOut registers within the same device are not required to have matching values.
Chapter 4
65
32559
Rev. 3.08 July 2007
The hardware will only allow values to be programmed into this field which are consistent with the width capabilities of the link as specified by MaxWidthOut. Attempts to write values inconsistent with the capabilities will result in this field not being updated 000b = 8-bit 001b = 16-bit 010b = reserved 011b = 32-bit 100b = 2-bit 101b = 4-bit 110b = reserved 111b = Link physically not connected Doubleword Flow Control Out Enable (DwFcOutEn)Bit 31. Since this function is not currently implemented, this bit is read-only 0.
4.3.13
LDTi Frequency/Revision Registers
These registers control the link frequency, specify the link frequency capabilities, and describe the HyperTransport technology specification level to which the specific link conforms. LDT0, LDT1, LDT2 Frequency/Revision Registers
31 LnkFreqCap 16 15 Error

12 11 Freq 8 7 5 4 MinRev 0
MajRev
Bits 3116 1512 118 75 40
Mnemonic LnkFreqCap Error Freq MajRev MinRev
Function Link Frequency Capability Error Link Frequency Major Revision Minor Revision
R/W R R R/W R R
Reset 0 0 001b 00010b
Field Descriptions Minor Revision (MinRev)Bits 40. Contains the minor revision of the HyperTransport technology specification to which the device conforms. Major Revision (MajRev)Bits 75. Contains the major revision of the HyperTransport technology specification level to which the device conforms. Link Frequency (Freq)Bits 118. Specifies the maximum operating frequency of the link's transmitter clock. The encoding of this field is shown below.
66
Chapter 4
32559
Rev. 3.08
July 2007
The link frequency field (Freq) is cleared by cold reset. Software can write a nonzero value to this register, and that value takes effect on the next warm reset or LDTSTOP_L disconnect sequence. The hardware will only allow values to be programmed into this field which are consistent with the frequency capabilities of the link as specified by LinkFreqCap. Attempts to write values inconsistent with the capabilities will result in this field not being updated. It is possible to program this field for a higher frequency than the maximum allowed by the processor. Refer to the processor data sheet for the maximum operating frequency allowed for a given processor implementation. See 3.6.1.2 HyperTransport Technology Transfer Speeds on page 31 for processor specific link frequency restrictions. 0000b = 200 MHz 0001b = reserved 0010b = 400 MHz 0011b = reserved 0100b = 600 MHz 0101b = 800 MHz 0110b = 1000 MHz 0111b = reserved 10001110b = reserved 1111b = 100 MHz Error (Error)Bits 1512. This function is not currently implemented. Link Frequency Capability (LnkFreqCap)Bits 3116. Indicates the clock frequency capabilities of the link. Each bit corresponds to one of the 16 possible link frequency encodings. While the frequency capabilities of different AMD NPT Family 0Fh Processors may vary, the 200-MHz and 100-MHz frequencies are supported by the design.
4.3.14
LDTi Feature Capability Registers
These registers identify features that are supported by the specific link. LDT0, LDT1, LDT2 Feature Capability Registers
31
Function 0: Offset 8Ch, ACh, CCh

9 8 ExtRegSet 7 4 3 ExtCTLRqd 2 CrcTstMode 1 LdtStopMode 0 IsocMode
reserved
reserved
Bits 319 8
Mnemonic reserved ExtRegSet
Function Extended Register Set
R/W R R
Reset 0 0
Chapter 4
67
32559
Rev. 3.08 July 2007
Bits 74 3 2 1 0
Mnemonic reserved ExtCTLRqd CrcTstMode LdtStopMode IsocMode
Function Extended CTL Required CRC Test Mode HyperTransport Stop Mode Isochronous Flow Control Mode
R/W R R R R R
Reset 0 0 0 1 0
Field Descriptions Isochronous Flow Control Mode (IsocMode)Bit 0. This is a read-only bit that reflects whether isochronous flow control operation is supported. This bit is set to 0, indicating no support for this feature. HyperTransport Stop Mode (LdtStopMode)Bit 1. This is a read-only bit that indicates whether the LDTSTOP_L protocol is supported. This bit is set to 1, indicating support for this feature. CRC Test Mode (CrcTstMode)Bit 2. This is a read-only bit that indicates whether CRC test mode is supported. This bit is set to 0, indicating no support for this feature. Extended CTL Required (ExtCTLRqd)Bit 3. This is a read-only bit that indicates whether extended CTL is required. This bit is set to 0, indicating that extended CTL is not required. Extended Register Set (ExtRegSet)Bit 8. This is a read-only bit that indicates whether the Enumeration Scratchpad, Error Handling and Memory Base/Limit Upper registers are supported. This bit is set to 0, indicating no support for this feature.
4.3.15
LDTi Buffer Count Registers
These registers specify the number of command and data buffers for each virtual channel available for use by the transmitter at the other end of the specific HyperTransport technology link. See XBAR Flow Control Buffers on page 156 for more information on command and data buffers. Note: The reset values for each of the LDTn Buffer Count registers depend on the link connection type (coherent HyperTransport or noncoherent HyperTransport technology). Because hardware attempts to choose optimal settings, this register should not, in general, need to be changed. LDT0, LDT1, LDT2 Buffer Count Registers
31 reserved 27 26 24 23 22 reserved 20 19 18 reserved 16 15 Probe
Function 0: Offset 90h, B0h, D0h

12 11 Rsp 8 7 PReq 4 3 Req 0
RspD
PReqD
ReqD
68
Chapter 4
32559
Rev. 3.08
July 2007
Bits 3127 2624 23 2220 19 1816 1512 118 74 30
Mnemonic reserved RspD reserved PReqD reserved ReqD Probe Rsp PReq Req
Function
R/W R
Coherent Link Reset 0 100b 0 001b 0 011b 5h 6h 1h 3h
Noncoherent Link Reset 0 010b 0 101b 0 001b 0h 4h 5h 6h
Response Data Buffer Count Posted Request Data Buffer Count Request Data Buffer Count Probe Buffer Count Response Buffer Count Posted Request Buffer Count Request Buffer Count
R/W R R/W R R/W R/W R/W R/W R/W
Field Descriptions Request Buffer Count (Req)Bits 30. Defines the number of request command buffers available for use by the transmitter at the other end of the link. See the register layout for the default reset settings for coherent and noncoherent HyperTransport technology links. Posted Request Buffer Count (PReq)Bits 74. Defines the number of Posted request command buffers available for use by the transmitter at the other end of the link. See the register layout for the default reset settings for coherent and noncoherent HyperTransport technology links. Response Buffer Count (Rsp)Bits 118. Defines the number of response buffers available for use by the transmitter at the other end of the link. See the register layout for the default reset settings for coherent and noncoherent HyperTransport technology links. Probe Buffer Count (Probe)Bits 1512. Defines the number of probe buffers available for use by the transmitter at the other end of the link. This field must be 0 for a noncoherent HyperTransport link. See the register layout for the default reset settings for coherent and noncoherent HyperTransport technology links. Request Data Buffer Count (ReqD)Bits 1816. Defines the number of request data buffers available for use by the transmitter at the other end of the link. See the register layout for the default reset settings for coherent and noncoherent HyperTransport technology links. Posted Request Data Buffer Count (PReqD)Bits 2220. Defines the number of posted request data buffers available for use by the transmitter at the other end of the link. See the register layout for the default reset settings for coherent and noncoherent HyperTransport technology links. Response Data Buffer Count (RspD)Bits 2624. Defines the number of response data buffers available for use by the transmitter at the other end of the link. See the register layout for the default reset settings for coherent and noncoherent HyperTransport technology links.
Chapter 4
69
32559
Rev. 3.08 July 2007
4.3.16
LDTi Bus Number Registers
If the specific link (LDT0, LDT1, or LDT2) is noncoherent HyperTransport technology, this register specifies the bus numbers downstream (behind) the host bridge. If the link is coherent HyperTransport technology, this register has no meaning. LDT0, LDT1, LDT2 Bus Number Registers
31 reserved 24 23 SubBusNum 16 15 SecBusNum

8 7 PriBusNum 0
Bits 3124 2316 158 70
Mnemonic reserved SubBusNum SecBusNum PriBusNum
Function Subordinate Bus Number Secondary Bus Number Primary Bus Number
R/W R R/W R/W R
Reset 0 0 0 0
Field Descriptions Primary Bus Number (PriBusNum)Bits 70. Defines the primary bus number. Because the primary bus is the coherent HyperTransport technology fabric, this field always reads 0. Secondary Bus Number (SecBusNum)Bits 158. Defines the secondary bus number. The Secondary Bus Number register is used to record the bus number of the bus segment to which the secondary interface of the host bridge is connected. Configuration software programs the value in this register. If this link contains the HyperTransport I/O hub, the secondary bus number must be programmed to 0. Subordinate Bus Number (SubBusNum)Bits 2316. Defines the subordinate bus number. The Subordinate Bus Number register is used to record the bus number of the highest numbered bus segment that is behind (or subordinate to) the host bridge. Configuration software programs the value in this register.
4.3.17
LDTi Type Registers
These registers designate the type of HyperTransport link attached to the specific link. The bits are set by hardware after link initialization.
70
Chapter 4
32559
Rev. 3.08
July 2007
LDT0, LDT1, LDT2 Type Registers

31

5 4 LinkConPend 3 UniP-cLDT 2 1 InitComplete 0 LinkCon
reserved
Bits 315 4 3 2 1 0
Mnemonic reserved LinkConPend UniP-cLDT NC InitComplete LinkCon
Function Link Connect Pending UniP-cLDT Non Coherent Initialization Complete Link Connected
R/W R R R R R R
Reset 0
Field Descriptions Link Connected (LinkCon)Bit 0. Indicates that the link is connected. It is valid once the LinkConPend bit is clear. 0 = Not connected 1 = Connected Initialization Complete (InitComplete)Bit 1. Set to 1 to indicate that the initialization of the link has completed. (It is a duplicate of Bit 5 in HyperTransport Link Control). The NC and UniPcLDT bits are invalid until link initialization is complete. 0 = Initialization not complete 1 = Initialization is complete. Non Coherent (NC)Bit 2. Defines the link type (coherent versus noncoherent). 0 = Coherent HyperTransport technology 1 = Noncoherent HyperTransport technology UniP-cLDT (UniP-cLDT)Bit 3. Further qualifies the NC link type bit. A 1 indicates that this link is a uniprocessor coherent HyperTransport link connected to an external Northbridge. 0 = Normal coherent HyperTransport or noncoherent HyperTransport link 1 = Uniprocessor coherent HyperTransport link to external Northbridge Link Connect Pending (LinkConPend)Bit 4. Qualifies the LinkCon bit and is set to 1 when hardware is attempting to determine whether a link is connected or not. 0 = Link connection determination complete 1 = Link connection still being determined
Chapter 4
NC
71
32559
Rev. 3.08 July 2007
4.4
Function 1Address Map
The address map defines the address spaces assigned to DRAM, memory-mapped I/O, PCI I/O, and configuration accesses and also specifies destination information for each to facilitate routing of each access to the appropriate target. Table 6 lists each Function 1 configuration register. Table 6.
Offset 00h 04h 08h

Register Name Device ID Status Base Class Code BIST
1
Reset Vendor ID (AMD) Command 1101_1022h 0000_0000h Revision ID Cache Line Size 0600_0000h
Access RO RO RO
Description page 47 page 74
Subclass Code Header Type
Program Interface Latency Timer
0Ch 10h 14h 18h 1Ch 20h 24h 28h 2Ch 30h 34h 38h 3Ch 40h 44h 48h 4Ch 50h 54h 58h 5Ch
0080_0000 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h
RO RO RO RO RO RO RO RO RO RO RO RO RO RW RW RW RW RW RW RW RW
page 75
Base Address 0 Base Address 1 Base Address 2 Base Address 3 Base Address 4 Base Address 5 Card Bus CIS Pointer Sub-System ID ROM Base Address Capabilities reserved Max Latency Min GNT Int Pin Int Line Sub-System Vendor ID
0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h
DRAM Base 0 DRAM Limit 0 DRAM Base 1 DRAM Limit 1 DRAM Base 2 DRAM Limit 2 DRAM Base 3 DRAM Limit 3
page 77 page 78 page 77 page 78 page 77 page 78 page 77 page 78
72
Chapter 4
32559
Rev. 3.08
July 2007
Table 6.
Offset 60h 64h 68h 6Ch 70h 74h 78h 7Ch 80h 84h 88h 8Ch 90h 94h 98h 9Ch A0h A4h A8h ACh B0h B4h B8h BCh C0h C4h C8h CCh D0h D4h D8h DCh

Register Name DRAM Base 4 DRAM Limit 4 DRAM Base 5 DRAM Limit 5 DRAM Base 6 DRAM Limit 6 DRAM Base 7 DRAM Limit 7 Memory-Mapped I/O Base 0 Memory-Mapped I/O Limit 0 Memory-Mapped I/O Base 1 Memory-Mapped I/O Limit 1 Memory-Mapped I/O Base 2 Memory-Mapped I/O Limit 2 Memory-Mapped I/O Base 3 Memory-Mapped I/O Limit 3 Memory-Mapped I/O Base 4 Memory-Mapped I/O Limit 4 Memory-Mapped I/O Base 5 Memory-Mapped I/O Limit 5 Memory-Mapped I/O Base 6 Memory-Mapped I/O Limit 6 Memory-Mapped I/O Base 7 Memory-Mapped I/O Limit 7 PCI I/O Base 0 PCI I/O Limit 0 PCI I/O Base 1 PCI I/O Limit 1 PCI I/O Base 2 PCI I/O Limit 2 PCI I/O Base 3 PCI I/O Limit 3 Reset Access RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW RW Description page 77 page 78 page 77 page 78 page 77 page 78 page 77 page 78 page 80 page 81 page 80 page 81 page 80 page 81 page 80 page 81 page 80 page 81 page 80 page 81 page 80 page 81 page 80 page 81 page 82 page 83 page 82 page 83 page 82 page 83 page 82 page 83
Chapter 4
73
32559
Rev. 3.08 July 2007
Table 6.
Offset E0h E4h E8h ECh F0h

Register Name Configuration Base and Limit 0 Configuration Base and Limit 1 Configuration Base and Limit 2 Configuration Base and Limit 3 DRAM Hole Address 0000_0000 Reset Access RW RW RW RW R/W Description page 84 page 84 page 84 page 84 page 85
4.4.1
31 Device ID 16 15 Vendor ID

0
Bits 3116 150
R/W R R
Reset 1101h 1022h
4.4.2
This register specifies the class code and revision for the Function 1 registers and is part of the standard PCI configuration header.
74
Chapter 4
32559
Rev. 3.08
July 2007

31 Base Class Code 24 23 Subclass Code 16 15 Programming Interface

8 7 Revision ID 0
Bits 3124 2316 158 70
R/W R R R R
Field Descriptions Revision ID (RevID)Bits 70. Programming Interface (PI)Bits 158. This read-only value is defined as 00h. Sub Class Code (SCC)Bits 2316. This read-only value is defined as 00h. Base Class Code (BCC)Bits 3124. This read-only value is defined as 06h for a host bridge device.
4.4.3

8 7 CLS 0
Bits 3124 2316 158 70
R/W R R R R
Field Descriptions CacheLineSize (CLS)Bits 70. This read-only value is defined as 00h. LatencyTimer (LatTimer)Bits 158. This read-only value is defined as 00h.
Chapter 4
75
32559
Rev. 3.08 July 2007
HeaderType (HType)Bits 2316. This read-only value is defined as 80h to indicate that multiple functions are present in the configuration header and that the header layout corresponds to a device header as opposed to a bridge header. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. BISTBits 3124. This read-only value is defined as 00h.
4.4.4
DRAM Address Map
These registers define sections of the memory address map for which accesses should be routed to DRAM. DRAM regions must not overlap each other. For addresses within the specified range of a base/limit pair, requests are routed to the memory controller on the node specified by the destination Node ID. System addresses are considered to be within the defined range if they are greater than or equal to the base and less than or equal to the limit. For the purposes of this comparison, the lower unspecified bits of the base are assumed to be 0s and the lower unspecified bits of the limit are assumed to be 1s. An address that maps to both DRAM and memory-mapped I/O will be routed to MMIO. Programming of the DRAM address maps must be consistent with the Top Of Memory and Memory Type Range registers (see Chapter 14, Processor Configuration Registers). Accesses from the CPU can only access the DRAM address maps if the corresponding CPU memory type is DRAM. For accesses from I/O devices, the lookup is based on address only. Each base/limit set of DRAM address maps is associated with a particular Node ID (see Node ID Register on page 52). The DRAM Base/Limit 0 registers specify the DRAM attached to node 0. Similarly, DRAM on nodes 17 is described by base/limit registers 17. Note that the destination NodeId field must still be written to ensure correct operation. When node interleaving is enabled, each nodes DRAM limit must be set to the Top Of Memory and each nodes DRAM base must be set to 0. The node to which an address is routed to when nodes are interleaved is defined by IntlvEn (Function 1, Offset 40h, 48h, etc.) and IntlvSel (Function 1, Offset 44h, 4Ch, etc.). Each node must be configured with the same amount of DRAM when node interleaving is enabled. Address routed to the DRAM controller (InputAddr) is calculated from the system address (SysAddr) in the following way: DramAddr[39:0] = {SysAddr[39:24] - DRAMBase[39:24], SysAddr[23:0]}, InputAddr[35:0] = {DramAddr[35:12], DramAddr[11:0]} when node memory is not interleaved, InputAddr[35:0] = {DramAddr[36:13], DramAddr[11:0]} when 2 nodes are interleaved, InputAddr[35:0] = {DramAddr[37:14], DramAddr[11:0]} when 4 nodes are interleaved, InputAddr[35:0] = {DramAddr[38:15], DramAddr[11:0]} when 8 nodes are interleaved.
76
Chapter 4
32559
Rev. 3.08
July 2007
4.4.4.1
DRAM Base i Registers Function 1: Offset 40h, 48h, 50h, 58h, 60h, 68h, 70h, 78h
16 15 DRAMBasei reserved 11 10 8 7 reserved 2 1 WE 0 RE
DRAM Base 07 Registers

31
IntlvEn
Bits 3116 1511 108 72 1 0
Mnemonic DRAMBasei reserved IntlvEn reserved WE RE
Function DRAM Base Address i Bits 3924 Interleave Enable Write Enable Read Enable
R/W R/W R R/W R R/W R/W
Reset X 0 X 0 0 0
X in the Reset column indicates that the field initializes to an undefined state after reset.
Field Descriptions Read Enable (RE)Bit 0. This bit enables reads to the defined address space. 0 = Disabled 1 = Enabled Write Enable (WE)Bit 1. This bit enables writes to the defined address space. 0 = Disabled 1 = Enabled Interleave Enable (IntlvEn)Bits 108. This field enables interleaving on a 4-Kbyte boundary between memory on different nodes. The bits are encoded as follows: 000b = No interleave 001b = Interleave on A[12] (2 nodes) 010b = reserved 011b = Interleave on A[12] and A[13] (4 nodes) 100b = reserved 101b = reserved 110b = reserved 111b = Interleave on A[12] and A[13] and A[14] (8 nodes) DRAM Base Address i Bits 3924 (DRAMBasei)Bits 3116. This field defines the upper address bits of a 40-bit address that defines the start of DRAM region i (where i = 0, 1, . . . 7).
Chapter 4
77
32559
Rev. 3.08 July 2007
4.4.4.2
DRAM Limit i Registers
DRAM Limit 07 Registers Function 1: Offset 44h, 4Ch, 54h, 5Ch, 64h, 6Ch, 74h, 7Ch
31 DRAMLimiti 16 15 reserved 11 10 8 7 reserved 3 2 0
IntlvSel
DstNode
Bits 3116 1511 108 73 20
Mnemonic DRAMLimiti reserved IntlvSel reserved DstNode
Function DRAM Limit Address i (3924) Interleave Select Destination Node ID
R/W R/W R R/W R R/W
Reset X 0 X 0 X
Field Descriptions Destination Node ID (DstNode)Bits 20. This field specifies the node that a packet is sent to if it is within the address range. The value of this field must correspond to the number of the register pair, and it should be set as follows: 000b for Function 1, Offset 44h register 001b for Function 1, Offset 4Ch register ... 111b for Function 1, Offset 7Ch register Interleave Select (IntlvSel)Bits 108. This field specifies the values of address bits A[14:12] to use with the Interleave Enable field (IntlvEn[2:0]) to determine which 4-Kbyte blocks are routed to this region (See Figure 3). IntlvSel[0] corresponds to A[12] IntlvSel[1] corresponds to A[13] IntlvSel[2] corresponds to A[14] DRAM Limit Address i (3924) (DRAMLimiti)Bits 3116. This field defines the upper address bits of a 40-bit address that defines the end of DRAM region n (where i = 0,1, . . . 7).
78
Chapter 4
32559
Rev. 3.08
July 2007
Node 0 IntlvEn[2:0] = IntlvSel[2:0] = A[13:12] =
011 000 00
011 001 01
011 010 10
011 011 11
Figure 3.
Interleave Example (IntlvEn Relation to IntlvSel)
4.4.5
Memory-Mapped I/O Address Map Registers
These registers define sections of the memory address map for which accesses should be routed to memory-mapped I/O. MMIO regions must not overlap each other. For addresses within the specified range of a base/limit pair, requests are routed to the noncoherent HyperTransport link specified by the destination Node ID and destination Link ID. Addresses are considered to be within the defined range if they are greater than or equal to the base and less than or equal to the limit. For the purposes of this comparison, the lower unspecified bits of the base are assumed to be 0s and the lower unspecified bits of the limit are assumed to be 1s. An address that maps to both DRAM and memory-mapped I/O is routed to MMIO. Programming of the MMIO address maps must be consistent with the Top Of Memory and Memory Type Range registers (see Chapter 14, Processor Configuration Registers). In particular, accesses from the CPU can only hit in the MMIO address maps if the corresponding CPU memory type is of type I/O. For accesses from I/O devices, the lookup is based on address only. 4.4.5.1 Extended Configuration Space Access
Chipset devices may support PCI-defined extended configuration space through an MMIO range. Typically, requests to the MMIO range for extended configuration space are required to use the nonposted channel. Therefore, the Non-Posted bit should be set for this MMIO range. Instructions used to read extended configuration space must be of the following form: mov EAX/AX/AL, <any_address_mode>; Instructions used to write extended configuration space must be of the following form: mov <any_address_mode>, EAX/AX/AL;
Chapter 4
79
32559
Rev. 3.08 July 2007
In addition, all such accesses are required not to cross any naturally aligned doubleword boundary. Access to extended configuration registers that do not meet these requirements result in undefined behavior. Extended configuration space is normally specified to be the uncacheable memory type. 4.4.5.2 Memory-Mapped I/O Base i Registers Function 1: Offset 80h, 88h, 90h, 98h, A0h, A8h, B0h, B8h
8 MMIOBasei 7 reserved 4 3 Lock 2 CpuDis 1 WE 0 RE
Memory-Mapped I/O Base 07 Registers

31
Bits 318 74 3 2 1 0
Mnemonic MMIOBasei reserved Lock CpuDis WE RE
Function Memory-Mapped I/O Base Address i (3916) Lock CPU Disable Write Enable Read Enable
R/W R/W R R/W R/W R/W R/W
Reset X 0 X X 0 0
Field Descriptions Read Enable (RE)Bit 0. This bit enables reads to the defined address space. 0 = Disabled 1 = Enabled Write Enable (WE)Bit 1. This bit enables writes to the defined address space. 0 = Disabled 1 = Enabled CPU Disable (CpuDis)Bit 2. When set, this bit causes this MMIO range to apply to I/O requests only, not to CPU requests. Lock (Lock)Bit 3. Setting this bit, along with either the WE or RE bits, makes the base/limit registers read-only. Memory-Mapped I/O Base Address i (3916) (MMIOBasei)Bit 318. This field defines the upper address bits of a 40-bit address that defines the start of memory-mapped I/O region n (where n = 0,1, . . . 7).
80
Chapter 4
32559
Rev. 3.08
July 2007
4.4.5.3
Memory-Mapped I/O Limit i Registers Function 1: Offset 84h, 8Ch, 94h, 9Ch, A4h, ACh, B4h, BCh
8 MMIOLimiti 7 NP 6 reserved 5 DstLink 4 3 reserved 2 0
Memory-Mapped I/O Limit i Registers

31
DstNode
Bits 318 7 6 54 3 20
Mnemonic MMIOLimiti NP reserved DstLink reserved DstNode
Function Memory-Mapped I/O Limit Address i Non-Posted Destination Link ID Destination Node ID
R/W R/W R/W R R/W R R/W
Reset X X 0 X 0 X
Field Descriptions Destination Node ID (DstNode)Bits 20. This field specifies the node that a packet is sent to if it is within the address range. The bits are encoded as follows: 000b = Node 0 001b = Node 1 ... 111b = Node 7 Destination Link ID (DstLink)Bits 54. This field specifies the HyperTransport link number to send the packet to once the packet reaches the destination node. 00b = Link 0 01b = Link 1 10b = Link 2 11b = reserved Non-Posted (NP)Bit 7. When set to 1, this bit forces CPU writes to this memory-mapped I/O region to be non-posted. This may be used to force writes to be non-posted for MMIO regions which map to the legacy ISA/LPC bus, or in conjunction with the DsNpReqLmt field in the HyperTransport Transaction Control register (page 54) in order to allow downstream CPU requests to be counted and thereby limited to a specified number. This latter use of the NP bit may be used to avoid loop deadlock in systems that implement a reflection region in an I/O device that reflects downstream accesses back upstream. See the HyperTransport I/O Link Specification summary of deadlock scenarios for more information. 0 = CPU writes may be posted 1 = CPU writes must be non-posted Chapter 4 Memory System Configuration 81
32559
Rev. 3.08 July 2007
Memory-Mapped I/O Limit Address i (3916) (MMIOLimiti)Bits 318. This field defines the upper address bits of a 40-bit address that defines the end of memory-mapped I/O region n (where n = 0,1, . . . 7).
4.4.6
PCI I/O Address Map Registers
These registers define sections of the PCI I/O address map for which accesses should be routed to each noncoherent HyperTransport technology chain. PCI I/O regions must not overlap each other. For PCI I/O addresses within the specified range of a base/limit pair, requests are routed to the noncoherent HyperTransport link specified by the destination Node ID and destination Link ID. Addresses are considered to be within the defined range if they are greater than or equal to the base and less than or equal to the limit. For the purposes of this comparison, the lower unspecified bits of the base are assumed to be 0s and the lower unspecified bits of the limit are assumed to be 1s. PCI I/O accesses are generated from x86 IN/OUT instructions. 4.4.6.1 PCI I/O Base i Registers Function 1: Offset C0h, C8h, D0h, D8h
12 11 PCIIOBasei reserved 6 5 IE 4 VE 3 reserved 2 1 WE 0 RE
PCI I/O Base 03 Registers

31 reserved 25 24
Bits 3125 2412 116 5 4 32 1 0
Mnemonic reserved PCIIOBasei reserved IE VE reserved WE RE
Function PCI I/O Base Address i ISA Enable VGA Enable Write Enable Read Enable
R/W R R/W R R/W R/W R R/W R/W
Reset 0 X 0 X X 0 0 0
Field Descriptions Read Enable (RE)Bit 0. This bit enables reads to the defined address space. 0 = Disabled 1 = Enabled Write Enable (WE)Bit 1. This bit enables writes to the defined address space. 0 = Disabled 82 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
1 = Enabled VGA Enable (VE)Bit 4. Forces addresses in the first 64 Kbytes of PCI I/O space and where A[9:0] is in the range 3B03BBh or 3C03DFh to match against this base/limit pair independent of the base/limit addresses (see the PCI-to-PCI Bridge Architecture specification for a description of this function). A WE or RE bit must also be set to enable this feature. The MMIO 000A_0000h to 000B_FFFFh matching function normally associated with the VGA enable bit in PCI is NOT included in the VE bit. To enable this behavior, an MMIO register pair must be used. ISA Enable (IE)Bit 5. Blocks addresses in the first 64 Kbytes of PCI I/O space and in the last 768 bytes in each 1-Kbyte block from matching against this base/limit pair (see the PCI-to-PCI Bridge Architecture specification for a description of this function). A WE or RE bit must also be set to enable this feature. PCI I/O Base Address i (PCIIOBasei)Bits 2412. This field defines the start of PCI I/O region n (where n = 0, 1, 2,3). 4.4.6.2 PCI I/O Limit i Registers Function 1: Offset C4h, CCh, D4h, DCh
12 11 PCIIOLimiti reserved 6 5 DstLink 4 3 reserved 2 0
PCI I/O Limit 03 Registers

31 reserved 25 24
DstNode
Bits 3125 2412 116 54 3 20
Mnemonic reserved PCIIOLimiti reserved DstLink reserved DstNode
Function PCI I/O Limit Address i Destination Link ID Destination Node ID
R/W R R/W R R/W R R/W
Reset 0 X 0 X 0 X
Field Descriptions Destination Node ID (DstNode)Bits 20. This field specifies the node to which a packet is sent if it is within the address range. The bits are encoded as follows: 000b = Node 0 001b = Node 1 ... 111b = Node 7
Chapter 4
83
32559
Rev. 3.08 July 2007
Destination Link ID (DstLink)Bits 54. This field specifies the link to send the packet to once the packet has reached the destination node. 00b = Link 0 01b = Link 1 10b = Link 2 11b = reserved PCI I/O Limit Address i (PCI I/OLimiti)Bits 2412. This field defines the end of PCI I/O region
4.4.7
Configuration Map Registers
These registers define sections of the configuration bus number map for which configuration accesses should be routed to each noncoherent HyperTransport technology chain. Configuration regions must not overlap each other. For configuration accesses with bus numbers within the specified range of a base/limit pair, requests are routed to the noncoherent HyperTransport link specified by the destination Node ID and destination Link ID. Bus numbers are considered to be within the defined range if they are greater than or equal to the base and less than or equal to the limit. In device number compare mode (see DevCmpEn bit), the base/limit fields are interpreted as device number base/limit values for Bus 0 configuration accesses. This may be used to support configurations where Bus 0 is split between two independent noncoherent HyperTransport technology chains. Configuration accesses to Bus 0, device 2431, are assumed to be targeting coherent HyperTransport technology devices and are routed to the corresponding coherent HyperTransport node irrespective of the values in these registers (see Chapter 4, Memory System Configuration). Note that the range of device numbers affected may be modified based on the number of nodes present (see HyperTransport Transaction Control Register on page 54). Configuration Base and Limit 03 Registers
31 BusNumLimiti 24 23 BusNumBasei 16 15 reserved
Function 1: Offset E0h, E4h, E8h, ECh

10 9 DstLink 8 7 reserved 6 4 3 reserved 2 DevCmpEn 1 WE 0 RE
DstNode
Bits 3124 2316 1510 98 7 64
Mnemonic BusNumLimiti BusNumBasei reserved DstLink reserved DstNode
Function Bus Number Limit i Bus Number Base i Destination Link ID Destination Node ID
R/W R/W R/W R R/W R R/W
Reset X X 0 X 0 X
84
Chapter 4
32559
Rev. 3.08
July 2007
Bits 3 2 1 0
Mnemonic reserved DevCmpEn WE RE
Function Device Number Compare Enable Write Enable Read Enable
R/W R R/W R/W R/W
Reset 0 X 0 0
Field Descriptions Read Enable (RE)Bit 0. This bit enables reads to the defined address space. 0 = Disabled 1 = Enabled Write Enable (WE)Bit 1. This bit enables writes to the defined address space. 0 = Disabled 1 = Enabled Device Number Compare Enable (DevCmpEn)Bit 2. When this bit is set, this register defines a device number range rather than a bus number range. To match, configuration cycles must be to bus number 0 and have device numbers between BusNumBase and BusNumLimit. This is used to enable multiple noncoherent HyperTransport chains to be configured as Bus 0. Destination Node ID (DstNode)Bits 64. This field specifies the node that a packet is sent to if it is within the address range. The bits are encoded as follows: 000b = Node 0 001b = Node 1 ... 111b = Node 7 Destination Link ID (DstLink)Bits 98. This field specifies the link to send the packet to once it has reached the destination node and unit. 00b = Link 0 01b = Link 1 10b = Link 2 11b = reserved Bus Number Base i (BusNumBasei)Bits 2316. This field defines the lowest bus number in configuration region i (where i = 0, 1, 2, 3). Bus Number Limit i (BusNumLimiti)Bits 3124. This bit field defines the highest bus number in configuration region i (where i = 0, 1, 2, 3).
4.4.8
DRAM Hole Address Register
This register is used to control memory hoisting. See Section 3.6.5 on page 35 for more information on memory hoisting. Chapter 4 Memory System Configuration 85
32559
Rev. 3.08 July 2007
DRAM Hole Address Register

31 24 23 16 15
Function 1: Offset F0h

8 7 1 0 DramHoleValid
DramHoleBase
reserved
DramHoleOffset
reserved
Bits 3124 2316 158 71 0
Mnemonic DramHoleBase reserved DramHoleOffset reserved DramHoleValid
Function DRAM Hole Base Address DRAM Hole Offset Address DRAM Hole Valid
R/W R/W R R/W R R/W
Reset 0 0 0 0 0
Field Descriptions DRAM Hole Valid (DramHoleValid)Bit 0. This bit should be set in the processor node(s) that own the DRAM address space that is hoisted above the 4GB address level. If node interleaving is enabled, then this should be set in all processors. 0 = Memory hoisting is not enabled. 1 = Memory hoisting is enabled in the processor node. DRAM Hole Offset Bits 31-24 (DramHoleOffset)Bits 158. When memory hoisting is enabled, this value is subtracted from the physical address of certain transactions before being passed to the DRAM controller. DRAM Hole Base Bits 31-24 (DramHoleBase)Bit 3124. This specifies the base address of the I/O hole, below the 4G address level, that is used in memory hoisting. Normally, DramHoleBase >= TOP_MEM[31:24].
86
Chapter 4
32559
Rev. 3.08
July 2007
4.5
Function 2DRAM Controller
Function 2 configuration registers are listed in Table 7.

Table 7. Function 2 Configuration Registers
Offset 00h 04h 08h 0Ch 10h 14h 18h 1Ch 20h 24h 28h 2Ch 30h 34h 38h 3Ch 40h 44h 48h 4Ch 50h 54h 58h 5Ch 60h Register Name Device ID Status Base Class Code BIST
1
Reset Vendor ID (AMD) Command 1102_1022h 0000_0000h Revision ID Cache Line Size 0600_0000h 0080_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h Sub-System Vendor ID 0000_0000h 0000_0000h 0000_0000h 0000_0000h
Access RO RO
RO RO RO RO RO RO RO RO RO RO RO RO RO RO RW RW RW RW RW RW RW RW RW page 90 page 90 page 90 page 90 page 90 page 90 page 90 page 90 page 95
Base Address 0 Base Address 1 Base Address 2 Base Address 3 Base Address 4 Base Address 5 Card Bus CIS Pointer Sub-System ID ROM Base Address Capabilities reserved Max Latency Min GNT Int Pin Int Line
0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h
DRAM CS Base 0 DRAM CS Base 1 DRAM CS Base 2 DRAM CS Base 3 DRAM CS Base 4 DRAM CS Base 5 DRAM CS Base 6 DRAM CS Base 7 DRAM CS Mask 0
Chapter 4
87
BIOS and Kernel Developers Guide for AMD NPT Family 0Fh Processors Table 7. Function 2 Configuration Registers
Offset 64h 68h 6Ch 78h 7Ch 80h 88h 8Ch 90h 94h 98h 9Ch A0h Register Name DRAM CS Mask 1 DRAM CS Mask 2 DRAM CS Mask 3 DRAM Control DRAM Initialization DRAM Bank Address Mapping DRAM Timing Low DRAM Timing High DRAM Configuration Low DRAM Configuration High DRAM Controller Additional Data Offset DRAM Controller Additional Data DRAM Controller Miscellaneous Reset 0000_0000h 0000_0000h 0000_0000h 0000_0006h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000
32559
Rev. 3.08 July 2007
Access RW RW RW RW RW RW RW RW RW RW RW RW R/W
Description page 95 page 95 page 95 page 95 page 96 page 98 page 104 page 106 page 109 page 109 page 114 page 116 page 126
4.5.1

0
Bits 3116 150
R/W R R
Reset 1102h 1022h
Field Descriptions Vendor ID (VenID)Bits 150. This read-only value is defined as 1022h for AMD. Device ID (DevID)Bits 3116. This read-only value is defined as 1102h for the HyperTransport technology configuration function. 88 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
4.5.2
31 BCC 24 23 SCC 16 15 PI

8 7 RevID 0
Bits 3124 2316 158 70
R/W R R R R
4.5.3

8 7 CLS 0
Bits 3124 2316 158 70
R/W R R R R
Chapter 4
89
32559
Rev. 3.08 July 2007
Field Descriptions CacheLineSize (CLS)Bits 70. This read-only value is defined as 00h. LatencyTimer (LatTimer)Bits 158. This read-only value is defined as 00h. HeaderType (Htype)Bits 2316. This read-only value is defined as 80h to indicate that multiple functions are present in the configuration header and that the header layout corresponds to a device header as opposed to a bridge header. BISTBits 3124. This read-only value is defined as 00h.
4.5.4
DRAM CS Base Address Registers
These registers define DRAM chip select (CS) address mapping. They are programmed based on data in the Serial Presence Detect (SPD) ROM on each DIMM. Function 1 address map registers define to which node to route a DRAM request. Function 2 address map registers define what DRAM chip select to access on that particular node. See DRAM Address Map on page 76. for more information on addresses routed to the DRAM controller (InputAddr). The memory size of each chip select bank is defined by a DRAM CS Base Address Register and a DRAM CS Mask register. The chip selects are formed as follows, using the field names from the DRAM CS Base Address Registers and DRAM CS Mask Registers, where , means concatenate / means not == means equals & means and ChipSelect[i] is asserted if the following is true:
CSBE[i] & ( { (InputAddr[36:27] & /AddrMaskHi[i][36:27]), ( InputAddr[21:13] & /AddrMaskLo[i][21:13]) } == { (BaseAddrHi[i][36:27] & /AddrMaskHi[i][36:27]), ( BaseAddrLo[i][21:13] & /AddrMaskLo[i][21:13]) } ) );
There are eight DRAM CS Base Address and four DRAM CS Mask register pairs. Table 8 and Table 9 show the mapping between the DRAM CS Base and Mask registers and the logical DIMM, the chip select, the CKE pins and the ODT pins.
90
Chapter 4
32559
Rev. 3.08
July 2007
Table 8. DRAM CS Base Address and DRAM CS Mask Registers

Registers N DRAM CS Base 0 DRAM CS Mask 0 DRAM CS Base 1 DRAM CS Mask 0 DRAM CS Base 2 DRAM CS Mask 1 DRAM CS Base 3 DRAM CS Mask 1 DRAM CS Base 4 DRAM CS Mask 2 DRAM CS Base 5 DRAM CS Mask 2 DRAM CS Base 6 DRAM CS Mask 3 DRAM CS Base 7 DRAM CS Mask 3 0 0 1 1 2
2
Logical DIMM
1
Chip Select
CKE
ODT
R4 S4 0 0 1 1 0 0 1 1 0 0 0 0 M[B,A]0_CS_L[0] M[B,A]0_CS_L[1] M[B,A]1_CS_L[0] M[B,A]0_CS_L[2] M[B,A]1_CS_L[1] M[B,A]0_CS_L[3]

3 5
M[B,A]_CKE[0] M[B,A]0_ODT[0] M[B,A]0_ODT[0] M[B,A]_CKE[1] M[B,A]1_ODT[0] M[B,A]0_ODT[1] M[B,A]_CKE[0] M[B,A]_CKE[1]

3 4 5
M[B,A]1_ODT[0] M[B,A]0_ODT[0] ,
3
4 5
3 5
M[B,A]1_ODT[0] M[B,A]0_ODT[1]
3 5
M[B,A]2_CS_L[0] M[B,A]2_CS_L[1] M[B,A]3_CS_L[0] M[B,A]3_CS_L[1]
M[B,A]_CKE[0] M[B,A]2_ODT[0] M[B,A]_CKE[1] M[B,A]2_ODT[0] M[B,A]_CKE[0] M[B,A]3_ODT[0] M[B,A]_CKE[1] M[B,A]3_ODT[0]
2 3 3
Notes: 1. N=Normal. R4=Four-rank registered DIMM only (Function 2 Offset 94h bit 18 = 1). S4=Four-rank SO-DIMM only (Function 2 Offset 94h bit 17 = 1); only 4 chip selects are supported. 2. Logical DIMM numbers 2 and 3 are not supported in the S1g1 and AM2 packages. 3. F(1207) and AM2 Packages. 4. SO-DIMM using an AM2 Package. 5. S1g1 Package.
Chapter 4
91
32559
Rev. 3.08 July 2007
Table 9.
DRAM CS Base Address and DRAM CS Mask Registers Mismatched DIMM Support
1
Registers
Logical DIMM
2
Chip Select
CKE
ODT
N DRAM CS Base 0 DRAM CS Mask 0 DRAM CS Base 1 DRAM CS Mask 0 DRAM CS Base 2 DRAM CS Mask 1 DRAM CS Base 3 DRAM CS Mask 1 DRAM CS Base 4 DRAM CS Mask 2 DRAM CS Base 5 DRAM CS Mask 2 DRAM CS Base 6 DRAM CS Mask 3 DRAM CS Base 7 DRAM CS Mask 3 0 0 1 1 2 2 3 3
S4 0 0 0 0 1 1 1 1 MA0_CS_L[0] MA0_CS_L[1] MA1_CS_L[0] MA0_CS_L[2] MA1_CS_L[1] MA0_CS_L[3] MB0_CS_L[0] MB0_CS_L[1] MB1_CS_L[0] MB0_CS_L[2] MB1_CS_L[1] MB0_CS_L[3]
3 5 3 5
MA_CKE[0] MA_CKE[1] MA_CKE[0] MA_CKE[1] MB_CKE[0] MB_CKE[1] MB_CKE[0] MB_CKE[1]
MA0_ODT[0] MA0_ODT[0] MA1_ODT[0] MA0_ODT[1]

3 4 5
MA1_ODT[0] MA0_ODT[0] ,
3
4 5
3 5
MA1_ODT[0] MA0_ODT[1]
3 5
MB0_ODT[0] MB0_ODT[0] MB1_ODT[0] MB0_ODT[1]

3 4 5
MB1_ODT[0] MB0_ODT[0] ,
3
4 5
3 5
MB1_ODT[0] MB0_ODT[1]
3 5
Notes: 1. Mod64Mux (Function 2 Offset A0h bit 4) = 1. 2. N=Normal. S4=Four-rank SO-DIMM only (Function 2 Offset 94h bit 17 = 1); only 4 chip selects are supported. 3. AM2 Packages. 4. SO-DIMM using an AM2 Package. 5. S1g1 Package.
DRAM memory can be assigned to chip select banks in two ways: non-interleaving, when contiguous addresses are assigned to each chip select bank, and interleaving, when non-contiguous addresses are assigned to each chip select bank. Non-interleaving mode can always be used. The BIOS must assign the largest DIMM chip-select range to the lowest address. As addresses increase, the chip select size must remain constant or decrease. This is necessary to keep DIMM chip select banks on aligned address boundaries as chipselect banks with different depths are added. The masking does not work otherwise. Memory interleaving mode can be used if the memory system is composed of a single type and size DRAM, and if the number of chip selects is a power of two. DRAM controller swaps low order address bits with high order address bits during decoding. As a result, some low order address bits are 92 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
used to determine chip select, and some high order address bits are used to determine DIMM row. It allows that a switch between DIMMs occurs before a page conflict within a DIMM. The one cycle turnaround saves time needed to close/open new pages. Algorithm for programming DRAM CS Base Address and DRAM CS Mask registers in interleaving mode are described in section DRAM Address Mapping in Interleaving Mode on page 100. The BIOS sets up the base address and base mask registers after determining the number of populated DIMMs, how many chip-select banks they have, and their sizes. A single base address is initialized for each chip select. A single mask is initialized for each chip select pair. Hardware decodes addresses and determines to which DIMM chip-select range the address maps, as a function of the base address and mask. DRAM CS Base Address Registers
31 29 28 BaseAddrHi 19 18 reserved
Function 2: Offset 40h, 44h, 48h, 4Ch, 50h, 54h, 58h, 5Ch
14 13 BaseAddrLo 5 4 reserved 3 2 TestFail 1 Spare 0 CSEnable
reserved
Bits 3129 2819 1814 135 43 2 1 0
Mnemonic reserved BaseAddrHi reserved BaseAddrLo reserved TestFail Spare CSEnable
Function Base Address (3627) Base Address (2113) Memory Test Failed Spare Rank Chip-Select Bank Enable
R/W R R/W R R/W R R/W R/W R/W
Reset 0 0 0 0 0 0 0 0
Field Descriptions Chip-Select Bank Enable (CSEnable)Bit 0. This bit enables access to the defined address space. The TestFail and Spare bits should never be set if this bit is set. Spare Rank (Spare)Bit 1. This bit identifies the chip select associated with the spare rank. The TestFail and CSEnable bits should never be set if this bit is set. This bit must be set prior to DRAM initialization. Memory Test Failed (TestFail)Bit 2. This bit is set by BIOS to indicate that a rank is present but the memory is bad. The CSEnable and Spare bits should never be set if this bit is set. This bit must be set prior to DRAM initialization. Base Address (2113) (BaseAddrLo)Bits 135. This field is only used for memory interleaving to avoid page conflicts. A new chip select bank is accessed before accessing a new row in the old chip select bank, delaying or avoiding a page conflict. Chapter 4 Memory System Configuration 93
32559
Rev. 3.08 July 2007
If contiguous addresses from address 0 are accessed, the controller would first access a row (e.g., row 0) in bank 0, chip select 0. As the address increases, the controller would access different columns in the following order for four bank DIMMs: row 0/bank 0/chip select 0 and then row 0/bank 1/chip select 0 and then row 0/bank 2/chip select 0 and then row 0/bank 3/chip select 0 and then row 0/bank 0/chip select 1. As the address increases, the controller would access different columns in the following order for eight bank DIMMs: row 0/bank 0/chip select 0 and then row 0/bank 1/chip select 0 and then row 0/bank 2/chip select 0 and then row 0/bank 3/chip select 0 and then row 0/bank 4/chip select 0 and then row 0/bank 5/chip select 0 and then row 0/bank 6/chip select 0 and then row 0/bank 7/chip select 0 and then row 0/bank 0/chip select 1. A chip select bank does not have a contiguous region of memory assigned to it. Memory is interleaved between two DIMMs, four DIMMs, or eight DIMMs. Base Address (3627) (BaseAddrHi)Bits 2819. These bits decode 128-Mbyte blocks of memory. In the non-interleaving mode (BaseAddrLo has a value of 0), physical addresses map to DRAM addresses in the following way: ColLowest order physical address bits BankSecond lowest order physical address bits (page miss is better than page conflict) RowSecond highest order physical address bits (page conflict before accessing new chip) Chip SelectHighest order physical address bits In the non-interleaving mode, contiguous addresses from 0 would first access a row (e.g., row 0) in bank 0, chip select 0. As the address increases, the next access is to a different column and eventually a different chip select bank, e.g.: row 0/bank 0/chip select 0 and then row 0/bank 1/chip select 0 and then row 0/bank 2/chip select 0 and then row 0/bank 3/chip select 0 and then row 1/bank 0/chip select 0 Memory controller accesses all rows and banks in a single chip select bank before accessing a new chip select bank in the non-interleaving mode. 94 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
4.5.5
DRAM CS Mask Registers
The purpose of this register is to exclude the corresponding address bits from the comparison with the DRAM Base address register. DRAM CS Mask Registers
31 29 28 AddrMaskHi 19 18 reserved reserved
Function 2: Offset 60h, 64h, 68h, 6Ch

14 13 AddrMaskLo 5 4 reserved 0
Bits 3129 2819 1814 135 40
Mnemonic reserved AddrMaskHi reserved AddrMaskLo reserved
Function Address Mask (3627) Address Mask (2113)
R/W R R/W R R/W R
Reset 0 0 0 0 0
Field Descriptions Address Mask (2113) (AddrMaskLo)Bits 135. This field specifies the addresses to be excluded for the memory interleaving mode described in the Base Address bit definitions. Address Mask (3627) (AddrMaskHi)Bits 2819. This field defines the top Address Mask bits. The bits with an address mask of 1 are excluded from the address comparison. This allows the memory block size to be larger than 128 Mbytes. If Address Mask bit 19 is set to 1, the memory block size is 256 Mbytes.
4.5.6
DRAM Control Register
This register is used to control the behavior of the DRAM controller. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this Register. DRAM Control Register
31 19 18 17 16 15 AltVidC3MemClkTriEn DllTempAdjTime DqsRcvEnTrain

7 6 RdPadRcvFifoDly 4 3 0
reserved
reserved
RdPtrInit
Chapter 4
95
32559
Rev. 3.08 July 2007
Bits 3119 18 17 16 157 64 30
Mnemonic reserved DqsRcvEnTrain DllTempAdjTime AltVidC3MemClkTriEn reserved RdPadRcvFifoDly RdPtrInit
Function DQS Receiver Enable Training Mode DLL Temperature Adjust Cycle Time AltVID Memory Clock Tristate Enable Read Delay from Pad Receive FIFO Read Pointer Initial Value
R/W R R/W R/W R/W R R/W R/W
Reset 0 0 0 0 0 7h 6h
Field Descriptions Read Pointer Initial Value (RdPtrInit)Bits 30. This field specifies the value of the read pointer for the pad transmit FIFO when the read pointer is reinitialized. Writing to this register reinitializes the read and write pointers in the pad transmit FIFO. Values of 3h-0h are reserved and should never programmed into this field. Read Delay from Pad Receive FIFO (RdPadRcvFifoDly)Bits 64. This field specifies the delay from the DQS receiver enable to the first data being read from the pad receive FIFO. 000b = 0.5 Memory Clocks 001b = 1 Memory Clock 010b = 1.5 Memory Clocks 011b = 2 Memory Clocks 100b = 2.5 Memory Clocks 101b = 3 Memory Clock 110b = 3.5 Memory Clocks 111b = 4 Memory Clocks AltVID Memory Clock Tristate Enable (AltVidC3MemClkTriEn)Bit 16. Enables the DDR memory clocks to be tristated when alternate VID mode is enabled. This bit has no effect if the DisNbClkRamp bit (Function 3, Offset 88h) is set. DLL Temperature Adjust Cycle Time (DllTempAdjTime)Bit 17. This bit selects the DLL temperature adjust Cycle Time. 0 = 5ms 1 = 1ms DQS Receiver Enable Training Mode (DqsRcvEnTrain)Bit 18. This bit enables the DQS receiver enable training mode. 0 = Normal DQS receiver enable operation. 1 = DQS receiver enable training mode.
4.5.7
DRAM Initialization Register
This register is used by BIOS to control the DRAM initialization sequence. The BIOS is responsible for not changing the command specified until the controller has completed the command. See 96 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this Register. DRAM Initialization Register
31 30 29 28 27 26 25 24 23 DeassertMemRstX SendAutoRefresh SendMrsCmd SendPchgAll EnDramInit AssertCke 19 18 16 15

0
reserved
reserved
MrsBank
MrsAddress
Bits 31 3029 28 27 26 25 24 2319 1816 150
Mnemonic EnDramInit reserved AssertCke DeassertMemRstX SendMrsCmd SendAutoRefresh SendPchgAll reserved MrsBank MrsAddress
Function Enable DRAM Initialization Assert CKE Deassert Memory Reset Send MSR/EMSR Command Send Autorefresh Command Send Precharge All Command Bank Address for MRS/EMRS Commands Address for MRS/EMRS Commands
R/W R/W R R/W R/W R/W R/W R/W R R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0
Field Descriptions Address for MRS/EMRS Commands (MrsAddress)Bits 150. This field specifies the data driven on the DRAM address pins 15-0 for MRS and EMRS commands. Bank Address for MRS/EMRS Commands (MrsBank)Bits 1816. This field specifies the data driven on the DRAM bank pins for MRS and EMRS commands. Send Precharge All Command (SendPchgAll)Bit 24. Setting this bit causes the DRAM controller to send a precharge all command. This bit is cleared by the hardware after the command completes. Send Auto Refresh Command (SendAutoRefresh)Bit 25. Setting this bit causes the DRAM controller to send an auto refresh command. This bit is cleared by the hardware after the command completes. Send MRS/EMRS Command (SendMrsCmd)Bit 26. Setting this bit causes the DRAM controller to send the MRS or EMRS command defined by the MrsAddress and MrsBank fields. This bit is cleared by the hardware after the command completes.
Chapter 4
97
32559
Rev. 3.08 July 2007
De-assert Memory Reset (DeassertMemRstX)Bit 27. Setting this bit causes the DRAM controller to de-assert the memory reset pin. This bit cannot be used to assert the memory reset pin. Assert CKE (AssertCke)Bit 28. Setting this bit causes the DRAM controller to assert the CKE pins. This bit cannot be used to de-assert the CKE pins. Enable DRAM Initialization (EnDramInit)Bit 31. Setting this bit puts the DRAM controller in a BIOS controlled DRAM initialization mode. Setting this bit causes the DRAM controller to assert memory reset for registered DIMMs and de-assert CKE for all DIMMs. BIOS must clear this bit after DRAM initialization is complete.
4.5.8
DRAM Bank Address Mapping Register
Each DIMM may have one or two chip-select banks. The address mode covers the whole DIMM, regardless of whether there is one or two chip-select banks. DRAM CS Address Mapping Register
31 reserved 16 15 CS7/6 12 11 CS5/4

8 7 CS3/2 4 3 CS1/0 0
Bits 3116 1512 118 74 30
Mnemonic reserved CS7/6 CS5/4 CS3/2 CS1/0
Function CS7/6 CS5/4 CS3/2 CS1/0
R/W R R/W R/W R/W R/W
Reset 0 0 0 0
Field Descriptions Chip Select 1/0 (CS1/0)Bits 30. This field specifies the memory module size. This field is programmed the same regardless of whether the DRAM interface is 64-bits or 128-bits wide. This field describes the type of DIMMs that compose the chip select. The bits are encoded as shown in Table 10 and Table 11. Chip Select 3/2 (CS3/2)Bits 74. This field specifies the memory module size. The bits are encoded as shown in Table 10 and Table 11. Chip Select 5/4 (CS5/4)Bits 118. This field specifies the memory module size. The bits are encoded as shown in Table 10 and Table 11. Chip Select 7/6 (CS7/6)Bits 1512. This field specifies the memory module size. The bits are encoded as shown in Table 10 and Table 11.
98
Chapter 4
32559
Rev. 3.08
July 2007
4.5.8.1
DRAM Address Mapping
The address line mapping with a 64-bit DRAM interface is shown in Table 10 and with a 128-bit DRAM interface in Table 11. Column A10 presents the precharge all signal (PC).
Table 10: DRAM address mapping 64-bit interface.

Device size, Bits CS Size width 2 x x x x 15 15 x 15 16 16 15 16 0000b 128 MB 256Mb, x16 0001b 256 MB 0010b 512 MB 0011b 512 MB 0100b 512 MB 0101b 0110b 0111b 1000b 1001b 1010b 1011b 1 GB 1 GB 2 GB 2 GB 4 GB 4 GB 8 GB 256Mb, x8 512Mb, x16 512Mb, x8 256Mb, x4 1G, x16 1G, x8 2G, x16 512Mb, x4 2Gb, x8 4Gb, x16 1Gb, x4 2Gb, x4 4Gb, x8 4Gb, x4 15 15 14 15 14 14 13 14 15 14 14 13 14 15 14 14 13 14 13 13 Bank 1 13 14 0 12 13 Row Col Row Col Row Col Row Col Row Col Row Col Row Col Row Col Row Col Row Col Row Col Row Col 15 x x x x x x x x x x x x x x x x x x x x 17 x 17 x 14 x x x x x x x x x x x x x x 17 x x x 17 x 16 x 32 x 13 x x x x 17 x x x x x 17 x 17 x 16 x 17 x 31 x 31 x 31 x 12 17 x 17 x 16 x 17 x 17 x 16 x 16 x 30 x 30 x 30 x 30 x 30 x 11 16 x 16 x 15 x 16 16 x 29 x 29 29 x 29 29 29 x 29 10 15 AP 15 28 28 28 28 28 28 28 28 28 28 Address 9 14 x 27 27 27 27 27 27 27 27 27 27 27 8 26 11 26 11 26 11 26 11 26 11 26 11 26 11 26 11 26 11 26 11 26 11 26 11 7 25 10 25 10 25 10 25 10 25 10 25 10 25 10 25 10 25 10 25 10 25 10 25 10 6 24 9 24 9 24 9 24 9 24 9 24 9 24 9 24 9 24 9 24 9 24 9 24 9 5 23 8 23 8 23 8 23 8 23 8 23 8 23 8 23 8 23 8 23 8 23 8 23 8 4 22 7 22 7 22 7 22 7 22 7 22 7 22 7 22 7 22 7 22 7 22 7 22 7 3 6 6 6 6 6 6 6 6 6 6 6 6 2 5 5 5 5 5 5 5 5 5 5 5 5 1 19 4 19 4 19 4 19 4 19 4 19 4 19 4 19 4 19 4 19 4 19 4 19 4 0 18 3 18 3 18 3 18 3 18 3 18 3 18 3 18 3 18 3 18 3 18 3 18 3 21 20 21 20 21 20 21 20 21 20 21 20 21 20 21 20 21 20 21 20 21 20 21 20
AP 12 AP 12
13 AP 12 AP 12 AP 12
13 AP 12 AP 12
13 AP 12 13 AP 12 AP 12
13 AP 12
Table 11: DRAM address mapping 128-bit interface.

Device size, Bits CS Size width 2 x x x 0000b 256 MB 256Mb, x16 0001b 512 MB 0010b 1 GB 256Mb, x8 512Mb, x16 512Mb, x8 15 14 Bank 1 14 15 0 13 14 Row Col Row Col Row Col 15 x x x x x x 14 x x x x x x 13 x x x x 18 x 12 18 x 18 x 17 x 11 17 x 17 x 16 x 10 16 AP 16 29 Address 9 15 x 28 28 8 27 12 27 12 27 12 7 26 11 26 11 26 11 6 25 10 25 10 25 10 5 23 9 24 9 24 9 4 23 8 23 8 23 8 3 7 7 7 2 6 6 6 1 20 5 20 5 20 5 0 19 4 19 4 19 4 22 21 22 21 22 21
AP 13 AP 13
Chapter 4
99
BIOS and Kernel Developers Guide for AMD NPT Family 0Fh Processors Table 11: DRAM address mapping 128-bit interface.
Device size, Bits 0011b 0100b 0101b 0110b 0111b 1000b 1001b 1010b CS Size 1 GB 1 GB 2 GB 2 GB 4 GB 4 GB 8 GB 8 GB width 256Mb, x4 1G, x16 1G, x8 2G, x16 512Mb, x4 2Gb, x8 4Gb, x16 1Gb, x4 2Gb, x4 4Gb, x8 4Gb, x4 17 17 16 17 16 16 15 16 15 15 14 15 x 16 16 15 15 14 2 x 16 16 Bank 1 16 15 15 0 15 14 14 Row Col Row Col Row Col Row Col Row Col Row Col Row Col Row Col 1011b 16 GB Row Col 15 x x x x x x x x x x x x x x 18 x 18 x 14 x x x x x x x x 18 x x x 18 x 17 x 33 x 13 x x x x 18 x 18 x 17 x 18 x 32 x 32 x 32 x 12 18 x 18 x 17 x 17 x 31 x 31 x 31 x 31 x 31 x 11 17 17 x 30 x 30 30 x 30 30 30 x 30 10 29 29 29 29 29 29 29 29 29 Address 9 28 28 28 28 28 28 28 28 28 8 27 12 27 12 27 12 27 12 27 12 27 12 27 12 27 12 27 12 7 26 11 26 11 26 11 26 11 26 11 26 11 26 11 26 11 26 11 6 25 10 25 10 25 10 25 10 25 10 25 10 25 10 25 10 25 10
32559
Rev. 3.08 July 2007
5 24 9 24 9 24 9 24 9 24 9 24 9 24 9 24 9 24 9
4 23 8 23 8 23 8 23 8 23 8 23 8 23 8 23 8 23 8
3 7 7 7 7 7 7 7 7 7
2 6 6 6 6 6 6 6 6 6
1 20 5 20 5 20 5 20 5 20 5 20 5 20 5 20 5 20 5
0 19 4 19 4 19 4 19 4 19 4 19 4 19 4 19 4 19 4
22 21 22 21 22 21 22 21 22 21 22 21 22 21 22 21 22 21
14 AP 13 AP 13 AP 13
14 AP 13 AP 13
14 AP 13 14 AP 13 AP 13
14 AP 13
4.5.8.2
DRAM Address Mapping in Interleaving Mode
In memory interleaving mode all DIMM chip-select ranges are the same size and type, and the number of chip selects is a power of two. A BIOS algorithm for programming DRAM CS Base Address and DRAM CS Mask registers in memory interleaving mode is: 1. Program all DRAM CS Base Address and DRAM CS Mask registers using contiguous mapping. 2. For each enabled chip select, swap BaseAddrHi bits with BaseAddrLo bits, as defined Table 12 and Table 13. 3. For each enabled chip select, swap AddrMaskHi bits with AddrMaskLo bits, as defined in Table 12 and Table 13.
Table 12.
Swapped physical address lines for interleaving with 64-bit interface

Swapped Base Address and Address Mask bits 8 way interleaving [29:27] and [16:14] [30:28] and [17:15] [31:29] and [17:15] 4 way interleaving [28:27] and [15:14] [29:28] and [16:15] [30:29] and [16:15] 2 way interleaving [27] and [14] [28] and [15] [29] and [15]
Chip Select Chip Select Size Mode 0000b 0001b 0010b 128-MB 256-MB 512-MB
100
Chapter 4
32559
Rev. 3.08
July 2007
Table 12.

Swapped Base Address and Address Mask bits 8 way interleaving [31:29] and [18:16] [31:29] and [18:16] [32:30] and [18:16] [32:30] and [18:16] [33:31] and [18:16] [33:31] and [19:17] [34:32] and [19:17] [34:32] and [18:16] [35:33] and [19:17] 4 way interleaving [30:29] and [17:16] [30:29] and [17:16] [31:30] and [17:16] [31:30] and [17:16] [32:31] and [17:16] [32:31] and [18:17] [33:32] and [18:17] [33:32] and [17:16] [34:33] and [18:17] 2 way interleaving [29] and [16] [29] and [16] [30] and [16] [30] and [16] [31] and [16] [31] and [17] [32] and [17] [32] and [16] [33] and [17]
Chip Select Chip Select Size Mode 0011b 0100b 0101b 0110b 0111b 1000b 1001b 1010b 1011b 512-MB 512-MB 1-GB 1-GB 2-GB 2-GB 4-GB 4-GB 8-GB
Table 13.

Swapped Base Address and Address Mask bits 8 way interleaving [30:28] and [17:15] [31:29] and [18:16] [32:30] and [18:16] [32:30] and [19:17] [32:30] and [19:17] [33:31] and [19:17] [33:31] and [19:17] [34:32] and [19:17] [34:32] and [20:18] [35:33] and [20:18] [35:33] and [19:17] [36:34] and [20:18] 4 way interleaving [29:28] and [16:15] [30:29] and [17:16] [31:30] and [17:16] [31:30] and [18:17] [31:30] and [18:17] [32:31] and [18:17] [32:31] and [18:17] [33:32] and [18:17] [33:32] and [19:18] [34:33] and [19:18] [34:33] and [18:17] [35:34] and [19:18] 2 way interleaving [28] and [15] [29] and [16] [30] and [16] [30] and [17] [30] and [17] [31] and [17] [31] and [17] [32] and [17] [32] and [18] [33] and [18] [33] and [17] [34] and [18]
Chip Select Chip Select Size Mode 0000b 0001b 0010b 0011b 0100b 0101b 0110b 0111b 1000b 1001b 1010b 1011b 256-MB 512-MB 1-GB 1-GB 1-GB 2-GB 2-GB 4-GB 4-GB 8-GB 8-GB 16-GB
Example. DRAM memory consists of two 2-sided DIMMs with 256 MBytes on each side. A 64-bit interface to DRAM is used. 1. Register settings for contiguous memory mapping are: Function 2, Offset 80h = 0000_0011h // CS0/1 = 256 MB; CS2/3 = 256 MB Function 2, Offset 40h = 0000_0001h // 0 MB base Function 2, Offset 44h = 0010_0001h // 256 MB base = 0 MB + 256 MB Function 2, Offset 48h = 0020_0001h // 512 MB base = 256 MB + 256 MB Chapter 4 Memory System Configuration 101
32559
Rev. 3.08 July 2007
Function 2, Offset 4Ch = 0030_0001h // 768 MB base = 512 MB + 256 MB Function 2, Offset 60h = 0008_3FF0h // CS0/CS1 = 256 MB Function 2, Offset 64h = 0008_3FF0h // CS2/CS3 = 256 MB 2. Base Address bits to be swapped are defined in Table 12 (64-bit interface), 256MByte row (chip select size), 4 way interleaving column (4 chip selects are used). Base Address bits [29:28] are defined with BaseAddrHi[21:20]. Base Address bits [16:15] are defined with BaseAddrLo[8:7]. Function 2, Offset 40h=0000_0001h Function 2, Offset 44h=0000_0081h Function 2, Offset 48h=0000_0101h Function 2, Offset 4Ch=0000_0181h 3. Address Mask bits to be swapped are the same as the Base Address bits defined in the previous step. Address Mask bits [29:28] are defined with AddrMaskHi[21:20]. Address Mask bits [16:15] are defined with AddrMaskLo[8:7]. Function 2, Offset 60h=0038_3E70h Function 2, Offset 64h=0038_3E70h
4.5.9
Address to Node/Chip Select Translation
The system address map consists of DRAM memory and memory mapped I/O space. A system address is mapped to a node with DRAM base registers (Function 1, Offsets 40h, 48h, etc.) and DRAM limit registers (Function 1, Offsets 44h, 4Ch, etc.). A system address (SysAddr) that is within the address range for a node, and is not a memory mapped I/O address, is normalized (InputAddr) and forwarded to the DRAM controller on that node (see 4.4.4 for details on SysAddr to InputAddr translation). A DRAM address (InputAddr) is mapped to DRAM chip selects with DRAM CS base address registers (Function 2, Offsets 40h, 44h, etc.) and DRAM CS mask registers (Function 2, Offsets 60h, 64h, etc.) (see 4.5.4 and 4.5.5). The following algorithm is designed to be used to determine the node and the chip select for a system address that maps to DRAM. SystemAddr is a 32 bit input variable that is equal to bits 39-8 of the system address. CSFound, NodeID, and CS are output variables. If CSFound is equal to 1, than NodeID and CS outputs are equal to the node and the chip select that corresponds to the input address. If the On-line Spare feature is enabled BIOS assigns one of the chip-selects, (constant) SPARE_RANK, to be the spare rank in the event of a DIMM failure precondition. If the DIMM failure precondition occurs and the data of the failing rank is copied over, the spare rank decodes to the same system address range as the failing rank (BadDramCs).
(int,int,int) TranslateSysAddrToCS((uint32)SystemAddr){ int CSFound, NodeID, CS, F1Offset, F2Offset, F2MaskOffset, Ilog; uint32 IntlvEn, IntlvSel; uint32 DramBase, DramLimit, DramEn; uint32 HoleOffset, HoleEn; uint32 CSBase, CSLimit, CSMask, CSEn;
102
Chapter 4
32559
Rev. 3.08
July 2007
uint32 InputAddr, Temp; uint32 OnlineSpareCTL; int SwapDone, BadDramCs; CSFound = 0; for(NodeID = 0; NodeID < 8; NodeID++){ F1Offset = 0x40 + (NodeID << 3); DramBase = Get_PCI(bus0, dev24 + NodeID, func1, F1Offset); DramEn = DramBase & 0x00000003; IntlvEn = (DramBase & 0x00000700) >> 8; DramBase = DramBase & 0xFFFF0000; DramLimit = Get_PCI(bus0, dev24 + NodeID, func1, F1Offset + 4); IntlvSel = (DramLimit & 0x00000700) >> 8; DramLimit = DramLimit | 0x0000FFFF; HoleEn = Get_PCI(bus0, dev24 + NodeID, func1, 0xF0); HoleOffset = (HoleEn & 0x0000FF00) << 8; HoleEn = (HoleEn &0x00000001); if(DramEn && DramBase <= SystemAddr && SystemAddr <= DramLimit){ If(HoleEn && SystemAddr > 0x00FFFFFF) InputAddr = SystemAddr - HoleOffset; Else InputAddr = SystemAddr - DramBase; if(IntlvEn){ if(IntlvSel == ((SystemAddr >> 4) & IntlvEn)){ if(IntlvEn == 1) Ilog = 1; else if(IntlvEn == 3) Ilog = 2; else if(IntlvEn == 7) Ilog = 3; else break; Temp = (InputAddr >> (4 + Ilog)) << 4; InputAddr = (Temp | (SystemAddr & 0x0000000F)); } else continue; } for(CS = 0; CS < 8; CS++){ F2Offset = 0x40 + (CS << 2); if ((CS % 2) == 0){ F2MaskOffset = 0x60 + (CS << 1); } else{ F2MaskOffset = 0x60 + ((CS-1) << 1); } CSBase = Get_PCI(bus0, dev24 + NodeID, func2, F2Offset); CSEn = CSBase & 0x00000001; CSBase = CSBase & 0x1FF83FE0; CSMask = Get_PCI(bus0, dev24 + NodeID, func2, F2MaskOffset); CSMask = (CSMask | 0x0007C01F) & 0x1FFFFFFF; if(CSEn && ((InputAddr & ~CSMask) == (CSBase & ~CSMask))){ CSFound = 1; OnlineSpareCTL = Get_PCI(bus0, dev24 + NodeID, func3, 0xB0); SwapDone = (OnlineSpareCTL >> 1) & 0x00000001; BadDramCs = (OnlineSpareCTL >> 4) & 0x00000007;
Chapter 4
103
if(SwapDone && CS == BadDramCs) CS=SPARE_RANK; break; } } } if(CSFound) break; } return(CSFound,NodeID,CS); }
32559
Rev. 3.08 July 2007
4.5.10
DRAM Timing Low Register
This register contains the normal timing parameters specified in a DRAM data sheet. DRAM Timing Low Register
31 MemClkDis 24 23 22 21 20 19 Trrd Twr Trc 16 15 Tras 12 11 10 9 reserved Trtp 8

7 reserved 6 5 4 3 reserved 2 1 Tcl 0
Trp
Trcd
Bits 3124 2322 2120 1916 1512 11 10 98 76 54 3 20
Mnemonic MemClkDis Trrd Twr Trc Tras Trtp reserved Trp reserved Trcd reserved Tcl
Function MEMCLK Disable Active-to-active (RAS#-to-RAS#) Delay Write Recovery Time Row Cycle Time Minimum RAS# Active Time Read CAS# to Precharge Time Row Precharge Time RAS#-active to CAS#-read/write Delay CAS# Latency
R/W R/W R/W R/W R/W R/W R/W R R/W R R/W R R/W
Reset FFh 0 0 0 0 0 0 0 0 0 0 0
Field Descriptions CAS# Latency (Tcl)Bits 20. Specifies the CAS#-to-read-data-valid. 000b = reserved 001b = reserved 010b = CL=3 011b = CL=4 100b = CL=5 101b = CL=6 11xb = reserved
104
Chapter 4
32559
Rev. 3.08
July 2007
RAS# Active to CAS# read/write Delay (Trcd)Bits 54. Specifies the RAS# to CAS# delay for a read/write command to the same bank. 00b = 3 Clocks 01b = 4 Clocks 10b = 5 Clocks 11b = 6 Clocks Row Precharge Time (Trp)Bits 98. Specifies the row precharge time (Precharge-to-Active or Auto-Refresh of the same bank.) 00b = 3 Clocks 01b = 4 Clocks 10b = 5 Clocks 11b = 6 Clocks Read to Precharge Time (Trtp)Bit 11. Specifies the read CAS# to precharge time. 0b = 2 Clocks for Burst Length of 32 Bytes 4 Clocks for Burst Length of 64 Bytes 1b = 3 Clocks for Burst Length of 32 Bytes 5 Clocks for Burst Length of 64 Bytes Minimum RAS# Active Time (Tras)Bits 1512. Specifies the minimum RAS# active time. 0000b = reserved 0001b = reserved 0010b = 5 bus clocks ... 1111b = 18 bus clocks Row Cycle Time (Trc)Bits 1916. RAS#-active to RAS#-active or auto refresh of the same bank. These bits are encoded as follows: 0000b = 11 bus clocks 0001b = 12 bus clocks ... 1110b = 25 bus clocks 1111b = 26 bus clocks Write Recovery Time (Twr)Bit 2120. Measures when the last write datum is safely registered by the DRAM. It measures from the last data to precharge (writes can go back-to-back). 00b = 3 bus clocks 01b = 4 bus clocks 10b = 5 bus clocks 11b = 6 bus clocks Active-to-active (RAS#-to-RAS#) Delay (Trrd)Bits 2322. Specifies the active-to-active delay (RAS#-to-RAS#) of different banks. 00b = 2 bus clocks Chapter 4 Memory System Configuration 105
32559
Rev. 3.08 July 2007
01b = 3 bus clocks 10b = 4 bus clocks 11b = 5 bus clocks MEMCLK disable (MemClkDis)Bits 3124. These bits disable the MEMCLK outputs for DRAM channel A. BIOS should leave all bits corresponding to unused clocks set to reduce power consumption. These bits are mapped to packages as follows:
:
Bit F(1207) Package AM2 Package

0 1 2 3 4 5 6 7 N/A N/A MA3_CLK MA2_CLK MA1_CLK MA0_CLK N/A N/A MA1_CLK1 MA0_CLK1 N/A N/A MA1_CLK0 MA0_CLK0 MA1_CLK2 MA0_CLK2
S1g1 Package
N/A MA0_CLK1 N/A N/A N/A MA0_CLK0 N/A MA0_CLK2
4.5.11
DRAM Timing High Register
This register contains the normal timing parameters specified in a DRAM data sheet. DRAM Timing High Register
31 29 28 26 25 23 22 20 19 18 17 16 15 14 13 12 11 10 9 DisAutoRefresh reserved

8 7 reserved 6 4 3 0
Trfc3
Trfc2
Trfc1
Trfc0
Tref
Trdrd Twrwr Twrrd
Twtr
TrwtTO
reserved
Bits 3129 2826 2523 2220 19 18 1716 1514 1312 1110 98
Mnemonic Trfc3 Trfc2 Trfc1 Trfc0 reserved DisAutoRefresh Tref Trdrd Twrwr Twrrd Twtr
Function Auto-Refresh Row Cycle Time for Logical DIMM3 Auto-Refresh Row Cycle Time for Logical DIMM2 Auto-Refresh Row Cycle Time for Logical DIMM1 Auto-Refresh Row Cycle Time for Logical DIMM0 Disable Automatic Refresh Refresh Rate Read to Read Timing Write to Write Timing Write to Read DIMM Termination Turnaround Internal DRAM write to Read Command Delay
R/W R/W R/W R/W R/W R R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0
106
Chapter 4
32559
Rev. 3.08
July 2007
Bits 7 64 30
Mnemonic reserved TrwtTO reserved
Function Read-to-Write Turnaround for Data, DQS Contention
R/W R R/W R
Reset 0 0 0
Field Descriptions Read-to-Write Turnaround for Data, DQS Contention (TrwtTO)Bits 64. This specifies the minimum number of cycles from the last clock of virtual CAS (defined as the clock in which CAS is asserted for the burst, N, plus the burst length, BL, minus 1; so the last clock of virtual CAS = N + BL - 1) of a first read operation to the clock in which CAS is asserted for a following write operation. Time may need to be inserted to insure there is no bus contention on bidirectional pins. 000b = 2 bus clocks 001b = 3 bus clocks 010b = 4 bus clocks 011b = 5 bus clocks 100b = 6 bus clocks 101b = 7 bus clocks 110b = 8 bus clocks 111b = 9 bus clocks Internal DRAM Write-to-Read Command Delay (Twtr)Bit 98. These bits specify the minimum write-to-read delay when both access the same chip select. 00b = reserved 01b = 1 bus clocks 10b = 2 bus clocks 11b = 3 bus clocks Write to Read DIMM Termination Turnaround (Twrrd)Bits 1110. These bits specify the minimum write-to-read delay when accessing two different DIMMs. 00b = 0 bus clocks 01b = 1 bus clocks 10b = 2 bus clocks 11b = 3 bus clocks Write to Write Timing (Twrwr)Bits 1312. This specifies the minimum number of cycles from the last clock of virtual CAS (defined as the clock in which CAS is asserted for the burst, N, plus the burst length, BL, minus 1; so the last clock of virtual CAS = N + BL - 1) of the first write-burst operation to the clock in which CAS is asserted for a following write-burst operation that changes the terminator that is enabled. Time must be inserted between consecutive writes to account for termination timing on DDR2 devices. 00b = 1 bus clocks (0 idle cycles on the bus) 01b = 2 bus clocks (1 idle cycle on the bus) Chapter 4 Memory System Configuration 107
32559
Rev. 3.08 July 2007
10b = 3 bus clocks (2 idle cycles on the bus) 11b = reserved Read to Read Timing (Trdrd)Bits 1514. This specifies the minimum number of cycles from the last clock of virtual CAS (defined as the clock in which CAS is asserted for the burst, N, plus the burst length, BL, minus 1; so the last clock of virtual CAS = N + BL - 1) of the first readburst operation to the clock in which CAS is asserted for a following read-burst operation that is to a different chip select than the first read-burst operation. Time must be inserted between consecutive reads to different chip selects to account for (1) turn-around timing and (2) termination timing. 00b = 2 bus clocks (1 idle cycle on the bus) 01b = 3 bus clocks (2 idle cycle on the bus) 10b = 4 bus clocks (3 idle cycle on the bus) 11b = 5 bus clocks (4 idle cycle on the bus) Refresh Rate (Tref)Bits 1716. This specifies the time required to refresh all DRAM devices. 00b = Undefined behavior. 01b = Reserved 10b = Refresh interval of 7.8 microseconds 11b = Refresh interval of 3.9 microseconds Disable Automatic Refresh (DisAutoRefresh)Bit 18. Setting this bit disables automatic refresh to the DIMMs. BIOS should normally leave this bit cleared. This function is sometimes useful during electrical characterization. Auto-Refresh Row Cycle Time for Logical DIMM0 (Trfc0)Bits 2220. This specifies the minimum time from an auto-refresh command to an activate command or another auto refresh command. The recommended programming of this register varies based on DRAM density and speed. 000b = 75 ns (all speeds, 256 Mbit) 001b = 105 ns all speeds, (512 Mbit) 010b = 127.5 ns (all speeds, 1 Gbit) 011b = 195 ns (all speeds, 2 Gbit) 100b = 327.5 ns (all speeds, 4 Gbit) 101b = reserved 110b = reserved 111b = reserved Auto-Refresh Row Cycle Time for Logical DIMM1 (Trfc1)Bits 2523. See Trfc0. Auto-Refresh Row Cycle Time for Logical DIMM2 (Trfc2)Bits 2826. See Trfc0. Auto-Refresh Row Cycle Time for Logical DIMM3 (Trfc3)Bits 3129. See Trfc0.
108
Chapter 4
32559
Rev. 3.08
July 2007
4.5.12
DRAM Configuration Registers
The following registers contain all the bits to configure the DRAMs for proper operation. 4.5.12.1 DRAM Configuration Low Register
This register controls drive strength, refresh, and initialization. DRAM Configuration Low Register
31 20 19 18 17 16 15 UnbuffDimm DimmEccEn reserved 12 11 10 9 BurstLength32 SelRefRateEn Width128 8 ParEn

7 DramDrvWeak 6 reserved 5 DramTerm 4 3 reserved 2 1 ExitSelfRef 0 InitDram
reserved
X4Dimm
Bits 3120 19 1817 16 1512 11 10 9 8 7 6 54 32 1 0
Mnemonic reserved DimmEccEn reserved UnbuffDimm X4Dimm Width128 BurstLength32 SelRefRateEn ParEn DramDrvWeak reserved DramTerm reserved ExitSelfRef InitDram
Function DIMM ECC Enable Unbuffered DIMM x4 DIMMs Width of DRAM Interface in 128-bit Mode DRAM Burst Length Set for 32 Bytes Faster Self Refresh Rate Enable Parity Enable DRAM Drivers Weak Mode DRAM Termination Exit Self Refresh Command Initialize DRAM
R/W R R/W R R/W R/W R/W R/W R/W R/W R/W R R/W R R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Field Descriptions Initialize DRAM (InitDram)Bit 0. Writing a 1 to this bit causes the DRAM controller to execute the DRAM initialization sequence described by the JEDEC specification. This command should be executed by BIOS when booting from an unpowered state (ACPI S4, S5 or G3; not S3 or suspend to RAM), after the DRAM controller configuration registers are properly initialized. This bit is read as a 1 while the DRAM initialization sequence is executing; it is read as 0 at all other times. The updated values of the other changed fields in this register are used in the initialization sequence when InitDram is written to 1 (they do not need to be set up correctly before writing InitDram to 1).
Chapter 4
109
32559
Rev. 3.08 July 2007
Exit Self Refresh Command (ExitSelfRef)Bit 1. Writing a 1 to this bit causes the DRAM controller to bring the DRAMs out of self refresh mode. This command should be executed by BIOS when returning from the suspend to RAM state, after the DRAM controller configuration registers are properly initialized. This bit is read as a 1 while the exit-selfrefresh command is executing; it is read as 0 at all other times. DRAM Termination (DramTerm)Bits 45. This specifies the programming of the DRAM termination value (Rtt) when the EMRS command is issued during DRAM initialization. 00b = On die termination disabled. 01b = 75 ohms. 10b = 150 ohms. 11b = 50 ohms. DRAM Drivers Weak Mode (DramDrvWeak)Bit 7. This specifies the programming of the DRAM data drive strength mode when the EMRS command is issued during DRAM initialization. 0 = Normal drive strength mode. 1 = Weak drive strength mode. Parity Enable (ParEn)Bit 8. This bit enables the DRAM command and address parity error input, M[B,A]ERR_L. Clearing this bit disables the input into the error reporting logic of the processor. Faster Self Refresh Rate Enable (SelfRefRateEn)Bit 9. This bit enables high temperature (two times normal) self refresh rate. This bit is reflected in the EMRS(2) command to the DRAM devices. DRAM Burst Length Set for 32 Bytes (BurstLength32)Bit 10. This specifies the burst length of DRAM accesses and, as a result, the number of data bytes exchanged in each access. 0 = 64-byte mode. 1 = 32-byte mode. 32-byte mode may be preferred in platforms that include graphics controllers that generate a lot of 32-byte system memory accesses. 32-byte mode is not supported when the DRAM interface is 128 bits wide; so this bit interacts with Width128 as follows:
BurstLength32 0 0 1 1 Width128 0 1 0 1 Description 8-beat burst length; 64-byte accesses 4-beat burst length; 64-byte accesses 4-beat burst length; 32-byte accesses 4-beat burst length; 64-byte accesses
Width of DRAM Interface (Width128)Bit 11. This controls the width of the DRAM interface. 0 = The controller DRAM interface is 64-bits wide 1 = The controller DRAM interface is 128-bits wide. 110 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
X4 DIMMs (X4Dimm)Bits 1512. Indicates whether each of the four logical DIMMs are x4 or not. 0 = DIMM is not x4 1 = x4 DIMM present Unbuffered DIMMs (UnBuffDimm)Bit 16. When set, the DRAM controller is only connected to unbuffered DIMMs. 0 = Buffered DIMMs 1 = Unbuffered DIMMs DIMM ECC Enable (DimmEccEn)Bit 19. ECC checking is capable of being enabled for all DIMMs on the DRAM controller (through Function 3 Offset 44h [EccEn]). This bit should not be set unless all populated DIMMs support ECC check bits. 4.5.12.2 DRAM Configuration High Register Function 2: Offset 94h
8 7 4 3 MemClkFreqVal 2 MemClkFreq 0 FourRankSODimm
DRAM Configuration High Register

31 FourActWindow 28 27 DcqBypassMax 24 23 22 21 20 19 18 17 16 15 14 13 12 11 BankSwizzleMode SlowAccessMode PowerDownMode DisDramInterface FourRankRDimm PowerDownEn
reserved
reserved
reserved
RDqsEn
reserved
MaxAsyncLat
Bits 3128 2724 22 21 20 19 18 17 16 15 14 13 12 118 74 3 20
Mnemonic FourActWindow DcqBypassMax BankSwizzleMode reserved SlowAccessMode reserved FourRankRDimm FourRankSODimm PowerDownMode PowerDownEn DisDramInterface reserved RDqsEn reserved MaxAsyncLat MemClkFreqVal MemClkFreq
Function Four Bank Activate Window DRAM Controller Queue Bypass Maximum Bank Swizzle Mode Slow Access Mode(2T Mode) Four Rank Registered DIMM Four Rank SO-DIMM Power Down Mode Power Down Mode Enable Disable the DRAM Interface Read DQS Enable Maximum Asynchronous Latency Memory Clock Frequency Valid Memory Clock Frequency
R/W R/W R/W R/W R R/W R R/W R/W R/W R/W R/W R R/W R R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Chapter 4
111
32559
Rev. 3.08 July 2007
Field Descriptions Memory Clock Frequency (MemClkFreq)Bits 20. This field specifies the frequency of the DRAM interface (MEMCLK). 000b = 200 MHz 001b = 266 MHz 010b = 333 MHz 011b = 400 MHz 1xxb = reserved Memory Clock Frequency Valid (MemClkFreqVal)Bit 3. System BIOS should set this bit when setting up MemClkFreq to the proper value. This indicates to the DRAM controller that it may start
driving MEMCLK at the proper frequency. BIOS must wait a minimum 20 MEMCLKs after setting this bit before initializing DRAM by setting InitDram (Function 2, Offset 90h), ExitSelfRef (Function 2, Offset 90h) or EnDramInit (Function 2, Offset 7Ch)
Maximum Asynchronous Latency (MaxAsyncLat)Bits 74. This field should be programmed by system BIOS to specify the maximum round trip latency in the system from the processor to the DRAM devices and back. The DRAM controller uses this to help determine when incoming DRAM read data can be safely transferred to the core clock domain. See 5.1.2.5 on page -190 for more information on how to program this field. 0000b = 0 ns ... 1111b = 15 ns Read DQS Enable (RDqsEn)Bit 12. When this bit is set, bit A11 is set when the EMRS(1) command is sent during DRAM initialization and the DRAM controller uses the DM pins as read DQS pins and disables data masking. This bit should only be set if x8 registered DIMMs are present in the system. 0 = DM pins function as data mask pins. 1 = DM pins function as read DQS pins. Disable the DRAM Interface (DisDramInterface)Bit 14. When this bit is set, the DRAM controller is disabled and the DRAM interface is placed into a low power state. This bit is normally set if there are no DIMMs connected to the DCT. To maximize power savings when this bit is set, all of the MemClkDis bits should also be set. 0 = Enabled (default) 1 = Disabled Power Down Mode Enable (PowerDownEn)Bit 15. When in power down mode, if all pages of the DRAMs associated with a CKE pin are closed, then these parts are placed in power down mode. Only pre-charge power down mode is supported, not active power down mode. 0 = Disabled (default) 1 = Enabled
112
Chapter 4
32559
Rev. 3.08
July 2007
Power Down Mode (PowerDownMode)Bit 16. This bit is initialized based on the type of system being implemented. For non-mobile systems, power down mode should be set to channel CKE control. A DIMM or a group of DIMMs enters power down mode by deasserting the corresponding clock enable signal when the DRAM controller detects that there are no transactions scheduled to any of the DIMMs connected to the clock enable signal. A DIMM or a group of DIMMs exits power down mode by asserting the corresponding clock enable signal when a transaction is scheduled to any DIMM connected to the clock enable signal. 0 = Channel CKE control. The DRAM channel is placed in power down when all chip selects associated with the channel are idle. 1 = Chip Select CKE control. A chip select or pair of chip selects is placed in power down when no transactions are pending for the chip select(s). Four Rank SO-DIMM (FourRankSODimm)Bit 17. This bit is set by BIOS to indicate that a four rank SO-DIMM is present. Four Rank Registered DIMM (FourRankRDimm)Bit 18. This bit is set by BIOS to indicate that a four rank registered DIMM is present. Slow Access Mode(2T Mode) (SlowAccessMode)Bit 20. This bit controls when 2T timing is used. 2T mode may be needed in order to meet electrical requirements of certain DIMM speed and loading configurations. 0 = DRAM address and control signals are driven for one MEMCLK cycle. 1 = One additional MEMCLK of setup time is provided on all DRAM address and control signals except CS, CKE, and ODT; i.e., these signals are driven for two MEMCLK cycles rather than one. Bank Swizzle Mode (BankSwizzleMode)Bit 22. This bit remaps the DRAM device bank address bits as a function of normalized physical address bits. Each of the bank address bits, as specified in Table 10 and Table 11, are remapped as follows: \
64-Bit Interface Bank Address 0 BA0 ^ A17 ^ A22 Bank Address 1 BA1 ^ A18 ^ A23 Bank Address 2 BA2 ^ A19 ^ A24 128-Bit Interface BA0 ^ A18 ^ A23 BA1 ^ A19 ^ A24 BA2 ^ A20^ A25
For example, encoding 06h of Table 10 would be remapped from bank[1:0]={A15, A14} to the following for a 64-bit DCT: Bank[1:0] = {A15 ^ A18 ^ A23, A14^ A17 ^ A22}. 0 = Disabled (default) 1 = Enabled Note: ^ is the logical exclusive-OR operator.
Chapter 4
113
32559
Rev. 3.08 July 2007
DRAM Controller Queue Bypass Maximum (DcqBypassMax)Bits 2724. The DRAM controller arbiter normally allows transactions to pass other transactions in order to optimize DRAM bandwidth. This field specifies the maximum number of times that the oldest memory-access request in the DRAM controller queue may be bypassed before the arbiter decision is overridden and the oldest memory-access request is serviced instead. 0000b = No bypass; the oldest request is never bypassed. 0001b = The oldest request may be bypassed no more than 1 time. ... ... 1111b = The oldest request may be bypassed no more than 15 times. Four Bank Activate Window (FourActWindow)Bits 3128. This specifies the rolling tFAW window during which no more than 4 banks in a 8-bank device are activated, per JEDEC DDR2 specification. For example, if this is set to 10 clocks and an activate command is issued in clock N, then no more than three further activate commands may be issued in clocks N+1 through N+9. To program this field, BIOS must convert the tFAW parameter into MEMCLK cycles by dividing the highest tFAW (in ns) across all DIMMs connected to the channel by the lowest period (highest frequency) of MEMCLK (in ns) over all P-states and rounding up to the next integer. 0000b = No tFAW window restriction. 0001b = 8 MEMCLK cycles 0010b = 9 MEMCLK cycles .... 1101b = 20 MEMCLK cycles 111xb = Reserved
4.5.13
DRAM Controller Additional Data Offset Register
This register is used to select the DRAM controller register that is accessed when reading or writing Function 2 Offset 9Ch. Dram Controller Additional Data Offset Register
31 30 29 DctAccessDone DctAccessWrite

0
DctOffset
Bits 31 30 290
Mnemonic DctAccessDone DctAccessWrite DctOffset
Function DRAM Controller Access Done DRAM Controller Read/Write Select DRAM Controller Offset
R/W R W R/W
Reset 1 0 0
114
Chapter 4
32559
Rev. 3.08
July 2007
Field Descriptions Dram Controller Offset (DctOffset)Bits 290. This field is used to select the DRAM controller register that is accessed when reading or writing Function 2 Offset 9Ch. Dram Controller Read/Write Select (DctAccessWrite)Bit 30. This bit indicates if the access to Function 2 Offset 9Ch is a read or write. 0b = Read access. 1b = Write access. Dram Controller Access Done (DctAccessDone)Bit 31. This bit indicates if the access to Function 2 Offset 9Ch is complete. 0b = Access in progress. 1b = No access is progress.
Table 14. DctOffset Field Encodings
DctOffset Channel A DctOffset Channel B Register
0000_0000h 0000_0020h Output Driver Compensation Control 0000_0001h 0000_0021h Write Data Timing Low 0000_0002h 0000_0022h Write Data Timing High 0000_0003h 0000_0023h Write ECC Timing 0000_0004h 0000_0024h Address/Command Timing Control 0000_0005h 0000_0025h Read DQS Timing Control Low 0000_0006h 0000_0026h Read DQS Timing Control High 0000_0007h 0000_0027h Read DQS ECC Timing Control 0000_0010h 0000_0030h DQS Receiver Enable Timing Control Chip Selects M[B:A]0_CS_L[1:0] 0000_0011h 0000_0031h Reserved DQS Receiver Enable Timing Control Chip Selects M[B:A]1_CS_L[1:0]
Note:
0000_0012h 0000_0032h Reserved 0000_0013h 0000_0033h
0000_0014h 0000_0034h Reserved 0000_0015h 0000_0035h Reserved 0000_0016h 0000_0036h DQS Receiver Enable Timing Control Chip Selects M[B:A]2_CS_L[1:0] 0000_0017h 0000_0037h Reserved 0000_0018h 0000_0038h Reserved 0000_0019h 0000_0039h DQS Receiver Enable Timing Control Chip Selects M[B:A]3_CS_L[1:0] 0000_001Ah 0000_003Ah Reserved 0000_001Bh 0000_003Bh Reserved
Note: For S1g1 Package this register controls the DQS Receiver Enable Timing for Chip Selects M[B:A]0_CS_L[3:2]
Chapter 4
115
32559
Rev. 3.08 July 2007
4.5.14
DRAM Controller Additional Data Port Register
This register is used to access the DRAM controller register that is selected by Function 2 Offset 98h. Access to these registers is accomplished as follows: Reads: Write the register number to Function 2 Offset 98h with Function 2 Offset 98h (DctAccessWrite)=0. Poll Function 2 Offset 98h (DctAccessDone) until it is high. Read the register contents from Function 2 Offset 9Ch. Writes: Write the register data to Function 2 Offset 9Ch. Write the register number to Function 2 Offset 98h (DctOffset) with Function 2 Offset 98h (DctAccessWrite)=1. Poll Function 2 Offset 98h (DctAccessDone) until it is high to ensure that the contents of the write have been delivered to the phy.
All software access to this register must be doubleword access. Byte and word access to this register are not supported. The MemClkFreqVal bit (Function 2, Offset 94h) must be set before attempting to read or write the DRAM Controller Additional Data registers. DRAM Controller Additional Data Port Register
31 DctOffsetData

0
Bits 310
Mnemonic DctOffsetData
Function DRAM Controller Offset Data
R/W R/W
Reset 0
Field Descriptions Dram Controller Offset Data (DctOffsetData)Bits 310. This field is used to access the DRAM controller register that is selected by Function 2 Offset 98h.
4.5.15
Output Driver Compensation Control Register
This register is used to program the drive strength and termination resistance for the DRAM channel.
116
Chapter 4
32559
Rev. 3.08
July 2007
Output Driver Compensation Control Register

31 30 29 28 27 22 21 20 19 18 17 16 15 14 13 12 11 10 9 AddrCmdDrvStren DataDrvStren DqsDrvStren ClkDrvStren
Function 2: Offset 9Ch Index 00h, 20h

8 7 6 5 CsOdtDrvStren 4 3 2 1 CkeDriveStren 0
reserved
reserved
reserved
reserved
reserved
reserved
Bits 3130 2928 2722 2120 1918 1716 1514 1312 1110 98 76 54 32 10
Mnemonic reserved ProcOdt reserved DqsDrvStren reserved DataDrvStren reserved ClkDrvStren reserved AddrCmdDrvStren reserved CsOdtDrvStrength reserved CkeDreStrength
Function Processor On-die Termination DQS Drive Strength Data Drive Strength Clock Drive Strength Address/Command Drive Strength CS/ODT Drive Strength CKE Drive Strength
R/W R R/W R R/W R R/W R R/W R R/W R R/W R R/W
Reset 0 0 0 01b 0 01b 0 01b 0 10b 0 10b 0 10b
Field Descriptions CKE Drive Strength (CkeDrvStren)Bits 10. This field is used to select the drive strength of the CKE pins. 00b = 1.0x. 01b = 1.25x 10b = 1.5x (Default) 11b = 2.0x CS/ODT Drive Strength (CsOdtDrvStren)Bits 54. This field is used to select the drive strength of the CS and ODT pins. 00b = 1.0x. 01b = 1.25x 10b = 1.5x (Default) 11b = 2.0x Address/Command Drive Strength (AddrCmdDrvStren)Bits 98. This field is used to select the drive strength of the address, RAS, CAS, WE, bank and parity pins. Chapter 4 Memory System Configuration 117
reserved
ProcOdt
32559
Rev. 3.08 July 2007
00b 01b 10b 11b
= 1.0x. = 1.25x = 1.5x (Default) = 2.0x
MEMCLK Drive Strength (ClkDrvStren)Bits 1312. This field is used to select the drive strength of the MEMCLK pins. 00b = 0.75x. 01b = 1.0x (Default) 10b = 1.25x 11b = 1.5x Data Drive Strength (DataDrvStren)Bits 1716. This field is used to select the drive strength of the Data pins. 00b = 0.75x. 01b = 1.0x (Default) 10b = 1.25x 11b = 1.5x DQS Drive Strength (DqsDrvStren)Bits 2120. This field is used to select the drive strength of the DQS pins. 00b = 0.75x. 01b = 1.0x (Default) 10b = 1.25x 11b = 1.5x Note: The DM[8:0] and DQS[17:9] functions share pins on the DIMM connector. The function selection is applied based on whether the DIMM is populated with by-4 (x4) DRAM devices, in which case the DQS[17:9] function is applied, or not (x8 or x16 DRAM devices), in which case the DM[8:0] function is applied. However, the DM function is associated with the data pin group and should therefore be controlled DataDrvStren. While the processor supports concurrent population of x4 and non-x4 DIMMs, the determination as to which field controls the drive strength of these pins is applied statically based on these rules: - If all DIMMs of a channel are populated with non-x4 devices, DataDrvStren is applied. - If any DIMMs of a channel is populated with x4 devices, DqsDrvStren is applied. Processor On-die Termination (ProcOdt)Bits 2928. This field is used to select the resistance of the on-die termination resistors. 00b = 300 ohms +/- 20% 01b = 150 ohms +/- 20%. 10b = 75 ohms +/- 20%. 11b = Reserved.
118
Chapter 4
32559
Rev. 3.08
July 2007
4.5.16
Write Data Timing Control Registers
These registers control the timing of write data with respect to DQS. Write Data Timing Low Timing Control Register
31 30 29 reserved 24 23 22 21 reserved 16 15 14 13 reserved WrDatTimeByte1

8 7 reserved 6 5 WrDatTimeByte0 0
WrDatTimeByte3
WrDatTimeByte2
Bits 3130 2924 2322 2116 1514 138 76 50
Mnemonic reserved WrDatTimeByte3 WrDatTimeByte2 reserved WrDatTimeByte1 reserved WrDatTimeByte0
Function Write Data Byte 3 Timing Control Write Data Byte 2 Timing Control Write Data Byte 1 Timing Control Write Data Byte 0 Timing Control
R/W R R/W R R/W R R/W R R/W
Reset 0 17h 0 17h 0 17h 0 17h
Field Descriptions Write Data Byte 0 Timing Control (WrDatTimeByte0)Bits 50. These bits select how much delay is added to byte 0 of the data with respect to DQS. 000000b = no delay 000001b = 1/96 MEMCLK delay 000010b = 2/96 MEMCLK delay ... 101111b = 47/96 MEMCLK delay 11xxxxb = reserved Write Data Byte 1 Timing Control (WrDatTimeByte1)Bits 138. These bits select how much delay is added to byte 1 of the data with respect to DQS. See WrDatTimeByte0 for bit definition. Write Data Byte 2 Timing Control (WrDatTimeByte2)Bits 2116. These bits select how much delay is added to byte 2 of the data with respect to DQS. See WrDatTimeByte0 for bit definition. Write Data Byte 3 Timing Control (WrDatTimeByte3)Bits 2924. These bits select how much delay is added to byte 3 of the data with respect to DQS. See WrDatTimeByte0 for bit definition.
Chapter 4
119
32559
Rev. 3.08 July 2007
Write Data Timing High Timing Control Register

31 30 29 reserved 24 23 22 21 reserved 16 15 14 13 reserved WrDatTimeByte5

8 7 reserved 6 5 WrDatTimeByte4 0
WrDatTimeByte7
WrDatTimeByte6
Bits 3130 2924 2322 2116 1514 138 76 50
Mnemonic reserved WrDatTimeByte7 WrDatTimeByte6 reserved WrDatTimeByte5 reserved WrDatTimeByte4
Function Write Data Byte 7 Timing Control Write Data Byte 6 Timing Control Write Data Byte 5 Timing Control Write Data Byte 4 Timing Control
Reset 0 17h 0 17h 0 17h 0 17h
Field Descriptions Write Data Byte 4 Timing Control (WrDatTimeByte4)Bits 50. These bits select how much delay is added to byte 4 of the data with respect to DQS. 000000b = no delay 000001b = 1/96 MEMCLK delay 000010b = 2/96 MEMCLK delay ... 101111b = 47/96 MEMCLK delay 11xxxxb = reserved Write Data Byte 5 Timing Control (WrDatTimeByte5)Bits 138. These bits select how much delay is added to byte 5 of the data with respect to DQS. See WrDatTimeByte4 for bit definition. Write Data Byte 6 Timing Control (WrDatTimeByte6)Bits 2116. These bits select how much delay is added to byte 6 of the data with respect to DQS. See WrDatTimeByte4 for bit definition. Write Data Byte 7 Timing Control (WrDatTimeByte7)Bits 2924. These bits select how much delay is added to byte 7 of the data with respect to DQS. See WrDatTimeByte4 for bit definition
120
Chapter 4
32559
Rev. 3.08
July 2007
Write Data ECC Timing Control Register

31 reserved

6 5 WrChkTime 0
Bits 316 50
Mnemonic reserved WrChkTime
Function Write Data ECC Timing Control
R/W R R/W
Reset 0 17h
Field Descriptions Write Data ECC Timing Control (WrChkTime)Bits 50. These bits select how much delay is added to the ECC with respect to DQS. 000000b = no delay 000001b = 1/96 MEMCLK delay 000010b = 2/96 MEMCLK delay ... 101111b = 47/96 MEMCLK delay 11xxxxb = reserved
4.5.17
Address Timing Control Register
This register controls the time of the address, command, chip select, ODT and clock enable pins with respect to MEMCLK. Address Timing Control Register
31 29 28 27 AtcDllMaxPhases 22 21 20 AddrCmdFineDelay AddrCmdSetup 16 15 14 13 12 CsOdtSetup

8 7 6 5 4 0
reserved
reserved
CsOdtFineDelay
CkeSetup 0 0 0 0 0
reserved
reserved
CkeFineDelay
Bits 3129 28 2722 21 2016
Mnemonic reserved AtcDllMaxPhases reserved AddrCmdSetup AddrCmdFineDelay
Function Address Timing Control DLL Phases, index 04h Reserved, index 24h Address/Command Setup Time Address/Command Fine Delay
R/W R R/W R R/W R/W
Reset
Chapter 4
121
32559
Rev. 3.08 July 2007
Bits 1514 13 128 76 5 40
Mnemonic reserved CsOdtSetup CsOdtFineDelay reserved CkeSetup CkeFineDelay
Function CS/ODT Setup Time CS/ODT Fine Delay CKE Setup Time CKE Fine Delay
R/W R R/W R/W R R/W R/W
Reset 0 0 0 0 0 0
Field Descriptions CKE Fine Delay (CkeFineDelay)Bits 40. These bits control how long the CKE pins are delayed from the default setup time. 00000b = no delay 00001b = 1/64 MEMCLK delay 00010b = 2/64 MEMCLK delay ... 11111b = 31/64 MEMCLK delay CKE Setup Time (CkeSetup)Bit 5. This bit selects the default setup time for the CKE pins versus MEMCLK. 0 = 1/2 MEMCLK 1 = 1 MEMCLK CS/ODT Fine Delay (CsOdtFineDelay)Bits 128. These bits control how long the CS and ODT pins are delayed from the default setup time. 00000b = no delay 00001b = 1/64 MEMCLK delay 00010b = 2/64 MEMCLK delay ... 11111b = 31/64 MEMCLK delay CS/ODT Setup Time (CsOdtSetup)Bit 13. This bit selects the default setup time for the CS and ODT pins versus MEMCLK. 0 = 1/2 MEMCLK 1 = 1 MEMCLK) Address/Command Fine Delay (AddrCmdFineDelay)Bits 2016. These bits control how long the address and command pins are delayed from the default setup time. 00000b = no delay 00001b = 1/64 MEMCLK delay 00010b = 2/64 MEMCLK delay ... 11111b = 31/64 MEMCLK delay
122
Chapter 4
32559
Rev. 3.08
July 2007
Address/Command Setup Time (AddrCmdSetup)Bit 21. This bit selects the default setup time for the address and command pins versus MEMCLK. 0 = 1/2 MEMCLK (1 1/2 MEMCLK for 2T timing) 1 = 1 MEMCLK(2 MEMCLKs for 2T timing) Address Timing Control DLL Phases(AtcDllMaxPhases)Bit 28. This bit selects the number of phases for the 8 DLLs controlling the timing of the address-command (AddrCmd, CsOdt, Cke) interface group. This bit exists only in index 04h and controls the DLL phases for both channel A and B. BIOS should program this bit prior to DRAM training. It is recommended that BIOS set this bit to 1b. 0 = 32 phases per bit-time 1 = 64 phases per bit-time
4.5.18
Read DQS Timing Control Registers
These registers control how much the read DQS is delayed with respect to the data. Read DQS Timing Low Timing Control Register
31 30 29 reserved 24 23 22 21 reserved 16 15 14 13 reserved RdDqsTimeByte1

8 7 reserved 6 5 RdDqsTimeByte0 0
RdDqsTimeByte3
RdDqsTimeByte2
Bits 3130 2924 2322 2116 1514 138 76 50
Mnemonic reserved RdDqsTimeByte3 RdDqsTimeByte2 reserved RdDqsTimeByte1 reserved RdDqsTimeByte0
Function Read DQS Byte 3 Timing Control Read DQS Byte 2 Timing Control Read DQS Byte 1 Timing Control Read DQS Byte 0 Timing Control
R/W R R/W R/W R R/W R R/W
Reset 0 17h 17h 0 17h 0 17h
Field Descriptions Read DQS Byte 0 Timing Control (RdDqsTimeByte0)Bits 50. These bits select how much delay is added to the DQS associated with byte 0 with respect to the data. 000000b = no delay 000001b = 1/96 MEMCLK delay 000010b = 2/96 MEMCLK delay ... 101111b = 47/96 MEMCLK delay 11xxxxb = reserved Chapter 4 Memory System Configuration 123
32559
Rev. 3.08 July 2007
Read DQS Byte 1 Timing Control (RdDqsTimeByte1)Bits 138. These bits select how much delay is added to the DQS associated with byte 1 with respect to the data. See RdDqsTimeByte0 for bit definition. Read DQS Byte 2 Timing Control (RdDqsTimeByte2)Bits 2116. These bits select how much delay is added to the DQS associated with byte 2 with respect to the data. See RdDqsTimeByte0 for bit definition. Read DQS Byte 3 Timing Control (RdDqsTimeByte3)Bits 2924. These bits select how much delay is added to the DQS associated with byte 3 with respect to the data. See RdDqsTimeByte0 for bit definition. Read DQS Timing High Timing Control Register
31 30 29 reserved 24 23 22 21 reserved 16 15 14 13 reserved RdDqsTimeByte5

8 7 reserved 6 5 RdDqsTimeByte4 0
RdDqsTimeByte7
RdDqsTimeByte6
Bits 3130 2924 2322 2116 1514 138 76 50
Mnemonic reserved RdDqsTimeByte7 RdDqsTimeByte6 reserved RdDqsTimeByte5 reserved RdDqsTimeByte4
Function Read DQS Byte 7 Timing Control Read DQS Byte 6 Timing Control Read DQS Byte 5 Timing Control Read DQS Byte 4 Timing Control
R/W R R/W R/W R R/W R R/W
Reset 0 17h 0 17h 0 17h 0 17h
Field Descriptions Read DQS Byte 4 Timing Control (RdDqsTimeByte4)Bits 50. These bits select how much delay is added to the DQS associated with byte 4 with respect to the data. 000000b = no delay 000001b = 1/96 MEMCLK delay 000010b = 2/96 MEMCLK delay ... 101111b = 47/96 MEMCLK delay 11xxxxb = reserved Read DQS Byte 5 Timing Control (RdDqsTimeByte5)Bits 138. These bits select how much delay is added to the DQS associated with byte 5 with respect to the data. See RdDqsTimeByte4 for bit definition.
124
Chapter 4
32559
Rev. 3.08
July 2007
Read DQS Byte 6 Timing Control (RdDqsTimeByte6)Bits 2116. These bits select how much delay is added to the DQS associated with byte 6 with respect to the data. See RdDqsTimeByte4 for bit definition. Read DQS Byte 7 Timing Control (RdDqsTimeByte7)Bits 2924. These bits select how much delay is added to the DQS associated with byte 7 with respect to the data. See RdDqsTimeByte4 for bit definition. Read DQS ECC Timing Control Register
31 reserved

6 5 RdDqsTimecheck 0
Bits 316 50
Mnemonic reserved RdDqsTimeCheck
Function Read DQS ECC Byte Timing Control
R/W R R/W
Reset 0 17h
Field Descriptions Read DQS ECC Byte Timing Control (RdDqsTimeCheck)Bits 50. These bits select how much delay is added to the DQS associated with the ECC byte with respect to the data. 000000b = no delay 000001b = 1/96 MEMCLK delay 000010b = 2/96 MEMCLK delay ... 101111b = 47/96 MEMCLK delay 11xxxxb = reserved
4.5.19
DQS Receiver Enable Timing Control Registers
These registers specify the delay in picoseconds from the assertion of MEMCLK at the CPU to the driving of the read preamble by the DIMM as perceived by the CPU. Writing to the DQS receiver enable timing causes the read and write pointers in the DRAM controller FIFOs to be reset. The DQS receiver enable delays programmed must meet the following restrictions: The maximum delay between two DIMMs that comprise a logical 128 bit DIMM = 1 MEMCLK. The delays for unused DIMMs should be programmed to 00h.
Chapter 4
125
32559
Rev. 3.08 July 2007
DQS Receiver Enable Timing Control Register Function 2: Offset 9Ch Index 10h, 13h, 16h, 19h, 30h, 33h, 36h, 39h
31 reserved 8 7 DqsRcvEnDelay 0
Bits 318 70
Mnemonic DqsRcvEnDelay
Function reserved DQS Receiver Enable Delay
R/W R R/W
Reset 0 0
Field Descriptions DQS Receiver Enable Delay (DqsRcvEnDelay)Bits 70. These bits specify the delay in 50 picosecond increments from the rising edge of MEMCLK at the CPU to the driving of the read preamble by the DIMM as perceived at the CPU. 00h = 0 ps 01h = 50 ps 02h = 100 ps ... AEh = 8.70 ns AFhFFh = Reserved
4.5.20
DRAM Controller Miscellaneous Register

Function 2: Offset A0h
11 10 9 PwrSavingsEn DramEnabled 8 ILD_lmt 6 5 DCC_EN 4 Mod64Mux 3 RdWrQByp 2 1 Disablejitter 0 reserved
DRAM Controller Miscellaneous Register

31 24 23
MemClkDis
reserved
Bits 3124 2311 10 9 86 5 4 32 1 0
Mnemonic MemClkDis reserved PwrSavingsEn DramEnabled ILD_lmt DCC_EN Mod64Mux RdWrQByp DisableJitter reserved
Function Memory Clock Disable Dram Power Savings Enable Dram enabled Idle Cycle Limit Dynamic Idle Cycle Counter Enable Mismatched DIMM Support Enable Read/Write Queue Bypass Count Disable Jitter
R/W R/W R R/W R R/W R/W R/W R/W R/W R
Reset FFh 0 0 0 0 0 0 0 0 X
126
Chapter 4
32559
Rev. 3.08
July 2007
Field Descriptions Disable Jitter (DisableJitter)Bit 1. When set the DDR compensation circuit will not change values unless the change is more than one step from the current value. Read/Write Queue Bypass Count (RdWrQByp)Bits 23. Specifies the number of times the oldest operation in the DCI read/write queue can be bypassed before the arbiter is overridden and the oldest operation is chosen. 00b = 2 01b = 4 10b = 8 11b = 16 Mismatched DIMM Support Enable (Mod64BitMux)Bit 4. This bit enables support for mismatched DIMMs when using 128-bit DRAM interface. When this bit is set, the Width128 bit has no effect. This bit only applies to the AM2 and S1g1 packages. This mode cannot be used with on-line spare. Dynamic Idle Cycle Counter Enable (DCC_EN)Bit 5. When set to 1, indicates that each entry in the page table dynamically adjusts the idle cycle limit based on Page Conflict/Page Miss (PC/PM) traffic. Idle Cycle Limit (ILD_lmt)Bits 86. Specifies the number of MemCLKs before forcibly closing (precharging) an open page. If DCC_EN (Function 2, Offset A0h) has a value of 0, the static counters are loaded with the ILD_lmt and decremented each clock. If DCC_EN has a value of 1, the dynamic counters are loaded with the ILD_lmt and modified as follows: IncrementWhen a Page Miss (PM) page hits on an invalid entry in the Page Table. The presumption is that in the past, that page table entry was occupied by the very same page that has a Page Miss. Had the old page been kept open longer, it would have been a Page Hit. Increment the Idle Cycle Limit count to increase the probability of getting a future Page Hit. DecrementWhen a Page Conflict (PC) arrives and hits on an idle entry (obviously an open page). This Page Conflict can be avoided if the open page is closed earlier. Decrement the Idle Cycle Limit count to increase the probability of avoiding a future Page Conflict. 000b = 0 cycles 001b = 4 cycles 010b = 8 cycles 011b = 16 cycles 100b = 32 cycles 101b = 64 cycles 110b = 128 cycles 111b = 256 cycles DRAM Enabled (DramEnabled)Bit 9. When set, this bit indicates that DRAM is enabled. This bit is set by hardware after DRAM initialization or on an exit from self refresh. The DRAM controller is initialized after the hardware-controlled initialization process (initiated by Chapter 4 Memory System Configuration 127
32559
Rev. 3.08 July 2007
Function 2: Offset 90h[DramInit]) completes or when the BIOS-controlled initialization process completes (Function 2: Offset 7Ch [EnDramInit] is written from 1 to 0). DRAM Power Savings Enable (PwrSavingsEn)Bit 10. When set, the DRAM DLLs and the on die DRAM termination are disabled when the DRAM is placed in self refresh. MEMCLK disable (MemClkDis)Bits 3124. These bits disable the MEMCLK outputs for DRAM channel B. BIOS should leave all bits corresponding to unused clocks set to reduce power consumption. These bits are mapped to packages as follows:
:
Bit F(1207) Package AM2 Package

0 1 2 3 4 5 6 7 N/A N/A MB3_CLK MB2_CLK MB1_CLK MB0_CLK N/A N/A MB1_CLK1 MB0_CLK1 N/A N/A MB1_CLK0 MB0_CLK0 MB1_CLK2 MB0_CLK2
S1g1 Package
N/A MB0_CLK1 N/A N/A N/A MB0_CLK0 N/A MB0_CLK2
4.6
Function 3Miscellaneous Control
This function is a collection of the remaining memory system configuration registers. As listed in Table 15, Function 3 includes the following registers: Machine check architecture for Northbridge errors DRAM scrubber Buffer counts for Northbridge flow control Power management GART Clocking/FIFO control Thermtrip status Northbridge capabilities
128
Chapter 4
32559
Rev. 3.08
July 2007
Table 15.
Offset 00h 04h 08h

Register Name Device ID Status Base Class Code BIST
1
Reset Vendor ID (AMD) Command 1103_1022h 0000_0000h Revision ID Cache Line Size 0600_0000h
Access RO RO RO
0Ch 10h 14h 18h 1Ch 20h 24h 28h 2Ch 30h 34h 38h 3Ch 40h 44h 48h 4Ch 50h 54h 58h 5Ch 60h 64h 68h 70h 74h
0080_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h
RO RO RO RO RO RO RO RO RO RO RO RO RO RW RW RW RW RW RW RW RW RW RW RW RW RW
page 131
Base Address 0 Base Address 1 Base Address 2 Base Address 3 Base Address 4 Base Address 5 Card Bus CIS Pointer Sub System ID ROM Base Address Capabilities reserved Max Latency Min GNT Int Pin Int Line Sub system Vendor ID
0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h page 139 page 142 page 145 page 145 0000_0000h page 153 page 154 page 154 page 155 5102_0111h 5000_8011h
MCA Northbridge Control MCA Northbridge Configuration MCA Northbridge Status Low MCA Northbridge Status High MCA Northbridge Address Low MCA Northbridge Address High Scrub Control DRAM Scrub Address Low DRAM Scrub Address High HTC Thermal Control SRI-to-XBAR Buffer Counts XBAR-to-SRI Buffer Counts
page 133 page 135 page 139 page 142 page 145 page 145 page 151 page 153 page 154 page 154 page 155 page 161 page 163
Chapter 4
129
32559
Rev. 3.08 July 2007
Table 15.
Offset 78h 7Ch 80h 84h 90h 94h 98h 9Ch A0h B0h D4h D8h DCh E4h E8h FCh

Register Name MCT-to-XBAR Buffer Counts Free List Buffer Counts Power Management Control Low Power Management Control High GART Aperture Control GART Aperture Base GART Table Base GART Cache Control Power Control Miscellaneous On-Line Spare Control Register Clock Power/Timing Low Clock Power/Timing High HyperTransport Read Pointer Optimization Thermtrip Status Northbridge Capabilities CPUID Family Model Reset 0800_3800h 0000_221Bh 0000_0000h 0000_0000h 0000_0000h page 169 page 170 0000_0000h 0000_0000h 0000_0000h page 174 page 176 page 177 0000_0000h 0000_0000h page 183 Access RW RW RW RW RW RW RW RW RW R/W RW RW RW RO RO R Description page 162 page 163 page 167 page 167 page 168 page 169 page 170 page 171 page 171 page 172 page 174 page 176 page 177 page 179 page 182 page 183
4.6.1

0
Bits 3116 150
R/W R R
Reset 1103h 1022h
130
Chapter 4
32559
Rev. 3.08
July 2007
4.6.2
31 BCC 24 23 SCC 16 15 PI

8 7 RevID 0
Bits 3124 2316 158 70
R/W R R R R
4.6.3

8 7 CLS 0
Chapter 4
131
32559
Rev. 3.08 July 2007
Bits 3124 2316 158 70
R/W R R R R
Field Descriptions CacheLineSize (CLS)Bits 70. This read-only value is defined as 00h. LatencyTimer (LatTimer)Bits 158. This read-only value is defined as 00h. HeaderType (Htype)Bits 2316. This read-only value is defined as 80h to indicate that multiple functions are present in the configuration header and that the header layout corresponds to a device header as opposed to a bridge header. BISTBits 3124. This read-only value is defined as 00h.
4.6.4
Machine Check Architecture (MCA) Registers
These registers are used to configure the Machine Check Architecture (MCA) functions of the onchip Northbridge (NB) hardware and to provide a method for the Northbridge hardware to report errors in a way compatible with MCA. All of the Northbridge MCA registers, except for the MCA NB Configuration register, are accessible through the MCA-defined MSR method, as well as through PCI configuration space. The MCA NB Configuration register is accessible only through PCI configuration space. The MCA NB Control register enables MCA reporting of each error checked by the Northbridge. The global MCA error enables must also be set through the global MCA MSRs. The error enables in this register only affect error reporting through MCA. Actions which the Northbridge may take in addition to MCA reporting are enabled through the MCA NB Configuration register. The MCA NB Status registers and MCA NB Address registers provide status and address information regarding an error which the Northbridge has logged. The values of error-logging registers are maintained through a warm reset allowing software to identify an error source that resulted in systemwide initialization.
132
Chapter 4
32559
Rev. 3.08
July 2007
4.6.4.1
MCA NB Control Register Function 3: Offset 40h

19 18 17 DramParEn 13 12 11 10 9 WchDogTmrEn AtomicRMWEn GartTblWkEn TgtAbrtEn 8 MstrAbrtEn 7 SyncPkt2En 6 SyncPkt1En 5 SyncPkt0En 4 CrcErr2En 3 CrcErr1En 2 CrcErr0En 1 UnCorrEccEn 0 CorrEccEn
MCA NB Control Register

31
reserved
reserved
Bits 3119 18 1713 12 11 10 9 8 7 6 5 4 3 2 1 0
Mnemonic reserved DramParEn reserved WchDogTmrEn AtomicRMWEn GartTblWkEn TgtAbrtEn MstrAbrtEn SyncPkt2En SyncPkt1En SyncPkt0En CrcErr2En CrcErr1En CrcErr0En UnCorrEccEn CorrEccEn
Function DRAM Parity Error Reporting enable Watchdog Timer Error Reporting Enable Atomic Read-Modify-Write Error Reporting Enable GART Table Walk Error Reporting Enable Target Abort Error Reporting Enable Master Abort Error Reporting Enable HyperTransport Link 2 Sync Packet Error Reporting Enable HyperTransport Link 1 Sync Packet Error Reporting Enable HyperTransport Link 0 Sync Packet Error Reporting Enable HyperTransport Link 2 CRC Error Reporting Enable HyperTransport Link 1 CRC Error Reporting Enable HyperTransport Link 0 CRC Error Reporting Enable Uncorrectable ECC Error Reporting Enable Correctable ECC Error Reporting Enable
R/W R R/W R R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Field Descriptions Correctable ECC Error Reporting Enable (CorrEccEn)Bit 0. Enables MCA reporting of correctable ECC errors which are detected in the Northbridge. Since correctable errors do not result in an MCA exception, the effect of this bit is limited to the setting of the error enable bit in the MCA NB Status High register. Uncorrectable ECC Error Reporting Enable (UnCorrEccEn)Bit 1. Enables MCA reporting of uncorrectable ECC errors which are detected in the Northbridge. Note that in some cases data may be forwarded to the CPU core prior to checking ECC in which case the check takes place in one of the other error reporting banks. HyperTransport Link 0 CRC Error Reporting Enable (CrcErr0En)Bit 2. Enables MCA reporting of CRC errors detected on HyperTransport link 0. The Northbridge floods its Chapter 4 Memory System Configuration 133
32559
Rev. 3.08 July 2007
outgoing HyperTransport links with sync packets after detecting a CRC error on an incoming link independent of the state of this bit. HyperTransport Link 1 CRC Error Reporting Enable (CrcErr1En)Bit 3. Enables MCA reporting of CRC errors detected on HyperTransport link 1. The Northbridge floods its outgoing HyperTransport links with sync packets after detecting a CRC error on an incoming link independent of the state of this bit. HyperTransport Link 2 CRC Error Reporting Enable (CrcErr2En)Bit 4. Enables MCA reporting of CRC errors detected on HyperTransport link 2. The Northbridge floods its outgoing HyperTransport links with sync packets after detecting a CRC error on an incoming link independent of the state of this bit. HyperTransport Link 0 Sync Packet Error Reporting Enable (SyncPkt0En)Bit 5. Enables MCA reporting of HyperTransport sync error packets detected on HyperTransport link 0. The Northbridge floods its outgoing HyperTransport links with sync packets after detecting a sync packet on an incoming link independent of the state of this bit. HyperTransport Link 1 Sync Packet Error Reporting Enable (SyncPkt1En)Bit 6. Enables MCA reporting of HyperTransport sync error packets detected on HyperTransport link 1. The Northbridge floods its outgoing HyperTransport links with sync packets after detecting a sync packet on an incoming link independent of the state of this bit. HyperTransport Link 2 Sync Packet Error Reporting Enable (SyncPkt2En)Bit 7. Enables MCA reporting of HyperTransport sync error packets detected on HyperTransport link 2. The Northbridge floods its outgoing HyperTransport links with sync packets after detecting a sync packet on an incoming link independent of the state of this bit. Master Abort Error Reporting Enable (MstrAbrtEn)Bit 8. Enables MCA reporting of master aborts (HyperTransport packets that return an error status with the Non Existent Address bit set). The Northbridge returns an error response back to the requestor with any associated data all 1s independent of the state of this bit. Target Abort Error Reporting Enable (TgtAbrtEn)Bit 9. Enables MCA reporting of target aborts (HyperTransport technology packets that return an error status with the Non Existent Address bit clear). The Northbridge returns an error response back to the requestor with any associated data all 1s independent of the state of this bit. GART Table Walk Error Reporting Enable (GartTblWkEn)Bit 10. Enables MCA reporting of GART cache table walks which encounter a GART PTE entry which is invalid. Atomic Read-Modify-Write Error Reporting Enable (AtomicRMWEn)Bit 11. Enables MCA reporting of atomic read-modify-write (RMW) HyperTransport technology commands received from an noncoherent HyperTransport link which do not target DRAM. Atomic RMW commands are not supported by AMD NPT Family 0Fh Processors. Atomic RMW commands to memory-mapped I/O are not supported in HyperTransport technology. An atomic RMW command that targets MMIO results in a HyperTransport error response being
134
Chapter 4
32559
Rev. 3.08
July 2007
generated back to the requesting I/O device. The generation of the HyperTransport error response is not affected by this bit. Watchdog Timer Error Reporting Enable (WchDogTmrEn)Bit 12. Enables MCA reporting of watchdog timer errors. The watchdog timer checks for Northbridge system accesses for which a response is expected and where no response is received. See the MCA NB Configuration register for information regarding configuration of the watchdog timer duration. Note that this bit does not affect operation of the watchdog timer in terms of its ability to complete an access that would otherwise cause a system hang. This bit only affects whether such errors are reported through MCA. DRAM Parity Error Reporting Enable (DramParEn)Bit 18. Enables MCA reporting of DRAM address parity errors. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. 4.6.4.2 MCA NB Configuration Register Function 3: Offset 44h
9 8 7 6 5 4 3 2 SyncOnUcEccEn 1 0
MCA NB Configuration Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 SyncOnDramAdrParErrEn DisMstAbtCpuErrRsp DisTgtAbtCpuErrRsp DisPciCfgCpuErrRsp NbMcaToMstCpuEn WdogTmrBaseSel SyncOnAnyErrEn
WdogTmrCntSel
SyncOnWdogEn
GenCrcErrByte1
GenCrcErrByte0
SyncPktPropDis
CpuRdDatErrEn
SyncPktGenDis
ChipKillEccEn
IoMstAbortDis
IoRdDatErrEn
Bits 31 30 29 28 27 26 25 24 23 22 21 20 1918
Mnemonic reserved
Function
R/W R R/W R/W R/W R/W R R/W R/W R/W R/W R/W R/W R
Reset 0 0 0 0 0 0 0 0 0 0 0 0 0
SyncOnDramAdrParErrEn Sync Flood on DRAM Address Parity Error Enable DisMstAbtCpuErrRsp DisTgtAbtCpuErrRsp NbMcaToMstCpuEn reserved DisPciCfgCpuErrRsp IoRdDatErrEn ChipKillEccEn EccEn SyncOnAnyErrEn SyncOnWdogEn reserved PCI configuration CPU Error Response Disable I/O Read Data Error Log Enable Chip-Kill ECC Mode Enable ECC Enable Sync Flood On Any Error Enable Sync Flood on Watchdog Timer Error Enable Master Abort CPU Error Response Disable Target Abort CPU Error Response Disable Northbridge MCA to CPU 0 Enable
Chapter 4
135
CpuEccErrEn
WdogTmrDis
LdtLinkSel
CpuErrDis
reserved
reserved
reserved
IoErrDis
EccEn
32559
Rev. 3.08 July 2007
Bits 17 16 1514 1312 119 8 7 6 5 4 3 2 1 0
Mnemonic GenCrcErrByte1 GenCrcErrByte0 LdtLinkSel WdogTmrBaseSel WdogTmrCntSel WdogTmrDis IoErrDis CpuErrDis IoMstAbortDis SyncPktPropDis SyncPktGenDis SyncOnUcEccEn CpuRdDatErrEn CpuEccErrEn
Function Generate CRC Error on Byte Lane 1 Generate CRC Error on Byte Lane 0 HyperTransport Link Select for CRC Error Generation Watchdog Timer Time Base Select Watchdog Timer Count Select Watchdog Timer Disable I/O Error Response Disable CPU Error Response Disable I/O Master Abort Error Response Disable Sync Packet Propagation Disable Sync Packet Generation Disable Sync Flood on Uncorrectable ECC Error Enable CPU Read Data Error Log Enable CPU ECC Error Log Enable
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Field Descriptions CPU ECC Error Log Enable (CpuEccErrEn)Bit 0. Enables reporting of ECC errors for data destined for the CPU on this node. This bit should be clear if ECC error logging is enabled for the remaining error reporting blocks in the CPU. Logging the same error in more than one block may cause a single error event to be treated as a multiple error event and cause the CPU to enter shutdown. CPU Read Data Error Log Enable (CpuRdDatErrEn)Bit 1. Enables reporting of read data errors (master aborts and target aborts) for data destined for the CPU on this node. This bit should be clear if read data error logging is enabled for the remaining error reporting blocks in the CPU. Logging the same error in more than one block may cause a single error event to be treated as a multiple error event and cause the CPU to enter shutdown. Sync Flood on Uncorrectable ECC Error Enable (SyncOnUcEccEn)Bit 2. Enables flooding of all HyperTransport links with sync packets on detection of an uncorrectable ECC error. Sync Packet Generation Disable (SyncPktGenDis)Bit 3. Disables flooding of all outgoing HyperTransport links with sync packets when a CRC error is detected on an incoming link. By default, sync packet generation for CRC errors is controlled through the HyperTransport technology LDTn Link Control registers (see page 62). Sync Packet Propagation Disable (SyncPktPropDis)Bit 4. Disables flooding of all outgoing HyperTransport links with sync packets when a sync packet is detected on an incoming link. Sync packets are propagated by default. I/O Master Abort Error Response Disable (IoMstAbortDis)Bit 5. Disables setting the NXA bit in HyperTransport response packets to I/O devices on detection of a master abort HyperTransport technology error condition. 136 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
CPU Error Response Disable (CpuErrDis)Bit 6. Disables generation of a read data error response to the CPU core on detection of a target or master abort HyperTransport technology error condition. I/O Error Response Disable (IoErrDis)Bit 7. Disables setting the Error bit in HyperTransport response packets to I/O devices on detection of a target or master abort HyperTransport technology error condition. Watchdog Timer Disable (WdogTmrDis)Bit 8. Disables the watchdog timer. The watchdog timer is enabled by default and checks for Northbridge system accesses for which a response is expected and where no response is received. If such a condition is detected the outstanding access is completed by generating an error response back to the requestor. An MCA error may also be generated if enabled in the MCA NB Control register. Watchdog Timer Count Select (WdogTmrCntSel)Bit 119. Selects the count used by the watchdog timer. The counter selected by WdogTmrCntSel determines the maximum count value in the time base selected by WdogTmrBaseSel. 000b = 4095 (default) 001b = 2047 010b = 1023 011b = 511 100b = 255 101b = 127 110b = 63 111b = 31 Watchdog Timer Time Base Select (WdogTmrBaseSel)Bit 1312. Selects the time base used by the watchdog timer. The counter selected by WdogTmrCntSel determines the maximum count value in the time base selected by WdogTmrBaseSel. 00b = 1 ms (default) 01b = 1 s 10b = 5 ns 11b = reserved HyperTransport Link Select for CRC Error Generation (LdtLinkSel)Bit 1514. Selects the HyperTransport link to be used for CRC error injection through GenCrcErrByte1/GenCrcErrByte0. 00b = HyperTransport link 0 01b = HyperTransport link 1 10b = HyperTransport link 2 11b = undefined Generate CRC Error on Byte Lane 0 (GenCrcErrByte0)Bit 16. Causes a CRC error to be injected on Byte Lane 0 of the HyperTransport link specified by LdtLinkSel. The data carried by the link is unaffected. This bit is cleared after the error has been generated.
Chapter 4
137
32559
Rev. 3.08 July 2007
Generate CRC Error on Byte Lane 1 (GenCrcErrByte1)Bit 17. Causes a CRC error to be injected on Byte Lane 1 of the HyperTransport link specified by LdtLinkSel. The data carried by the link is unaffected. This bit is cleared after the error has been generated. Sync Flood on Watchdog Timer Error Enable (SyncOnWdogEn)Bit 20. Enables flooding of all HyperTransport links with sync packets on detection of a watchdog timer error. Sync Flood On Any Error Enable (SyncOnAnyErrEn)Bit 21. Enables flooding of all HyperTransport links with sync packets on detection of any MCA error that is uncorrectable. Note: Individual errors cannot be disabled from generating sync packets by clearing bits in the MCA NB Control Register. Some individual errors may be disabled from generating sync packets by disabling their error sources in the MCA NB Configuration Register. ECC Enable (EccEn)Bit 22. Enables ECC check/correct mode. This bit must be set in order for any ECC checking/correcting to be enabled for any processor block. If set, ECC is checked and correctable errors are corrected irrespective of whether machine check ECC reporting is enabled. See ECC and Chip Kill Error Checking and Correction on page 148 for more details. The hardware only allows values to be programmed into this field which are consistent with the ECC capabilities of the device as specified in the Northbridge Capabilities register (page 182). Attempts to write values inconsistent with the capabilities results in this field not being updated. Chip-Kill ECC Mode Enable (ChipKillEccEn)Bit 23. Enables chip-kill ECC mode. Setting this bit causes ECC checking to be based on a 128/16 data/ECC rather than on a 64/8 data/ECC. Chip-kill ECC can only be enabled in 128-bit DRAM data width mode. It allows correction of an error affecting a single symbol rather than the single bit correction available in 64/8 ECC checking. When this bit is set, EccEn must be set. See ECC and Chip Kill Error Checking and Correction on page 148 for more details. The hardware only allows values to be programmed into this field which are consistent with the ECC and Chip Kill ECC capabilities of the device as specified in the Northbridge Capabilities register (page 182). Attempts to write values inconsistent with the capabilities results in this field not being updated. I/O Read Data Error Log Enable (IoRdDatErrEn)Bit 24. Enables reporting of read data errors (master aborts and target aborts) for data destined for I/O devices. If this bit is clear (default state), master and target aborts for transactions from I/O devices are not logged by MCA, although error responses may still be generated to the requesting I/O device. PCI Configuration CPU Error Response Disable (DisPciCfgCpuErrRsp)Bit 25. Disables master abort and target abort reporting through the CPU error-reporting banks for PCI configuration accesses. It is recommended that this bit be set in order to avoid MCA exceptions being generated from master aborts for PCI configuration accesses (which can be common during device enumeration).
138
Chapter 4
32559
Rev. 3.08
July 2007
Northbridge MCA to CPU 0 Enable (NbMcaToMstCpuEn)Bit 27. Enables reporting of all MCA errors to CPU 0 including errors from CPU 1. In addition, when this bit is set, reading MC4_CTL, MC4_ADDRESS and MC4_STATUS through MSR accesses from CPU 1 returns zeros and writes that are not all zeros cause a GP fault. This bit is only valid for dual core processors. Target Abort CPU Error Response Disable ((DisTgtAbtCpuErrRsp)Bit 28. Disables target abort reporting through the CPU error-reporting banks. See Register Differences in Revisions of AMD NPT Processors on page 27 for revision information about this field. Master Abort CPU Error Response Disable (DisMstAbtCpuErrRsp)Bit 29. Disables master abort reporting through the CPU error-reporting banks. See Register Differences in Revisions of AMD NPT Processors on page 27 for revision information about this field. Sync Flood on DRAM Address Parity Error Enable (SyncOnDramAdrParErrEn)Bit 30. Enables flooding of all HyperTransport links with sync packets on detection of a DRAM Address Parity Error. See Register Differences in Revisions of AMD NPT Processors on page 27 for revision information about this field. 4.6.4.3 MCA NB Status Low Register Function 3: Offset 48h
20 19 reserved 16 15 ErrorCode 0
MCA NB Status Low Register

31 Syndrome[15:8] 24 23
ErrorCodeExt
Bits 3124 2320 1916 150
Mnemonic Syndrome[15:8] reserved ErrorCodeExt ErrorCode
Function Syndrome Bits 158 for Chip Kill ECC Mode Extended Error Code Error Code
R/W R/W R R/W R/W
Reset X 0 X X
Field Descriptions Error Code (ErrorCode)Bits 150. Logs an error code when an error is detected. See Table 23 on page 141 for the error codes. Table 16 on page 140 describes the possible error code formats. Extended Error Code (ErrorCodeExt)Bits 1916. Logs an extended error code when an error is detected. See Table 23 on page 141 for the extended error codes. Syndrome Bits 158 for Chip Kill ECC Mode (Syndrome[158])Bits 3124. Logs the upper eight syndrome bits when an ECC error is detected. Only valid for ECC errors in Chip-Kill ECC mode.
Chapter 4
139
32559
Rev. 3.08 July 2007
Table 16.
Error Value
Error Code Field Formats

Error Type TLB errors Description Errors in the GART TLB cache. TT = Transaction Type (Table 17 on page 140) LL = Cache Level (Table 18 on page 140)
0000 0000 0001 TTLL
0000 0001 RRRR TTLL
Memory errors
Errors in the cache hierarchy (not applicable for Northbridge errors) RRRR = Memory Transaction Type (Table 19 on page 140) TT = Transaction Type (Table 17 on page 140) LL = Cache Level (Table 18 on page 140)
0000 1PPT RRRR IILL
Bus errors
General bus errors including errors in the HyperTransport link or DRAM. PP = Participation Processor (Table 20 on page 141) T = Time-out (Table 21 on page 141) RRRR = Memory Transaction Type (Table 19 on page 140) II = Memory or I/O (Table 22 on page 141) LL = Cache Level (Table 18 on page 140)
Table 17.
00b 01b 10b 11b
Transaction Type Bits (TT)

Instruction Data Generic reserved
Table 18.
00b 01b 10b 11b
Cache Level Bits (LL)

Level 0 (L0) Level 1 (L1) Level 2 (L2) Generic (LG)
Table 19.
0000b 0001b 0010b 0011b 0100b 0101b
Memory Transaction Type Bits (RRRR)

Generic error (GEN) Generic read (RD) Generic write (WR) Data read (DRD) Data write (DWR) Instruction fetch (IRD)
140
Chapter 4
32559
Rev. 3.08
July 2007
Table 19.
0110b 0111b 1000b
Memory Transaction Type Bits (RRRR) (Continued)

Prefetch Evict Snoop (probe)
Table 20.
00b 01b 10b 11b
Participation Processor Bits (PP)

Local node originated the request (SRC) Local node responded to the request (RES) Local node observed error as 3rd party (OBS) Generic
Table 21.
0 1
Time-Out Bit (T)

Request did not time out Request timed out (TIMOUT)
Table 22.
00b 01b 10b 11b
Memory or I/O Bits (II)

DRAM Memory access (MEM) reserved MMIO or I/O Space access Generic (GEN)
Table 23 on page 141 lists the errors reported by the Northbridge, along with the error codes for each error. The error code fields are defined in the table above. See the NB Control register for more information on errors detected by the Northbridge. Table 23.
Error Type ECC error CRC error Sync error Mst Abort Tgt Abort GART error RMW error
Northbridge Error Codes

Extended Error Code 0000 0001 0010 0011 0100 0101 0110 Error Code Type BUS BUS BUS BUS BUS TLB BUS PP SRC/RSP OBS OBS SRC/OBS SRC/OBS OBS T 0 0 0 0 0 0 RRRR RD/WR GEN GEN RD/WR RD/WR GEN II MEM GEN GEN MEM/IO MEM/IO IO LL LG LG LG LG LG LG LG TT GEN -
Chapter 4
141
32559
Rev. 3.08 July 2007
Table 23.
Error Type Wdog error ChipKill ECC error DEV Error DRAM Parity Error
Northbridge Error Codes

Extended Error Code 0111 1000 1001 1101 Error Code Type BUS BUS BUS BUS PP GEN SRC/RSP SRC/OBS OBS T 1 0 0 0 RRRR GEN RD/WR RD/WR GEN II GEN MEM MEM/IO MEM LL LG LG LG LG TT -
4.6.4.4
MCA NB Status High Register Function 3: Offset 4Ch

15 14 13 12 UnCorrECC CorrECC 10 9 DramChannel 8 ErrScrub 7 reserved 6 4 3 reserved 2 1 ErrCPU1 0 ErrCPU0
MCA NB Status High Register

31 30 29 28 27 26 25 24 23 22 ErrAddrVal ErrMiscVal ErrUnCorr reserved ErrOver ErrValid ErrEn
PCC
ECC_Synd (70)
reserved
LDTLink
Bits 31 30 29 28 27 26 25 2423 2215 14 13 1210 9 8 7 64 32 1 0
Mnemonic ErrValid ErrOver ErrUnCorr ErrEn ErrMiscVal ErrAddrVal PCC reserved ECC_Synd (70) CorrECC UnCorrECC reserved DramChannel ErrScrub reserved LDTLink reserved ErrCPU1 ErrCPU0
Function Error Valid Error Overflow Error Uncorrected Error Enable Miscellaneous Error Register Valid Error Address Valid Processor Context Corrupt Syndrome Bits (70) for ECC Errors Correctable ECC Error Uncorrectable ECC Error DRAM Error Channel Error Found by DRAM Scrubber HyperTransport Link Number Error Associated with CPU 1 Error Associated with CPU 0
R/W R/W R/W R/W R/W R R/W R/W R R/W R/W R/W R R/W R/W R R/W R R/W R/W
Reset X X X X X X X 0 X X X 0 X X 0 X 0 X X
142
Chapter 4
32559
Rev. 3.08
July 2007
Field Descriptions Error Associated with CPU 0 (ErrCPU0)Bit 0. If set to 1, this bit indicates that the error was associated with CPU0 core. Error Associated with CPU 1 (ErrCPU1)Bit 1. If set to 1, this bit indicates that the error was associated with CPU1 core. HyperTransport Link Number (LDTLink[2:0])Bits 64. For errors associated with a HyperTransport link (e.g., CRC errors), this field indicates which link was associated with the error. LDTLink[0] = Error associated with HyperTransport link 0 LDTLink[1] = Error associated with HyperTransport link 1 LDTLink[2] = Error associated with HyperTransport link 2 Error Found by DRAM Scrubber (ErrScrub)Bit 8. If set to 1, this bit indicates that the error was found by the DRAM scrubber. DRAM Error Channel (Dram Channel)Bit 9. This bit indicates the channel that the DRAM parity error occurred on. 0 = Error occurred on DRAM Channel A. 1 = Error occurred on DRAM Channel B. Uncorrectable ECC Error (UnCorrECC)Bit 13. If set to 1, this bit indicates that the error was an uncorrectable ECC error. Correctable ECC Error (CorrECC)Bit 14. If set to 1, this bit indicates that the error was a correctable ECC error. Syndrome Bits 70 for ECC Errors (ECC_Synd[7:0])Bits 2215. Logs the lower eight syndrome bits when an ECC error is detected. Processor Context Corrupt (PCC)Bit 25. If set to 1, this bit indicates that the state of the processor may be corrupted by the error condition. Reliable restarting might not be possible. 0 = Processor not corrupted 1 = Processor may be corrupted Error Address Valid (ErrAddrVal)Bit 26. If set to 1, this bit indicates that the address saved in the address register is the address where the error occurred (i.e. it validates the address in the address register). 0 = Address register not valid 1 = Address register valid Miscellaneous Error Register Valid (ErrMiscVal)Bit 27. If set to 1, this bit indicates whether the Miscellaneous Error register contains valid information for this error. This bit is set when CntEn (MSR413) is set and the current error is an ECC error.
Chapter 4
143
32559
Rev. 3.08 July 2007
Error Enable (ErrEn)Bit 28. If set to 1, this bit indicates that MCA error reporting is enabled in the MCA Control register. 0 = MCA error reporting not enabled 1 = MCA error reporting enabled Error Uncorrected (ErrUnCorr)Bit 29. If set to 1, this bit indicates that the error was not corrected by hardware. 0 = Error corrected 1 = Error not corrected Error Overflow (ErrOver)Bit 30. Set to 1 if the Northbridge detects an error with the valid bit of this register already set. Enabled errors are written over disabled errors, uncorrectable errors are written over correctable errors. Uncorrectable errors are not overwritten. 0 = No error overflow 1 = Error overflow Error Valid (ErrValid)Bit 31. If set to 1, this bit indicates that a valid error has been detected. This bit should be cleared to 0 by software after the register is read. 0 = No valid error detected 1 = Valid error detected Table 24 lists the errors reported by Northbridge along with the status bits logged for each error. The status bits are defined in the table above. See the Northbridge Control register for more information on errors detected by Northbridge. Table 24.
Error Type ECC error CRC error Sync error Mst Abort Tgt Abort GART error RMW error
Northbridge Error Status Bit Settings

ErrAddr Val 1 PCC If multi-bit and src CPU 1 1 If src CPU If src CPU If src CPU 0 ECC_Synd Valid 1 Corr ECC UnCorr Err ECC Scrub 1/0 LDTLink Valid 0 Err CPU 1/0
ErrUnCorr If multi-bit
If single- If multibit bit 0 0 0 0 0 0 0 0 0 0 0 0
1 1 1 1 1 1
0 0 1 1 1 1
0 0 0 0 0 0
0 0 0 0 0 0
1 1 1 1 0 1
0 0 1/0 1/0 1/0 0
144
Chapter 4
32559
Rev. 3.08
July 2007
Table 24.
Error Type Wdog error ChipKill ECC error DEV Error DRAM Parity Error
Northbridge Error Status Bit Settings (Continued)

ErrAddr Val 0 1 PCC 1 If multisymbol and src CPU 0 1 ECC_Synd Valid 0 1 Corr ECC 0 UnCorr Err ECC Scrub 0 0 1/0 LDTLink Valid 0 0 Err CPU X 1/0
ErrUnCorr 1 If multisymbol
If single- If multisymbol symbol
1 1
1 0
0 0
0 0
0 0
0 0
1 0
X X
4.6.4.5
MCA NB Address Low Register Function 3: Offset 50h

3 ErrAddrLo 2 0
MCA NB Address Low Register

31
reserved
Bits 313 20
Mnemonic ErrAddrLo reserved
Function Error Address Bits 313
R/W R/W R
Reset X 0
Field Descriptions Error Address Bits 313 (ErrAddrLo)Bits 313. This field specifies bits 313 of the address associated with a machine check error. 4.6.4.6 MCA NB Address High Register Function 3: Offset 54h
8 reserved 7 ErrAddrHi 0
MCA NB Address High Register

31
Chapter 4
145
32559
Rev. 3.08 July 2007
Bits 318 70
Mnemonic reserved ErrAddrHi
Function Error Address Bits 3932
R/W R R/W
Reset 0 X
Field Descriptions Error Address Bits 3932 (ErrAddrHi)Bits 70. This field specifies bits 3932 of the address associated with a machine check error. 4.6.4.7 Watchdog Timer Errors
For watchdog timer errors the ErrAddrHi and ErrAddrLo fields in the MCA NB Address High/Low registers log the hardware state information of the request for which the response timed out. MCA NB Address Low Register for Watchdog Timer Error
31 30 29 28 27 WaitCode reserved WaitPW 25 24 23 22 DstUnit 20 19 18 17 SrcUnit 15 14 NextAction 12 11 9 8 LdtCmd

3 2 0
DstNode
SrcNode
SrcPtr
OpType
reserved
Bit 31-30 29 28 2725 2423 2220 1918 1715 1412 119 83 20
Mnemonic WaitCode WaitPW reserved DstNode DstUnit SrcNode SrcUnit SrcPtr NextAction OpType LdtCmd reserved
Function Wait Code (Bits 1-0 of WaitCode) Wait For Posted Write Destination Node ID Destination Unit ID Source Node ID Source Unit ID Source Pointer Next Action Operation Type HyperTransport Command
R/W
Reset
Field Descriptions HyperTransport Command (LdtCmd)Bit 83. The initial command (in HyperTransport technology format) for the stalled operation. Operation Type (OpType)Bit 119. Indicates what type of operation is stalled. 000b = Normal operation (i.e., none of the other types listed) 146 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
001b 010b 011b 100b 101b 110b 111b
= Bus lock = Local APIC access = Interrupt request = System management request = Interrupt broadcast = Stop grant = SMI ack
Next Action (NextAction)Bit 1412. Indicates what the stalled operation would have done next had it not become stalled. 000b = Complete 001b = reserved 010b = Send request 011b = Send second request (if any) 100b = Send final response back to requestor 101b = Send initial response (if any) back to requestor 110b = Send final response to the memory controller 111b = Send initial response (if any) to mem controller Source Pointer (SrcPtr)Bit 1715. Source pointer of the stalled operation. 000b = Host bridge on local node 001b = CPU on local node 010b = Mem controller on local node 101b = HyperTransport link 0 110b = HyperTransport link 1 111b = HyperTransport link 2 Source Unit ID (SrcUnit)Bit 1918. Source Unit ID of the stalled operation. Source Node ID (SrcNode)Bit 2220. Source Node ID of the stalled operation. Destination Unit ID (DstUnit)Bit 2423. Destination Unit ID of the stalled operation. Destination Node ID (DstNode)Bit 2725. Destination Node ID of the stalled operation. Wait For Posted Write (WaitPW)Bit 29. When set, indicates that a response for the stalled operation is waiting for one or more prior posted writes to complete. Wait Code (WaitCode)Bit 31-30. WaitCode is a 5 bit field; WaitCode[1:0] is in Function 3, Offset 50h; WaitCode[4:2] is in Function 3, Offset 54h. WaitCode encodings are implementation specific. They indicate the reason the hardware is waiting. All zeros mean that there is not a waiting condition. 00000b = No waiting condition 00100b = XBAR waiting for probe response from CPU0 00101b = XBAR waiting for probe response from CPU1 01000b = XBAR waiting because of GART/DEV miss 01001b = XBAR waiting for GART table walk state machine to free up Chapter 4 Memory System Configuration 147
32559
Rev. 3.08 July 2007
01010b 1x000b 1x001b 1x010b 1x011b 1x100b 1x101b 1x110b
= XBAR or CPU waiting for bus to unlock (locked instruction) = SRI waiting for available XBAR upstream request buffer = SRI waiting for available XBAR upstream posted request buffer = SRI waiting for available XBAR downstream request buffer = SRI waiting for available XBAR downstream posted request buffer = SRI waiting for available XBAR upstream response buffer = SRI waiting for available XBAR downstream response buffer = SRI waiting for available XBAR display refresh buffer Function 3: Offset 54h
8 7 GartTblWkInProg 6 3 2 0
MCA NB Address High Register for Watchdog Timer Error

31
reserved
RspCnt
WaitCode
Bits 318 7 63 2-0
Mnemonic reserved GartTblWkInProg RspCnt WaitCode
Function GART Table Walk In Progress (Bit 39 of ErrAddr) System Response Count (Bits 3835 of ErrAddr) Wait Code (Bits 4-2 of WaitCode)
R/W
Reset
Field Descriptions Wait Code (WaitCode)Bit 2-0. See WaitCode definition in MCA NB Address Low Register for Watchdog Timer Error (Function 3, Offset 50h). System Response Count (RspCnt)Bits 63 (Bits 3835 of ErrAddr). Number of outstanding system responses for which the stalled operation is waiting. GART Table Walk In Progress (GartTblWkInProg)Bit 7 (Bit 39 of ErrAddr). Indicates that a GART table walk was in progress at the time the error was logged.
4.6.5
ECC and Chip Kill Error Checking and Correction
The memory controller implements two ECC modes: normal ECC and Chip Kill ECC. These error checking and correction modes can only be used if all DIMMs are ECC capable. 4.6.5.1 Normal ECC
Normal ECC mode is a 64/8 SEC/DED (Single Error Correction/Double Error Detection) Hamming code. It can detect one bit error and correct it, it can detect two bit errors but it can not correct them, and it may detect more than two bit errors depending on the position of corrupted bits. Normal ECC 148 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
mode is enabled when EccEn, Function 3, Offset 44h is set and ChipKillEccEn, Function 3, Offset 44h is clear. Error address (Function 3, Offsets 50h and 54h) is valid and uniquely identifies the DIMM with a correctable or uncorrectable error for 64-bit and 128-bit data width configurations. In a 128-bit data width configuration, ECC is independently applied to the lower 64 and upper 64 data bits. In this configuration ECC may detect and correct two bit errors if one error happens on data bits 63-0 and one error happens on data bits 127-64. Status registers (Function 3, Offsets 48h and 4Ch) and error address registers contains information about the second detected error. Bit position of a correctable error can be determined by matching Ecc_Synd, Function 3, Offset 4Ch with a value from Table 25. Bits 63-0 correspond to data bits, and bits 73-64 correspond to ECC check bits. Table 25.
Bit (0+n) Bit (8+n) Bit (16+n) Bit (24+n) Bit (32+n) Bit (40+n) Bit (48+n) Bit (56+n) Bit (64+n)
ECC Syndromes
n=0 ce 23 0e e3 4f a2 8f 62 01 n=1 cb 25 0b e5 4a a4 8a 64 02 n=2 d3 26 13 e6 52 a7 92 67 04 n=3 d5 29 15 e9 54 a8 94 68 08 n=4 d6 2a 16 ea 57 ab 97 6b 10 n=5 d9 2c 19 ec 58 ad 98 6d 20 n=6 da 31 1a f1 5b b0 9b 70 40 n=7 dc 34 1c f4 5d b5 9d 75 80
4.6.5.2
Chip Kill
Chip Kill ECC mode is a 128/16 SSC/DSD (Single Symbol Correction/Double Symbol Detection) BCH code. A symbol is a group of 4 bits which are 4 bit aligned (bits 03 make symbol 0, bits 47 make symbol 1, etc.). A single symbol error is any bit error combination within one symbol (1 to 4 bits can be corrupted). Chip Kill ECC mode can detect one symbol error and correct it, it can detect two symbol errors but it can not correct them, and it may detect more than two symbol errors depending on the position of corrupted symbols. Chip Kill ECC mode is enabled when EccEn and ChipKillEccEn, Function 3, Offset 44h are set. Chip Kill mode can only be used in a 128-bit data width configuration. A DIMM with a correctable error is uniquely identified with error address (Function 3, Offsets 50h and 54h) and Chip Kill syndrome (Syndrome, Function 3, Offset 48h and Ecc_Synd, Function 3, Offset 4Ch). Error address identifies two DIMMs and Chip Kill syndrome identifies one of them. Chip Kill syndromes for symbols 00h-0fh map to data bits 0-63; Chip Kill syndromes for symbols 2021h map to ECC check bits for data bits 0-63; Chip Kill syndromes for symbols 10h-1fh map to data Chapter 4 Memory System Configuration 149
32559
Rev. 3.08 July 2007
bits 64-127; Chip Kill syndromes for symbols 22-23h map to ECC check bits for data bits 64-127 (see Table 26). Corrupted bit positions within a symbol are determined from the Chip Kill syndrome column number in Table 26. Bits set in the column number identify corrupted bits within a symbol. For example, if 6913h is a Chip Kill syndrome, symbol 05h has an error, and bits 0 and 1 within that symbol are corrupted, since the syndrome is in column 3h. Symbol 05h maps to bits 23-20, so the corrupted bits are 20 and 21. A DIMM with an uncorrectable error can not be uniquely identified. The error address is valid and maps to two DIMMs. Chip Kill mode can be used with any device width. Chip Kill mode can correct all errors in an x4 device, since the device width is one symbol. Chip Kill mode can correct a subset of errors in an x8 or x16 device. Table 26.
Symbol 1h 00h 01h 02h 03h 04h 05h 06h 07h 08h 09h 0ah 0bh 0ch 0dh 0eh 0fh 10h 11h 12h 13h 14h 15h 16h
Chip Kill ECC Syndromes

2h 3h 4h 5h 6h 7h 8h 9h ah bh ch dh eh fh f2df dd88 35a9 a1ba 499b 66cc 8eed 1afe 4cda 11eb 7f4c
e821 7c32 9413 bb44 5365 c776 2f57 5d31 a612 fb23
9584 c8b5 3396 6ea7 eac8 b7f9
227d d95e 846f d0df 7adf
0001 0002 0003 0004 0005 0006 0007 0008 0009 000a 000b 000c 000d 000e 000f 2021 3032 1013 4044 6065 7076 5057 8088 a0a9 b0ba 909b c0cc e0ed f0fe 5041 a082 f0c3 4951 8ea2 c7f3 15c1 2a42 3f83 9054 c015 30d6 6097 e0a8 b0e9 402a 106b 70fc f676 5394 1ac5 dd36 9467 a1e8 e8b9 2f4a f3aa cef4 db35 e4b6 f177 8706 1f07 661b f27c be21 d732 6913 2144 9f65 4857 3288 8ca9 e5ba 5b9b 13cc aded c4fe 20bd d07e 803f bb2d 7cde 358f
74e1 9872 ec93 d6b4 a255 4ec6 3a27 6bd8 1f39 3d01 1602 2b03 8504 b805 9306 ae07 ca08 f709 9801 ec02 7403 6b04 f305 d131 6212 b323 3884 e9b5 5a96 8ba7 1cc8 cdf9 6051 b0a2 d0f3 a4c1 f842
874b bd6c c98d 251e 51ff 720d 590e 640f 465e 976f 164e f79f
4758 5299 6d1a 78db 89ac 9c6d a3ee b62f dc0a e10b 4f0c 7eda afeb bd08 2509 510a c90b d60c 4e0d 3a0e a20f 244c f57d 642c 85fd
e1d1 7262 93b3 b834 59e5 ca56 2b87 dc18 3dc9 ae7a 4fab 1094 70c5 a036 c067 20e8 40b9 904a f01b 4235 1eb6 ba77 7b58 df99 d935 eab6 fb77 4c58 5d99 6e1a 7fdb 5c83 e6f4
307c 502d 80de e08f 84ac 956d a6ee b72f bcfd 734e 369f 8d1e eeff 497e 943f f74e dc9f 9dee 1e2f 924e 1d9f 3c6c 5f8d f2bd
831a 27db 9dac 396d 65ee c12f
11c1 2242 3383 c8f4 45d1 8a62 cfb3
5e34 1be5 d456 9187 a718 e2c9 2d7a 68ab f92c 4c97 33a8 84e9 ea2a 5d6b 11fc
63e1 b172 d293 14b4 7755 a5c6 c627 28d8 4b39 99aa fa4b b741 d982 6ec3 2254 9515 fbd6 2bd1 3d62 16b3 4f34 83c1 c142 4283 a4f4 8fd1
a6bd c87e 7f3f
dd41 6682 bbc3 3554 e815 53d6 8e97 1aa8 c7e9 7c2a a16b 2ffc 2735 65b6 e677 f858
64e5 7256 5987 8518 aec9 b87a 93ab ca2c e1fd 7b99 391a badb 5cac df6d 71c9 3b7a b4ab 572c d8fd
c562 4ab3 a934 26e5 6c56 e387 fe18
150
Chapter 4
32559
Rev. 3.08
July 2007
Table 26.
Symbol 1h 17h 18h 19h 1ah 1bh 1ch 1dh 1eh 1fh 20h 21h 22h 23h
Chip Kill ECC Syndromes

2h 3h 4h 5h 6h 7h 8h 9h ah bh ch dh eh fh db86 9c17 a3b8 e429 2a5a 6dcb f1dc 4397 3ea8 81e9 eb2a 546b 17fc b64d 783e 3faf a8bd c27e 7d3f 3d3e aeaf 23ff
4791 89e2 ce73 5264 15f5 5781 a9c2 fe43 bf41
92a4 c525 3b66 6ce7 e3f8
b479 4a3a 1dbb 715c 26dd d89e 8f1f
d582 6ac3 2954 9615 fcd6
9391 e1e2 7273 6464 f7f5 cce1 4472 8893 fdb4 a761 f9b2 ff61 6fc1
8586 1617 b8b8 2b29 595a cacb dcdc 4f4d
3155 b9c6 7527 56d8 9a39 12aa de4b ab6c 678d ef1e
5ed3 e214 4575 1ba6 bcc7 7328 d449 8a9a 2dfb 9694 c2c5 3e36 6a67 ebe8 bfb9 f606
913c 365d 688e cfef e73c 185d b28e 4def 37ac 586d 82ee ed2f f40c b50d 760e 370f 69bd e57e 213f 58df
55b2 aad3 7914 8675 2ca6 d3c7 9e28 6149 cb9a 34fb b542 da83 19f4 7635 acb6 c377 2e58 4199 9b1a f4db
5451 a8a2 fcf3
434a 171b 7d7c 292d d5de 818f
be01 d702 6903 2104 9f05 c441 4882 8cc3 f654
4807 3208 8c09 e50a 5b0b 130c ad0d c40e 7a0f 132a d76b adfc
4101 8202 c303 5804 1905 da06 9b07 ac08 ed09 2e0a 6f0b 3215 bed6 7a97 5ba8 9fe9 7621 9b32 ed13 da44 ac65 4176 3757 6f88 19a9 f4ba
829b b5cc c3ed 2efe
4.6.6
Scrub Control Register
This register specifies the scrub rate for sequential ECC scrubbing of DRAM, the L2 cache and the DCACHE. The scrub rate specifies the duration between successive scrub events. A value of 0 disables scrubbing. Each scrub event corresponds to: 64 bytes for the DRAM scrubber. 64 bits for the data cache scrubber. One single L2 cache line tag address for the L2 scrubber. Function 3: Offset 58h
21 20 reserved 16 15 13 12 L2Scrub 8 7 5 4 DramScrub 0
Scrub Control Register

31
DcacheScrub
reserved
reserved
Bits 3121 2016 1513 128
Mnemonic reserved DcacheScrub reserved L2Scrub
Function Data Cache Scrub Rate L2 Cache Scrub Rate
R/W R R/W R R/W
Reset 0 0 0 0
Chapter 4
151
32559
Rev. 3.08 July 2007
Bits 75 40
Mnemonic reserved DramScrub
Function DRAM Scrub Rate
R/W R R/W
Reset 0 0
Field Descriptions DRAM Scrub Rate (DramScrub)Bits 40. Specifies the scrub rate for the DRAM. See Table 27. For example, if 256MByte memory is scrubbed every 12 hours, a 64-byte memory block should be scrubbed every 10ms, and scrubbing rate of 10.49ms should be selected. L2 Cache Scrub Rate (L2Scrub)Bits 128. Specifies the scrub rate for the L2 cache. See Table 27. Data Cache Scrub Rate (DcacheScrub)Bits 2016. Specifies the scrub rate for the data cache. See Table 27. Table 27.
Scrub Rate Code 00000b 00001b 00010b 00011b 00100b 00101b 00110b 00111b 01000b 01001b 01010b 01011b
Scrub Rate Control Values

Scrub Rate Do not scrub 40.0 ns 80.0 ns 160.0 ns 320.0 ns 640.0 ns 1.28 s 2.56 s 5.12 s 10.2 s 20.5 s 41.0 s Scrub Rate Code 01100b 01101b 01110b 01111b 10000b 10001b 10010b 10011b 10100b 10101b 10110b All others Scrub Rate 81.9 s 163.8 s 327.7 s 655.4 s 1.31 ms 2.62 ms 5.24 ms 10.49 ms 20.97 ms 42.00 ms 84.00 ms reserved
4.6.7
DRAM Scrub Address Registers
These registers specify the next address to be scrubbed by the DRAM scrubber. They should be initialized by the BIOS before the scrubber is enabled (the DRAM scrubber is enabled by writing a valid scrub rate to DramScrub field, Function 3, Offset 58h). They are then updated by the scrubber hardware as it scrubs successive 64-byte blocks of DRAM. Once the scrubber reaches the DRAM limit address for the node (see DRAM Address Map on page 76), it wraps to the DRAM base address.
152
Chapter 4
32559
Rev. 3.08
July 2007
The initial scrubbing address specified by these registers must be between the base and limit address for the node, defined through the DRAM address maps (see DRAM Address Map on page 76). The recommended initial scrubbing address is the DRAM base address. In addition to sequential DRAM scrubbing, the DRAM scrubber has a redirect mode for scrubbing DRAM locations accessed during normal operation. This is enabled by setting ScrubReDirEn (Function 3, Offset 5Ch). When a DRAM read is generated by any agent other than the DRAM scrubber, correctable ECC errors are corrected as the data is passed to the requestor, but the data in DRAM is not corrected if redirect scrubbing mode is disabled. In scrubber redirect mode, correctable errors detected during normal DRAM read accesses redirect the scrubber to the location of the error. After the scrubber corrects the location in DRAM, it resumes scrubbing from where it left off. DRAM scrub address registers are not modified by the redirect scrubbing mode. Sequential scrubbing and scrubber redirection can be enabled independently or together. ECC errors detected by the scrubber are logged in the MCA registers (see Machine Check Architecture Registers on page 206). 4.6.7.1 DRAM Scrub Address Low Register Function 3: Offset 5Ch
6 5 1 0 ScrubReDirEn
DRAM Scrub Address Low Register

31
ScrubAddrLo
reserved
Bits 316 51 0
Mnemonic ScrubAddrLo reserved ScrubReDirEn
Function DRAM Scrub Address Bits 316 DRAM Scrubber Redirect Enable
R/W R/W R R/W
Reset X 0 0
Field Descriptions DRAM Scrubber Redirect Enable (ScrubReDirEn)Bit 0. If this bit is set, the scrubber is redirected to correct errors found during normal operation. DRAM Scrub Address Bits 316 (ScrubAddrLo)Bits 316. This field specifies the low address bits 316 of the next 64-byte block to be scrubbed.
Chapter 4
153
32559
Rev. 3.08 July 2007
4.6.7.2
DRAM Scrub Address High Register Function 3: Offset 60h

8 reserved 7 ScrubAddrHi 0
DRAM Scrub Address High Register

31
Bits 318 70
Mnemonic reserved ScrubAddrHi
Function DRAM Scrub Address Bits 3932
R/W R R/W
Reset 0 X
Field Descriptions DRAM Scrub Address Bits 3932 (ScrubAddrHi)Bits 70. This field specifies the high address bits 3932 of the next 64-byte block to be scrubbed.
4.6.8
HTC Register
This register is not accessible if HtcCap bit (Function 3, Offset E8h) is not set. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. HTC Register
31 reserved

6 5 HtcActSts 4 HtcAct 3 reserved 2 1 HtcSbcEn 0 HtcEn
Bits 316 5 4 32 1 0
Mnemonic reserved HtcActSts HtcAct reserved HtcSbcEn HtcEn
Function HTC Active Status HTC Active HTC Special Bus Cycle Enable HTC Enable
R/W R R/W R R R/W R/W
Reset 0 0 0 0 0 0
Field Descriptions HTC Enable (HtcEn)Bit 0. HTC is enabled when this bit is set.
154
Chapter 4
32559
Rev. 3.08
July 2007
HTC Special Bus Cycle Enable (HtcSbcEn)Bit 1. Enables special bus cycles generation when HTC active state is entered or exited. When this bit is set, HTC clock active must be a minimum of 1024 ns. HTC Active (HtcAct)Bit 4. When this bit is set, HTC is active. HTC Active Status (HtcActSts)Bit 5. This bit is set by hardware when HTC is active. It is cleared by writing a 1.
4.6.9
Thermal Control Register
This register is not accessible if HtcCap bit in Function 3, Offset E8h register is not set. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. Thermal Control Register
31 28 27 24 23 21 20 16 15 14 12 11 10 8

7 StcTmpLoSts 6 StcTmpHiSts 5 reserved 4 3 StcApcTmpLoEn 2 StcApcTmpHiEn 1 StcSbcTmpLoEn 0 StcSbcTmpHiEn
reserved
StcHystLmt
reserved
StcTmpLmt
reserved
Bits 31-28 27-24 2321 20-16 158 7 6 54 3 2 1 0
Mnemonic reserved StcHystLmt reserved StcTmpLmt reserved StcTmpLoSts StcTmpHiSts reserved StcApcTmpLoEn StcApcTmpHiEn StcSbcTmpLoEn StcSbcTmpHiEn
Function STC Hysteresis Limit STC Temperature Limit STC Temperature Low Status STC Temperature High Status STC APIC Temperature Low Interrupt Enable STC APIC Temperature High Interrupt Enable STC Special Bus Cycle Temperature Low Enable STC Special Bus Cycle Temperature High Enable
R/W R R/W R R/W R R/W R/W R R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0
Field Descriptions STC Special Bus Cycle Temperature High Enable (StcSbcTmpHiEn)Bit 0. Enables STC special bus cycle generation when temperature is higher than the STC high temperature. STC Special Bus Cycle Temperature Low Enable (StcSbcTmpLoEn)Bit 1. Enables STC special bus cycle generation when temperature is lower than the STC low temperature. Chapter 4 Memory System Configuration 155
32559
Rev. 3.08 July 2007
STC APIC Temperature High Enable (StcApcTmpHiEn)Bit 2. Enables STC APIC interrupt when temperature is higher than the STC high temperature. STC APIC Temperature Low Enable (StcApcTmpLoEn)Bit 3. Enables STC APIC interrupt when temperature is lower than the STC low temperature. STC Temperature High Status (StcTmpHiSts)Bit 6. This bit is set by hardware when die temperature is higher than the STC high temperature. It is cleared by writing a 1. STC Temperature Low Status (StcTmpLoSts)Bit 7. This bit is set by hardware when die temperature is lower than the STC low temperature. It is cleared by writing a 1. STC Temperature Limit (StcTmpLmt)Bits 20-16. Interrupt is generated or special bus cycle is sent when die temperature is greater than HtcTmpLmt. 00000b = 52C 00001b = 54C ... 11111b = 114C STC Hysteresis Limit (StcHystLmt)Bits 27-24. Interrupt is generated or special bus cycle is sent when die temperature is less than HtcTmpLmt minus HtcHystLmt. 0000b = 2C 0001b = 4C ... 1111b = 32C
4.6.10
XBAR Flow Control Buffers
The Northbridge interfaces with the CPU core, DRAM controller, and, through three HyperTransport links, to external chips. The major Northbridge blocks are: System Request Interface (SRI), Memory Controller (MCT), and Cross Bar (XBAR). SRI interfaces with the CPU core and connects coherent HyperTransport links and noncoherent HyperTransport links. MCT maintains cache coherency and interfaces with the DRAM. XBAR is a five port switch which routes the command packets between SRI, MCT, and the three HyperTransport links. Not all HyperTransport links have to be active. The number of buffers available for each link at the XBAR input is shown in Table 28. Table 28.
Link
XBAR Input Buffers

Number of Command Buffers 10 12 Number of Data Buffers 8x3 5 8
HyperTransport Link 0-2 16 x 3 SRI MCT
156
Chapter 4
32559
Rev. 3.08
July 2007
Table 28.
Link Total
XBAR Input Buffers

Number of Command Buffers 70 Number of Data Buffers 37
XBAR command and data buffers are independent. There are 70 command buffers available, but a maximum of 64 can be used at any given time. Number of used data buffers is not restricted. The default allocation of command buffers when all HyperTransport links are present is shown in Table 29. Table 29.
Link
Default XBAR Command Buffer Allocation

Number of Command Buffers 8 11 64
HyperTransport Link 0-2 15 x 3 SRI MCT Total
The HyperTransport I/O Link Specification defines four virtual channels: requests, posted requests, responses, and probes. The default virtual channel command buffer allocation is shown in Table 30. At least one command buffer should be allocated to each used virtual channel to avoid deadlock. Command buffers do not need to be allocated for a virtual channel that is not used by a link. For example, MCT does not initiate any requests, so there is no need to allocate request or posted request command buffers for an MCT link. A link should not send transactions through a virtual channel that does not have at least one command buffer allocated. Default allocation of SRI buffers can be found in SRI-to-XBAR Buffer Count Register on page 161 and Free List Buffer Count Register on page 163. In the SRI-to-Xbar Buffer Count Register (Function 3, Offset 70h), the virtual channels are subdivided into upstream (coherent HyperTransport) and downstream (noncoherent HyperTransport) directions, since SRI must send packets into both coherent HyperTransport and noncoherent HyperTransport links. Some buffers are allocated to a free list pool to use the available resources more efficiently. The buffers in the free list pool are shared by packets in multiple virtual channels. By default, no more than one buffer is allocated to a virtual channel and the rest are allocated to the free list pool in the Free List Buffer Count Register (Function 3, Offset 7Ch). Any buffer increase to SRI can be added to the free list or allocated directly to a virtual channel. The total number of request command buffers allocated through the Free List Buffer Count Register and the SRI-to-Xbar Buffer Count Register cannot exceed 10.
Chapter 4
157
32559
Rev. 3.08 July 2007
Table 30.
Link
Default Virtual Channel Command Buffer Allocation

Request 3 6 2 0 Posted Request 1 5 3 0 Response Probe 6 4 3 8 5 0 0 3 Number of Command Buffers 15 15 8 11
Coherent HyperTransport Links Non-coherent HyperTransport Links SRI MCT
Software can reallocate buffers by modifying buffer count registers (Function 0, Offsets 90h, B0h, D0h, Function 3, Offsets 70h, 78h, 7Ch). The new buffer counts take effect after a warm reset. Since hardware attempts to choose optimal settings, in general, these registers should not be changed. The following information should be used for buffer reallocation: 1. The number of allocated command buffers for each link should not be greater than the number of available command buffers for that link (see Table 28). Sum of allocated command buffers for three HyperTransport links, SRI, and MCT should not be greater than 64. 2. The number of allocated data buffers for each link should not be greater than the number of available data buffers for that link (see Table 28). There is no restriction on the sum of allocated data buffers for three HyperTransport links, SRI, and MCT. 3. Each present HyperTransport link must have at least one command response buffer (Rsp) allocated, even if the link is designated as a request only link in the routing table. 4. If one HyperTransport link is not present, then other links can use all available buffers for each of them (MCT and HyperTransport links can use one additional buffer, and SRI can use two additional buffers). 5. If all links are present, some buffers from a noncoherent HyperTransport link can be reallocated to coherent HyperTransport links. Table 31 shows command buffer allocation after two response buffers from noncoherent HyperTransport link 0 are reallocated to coherent HyperTransport links 1 and 2. Table 31. An Example of a Non Default Virtual Channel Command Buffer Allocation
Link Request Posted Request 5 1 1 3 0 Response 2 7 7 3 8 Probe 0 5 5 0 3 Number of Command Buffers 13 16 16 8 11
HyperTransport Link 0 6 HyperTransport Link 1 3 HyperTransport Link 2 3 SRI MCT 2 0
6. If there are no noncoherent HyperTransport links, one or more response buffers can be reallocated from MCT to HyperTransport links. 158 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
7. In a multiprocessor system, number of coherent HyperTransport response buffers should be increased first. The default buffer allocation is sufficient for a uniprocessor system. 8. If a two processor system is implemented with two coherent HyperTransport links between the processors, AMD recommends configuring one link for requests and the other link for responses using the buffer allocations in Table 32. a. The request link may have zero response data buffers (RspD). b. The response link may have zero request and posted-request data buffers. Table 32.
Link Coherent HyperTransport Request Link
1,2
2P Dual Coherent Link XBAR Command Buffer Allocation

Request (Req) 6 0 6 1 1 1 Posted Request (PReq) 3 0 5 1 1 Response Probe (Rsp) (Prb) 1 15 2 1 2 8 6 0 0 3 11 Number of Command Buffers 16 15 13 8
Coherent HyperTransport Response Link

1
Non-coherent HyperTransport Links SRI Upstream (Dev:3x70) SRI Downstream (Dev:3x70) Free List (Dev:3x7C) MCT (Dev:3x78)
Note: 1. Dev:0x90, Dev:0xB0, or Dev:0xD0. 2. Route broadcasts to the request link.
9. AMD recommends using different buffer allocations for inner and outer nodes when an eight processor system is implemented using either a ladder (Figure 4) or a twisted ladder (Figure 5) configuration. An outer node is a node with a noncoherent HyperTransport link. Unconnected links are never considered noncoherent HyperTransport links when determining if the node is an inner or outer node. Table 33 and Table 34 show the recommended buffer allocations.
Node 0 Node 2 Node 4 Node 6
I/O
I/O Link
I/O Link
I/O
I/O
I/O Link
Node 1
Node 3
Node 5
Node 7
I/O Link
I/O
Figure 4.
8P Ladder Configuration
Chapter 4
159
32559
Rev. 3.08 July 2007
I/O
I/O Link
Node 0
Node 2
Node 4
Node 6
I/O Link
I/O
I/O
I/O Link
Node 1
Node 3
Node 5
Node 7
I/O Link
I/O
Figure 5. Table 33.

Link
8P Twisted Ladder Configuration 8P Outer Node Virtual Channel XBAR Command Buffer Allocation
Request (Req)
1
Posted Request (PReq) 1 5 1 1 -
Response Probe (Rsp) (Prb) 8 4 1 2 8 5 0 3
Number of Command Buffers 15 15 8
Coherent HyperTransport Links SRI Upstream (Dev:3x70) SRI Downstream (Dev:3x70) Free List (Dev:3x7C) MCT (Dev:3x78)
1
1
Non-coherent HyperTransport Links
6 1 1 1 -
11
Note: Dev:0x90, Dev:0xB0, or Dev:0xD0.
Table 34.
Link
8P Inner Node Virtual Channel XBAR Command Buffer Allocation

Request (Req) 3 1 1 1 Posted Request (PReq) 1 1 1 Response Probe (Rsp) (Prb) 6 1 2 8 5 3 11 Number of Command Buffers 15 8
Coherent HyperTransport Links (Dev:0x90, Dev:0xB0, Dev:0xD0) SRI Upstream (Dev:3x70) SRI Downstream (Dev:3x70) Free List (Dev:3x7C) MCT (Dev:3x78)
10. AMD recommends that UMA graphics systems use the flow control buffer allocation in Table 35 to meet the display refresh bandwidth requirements of typical UMA systems. Note that some chipsets may require different values from those specified here. For chipsets that require less display refresh bandwidth a single DispRefReq credit in SRI (Dev:3x74) may be reallocated to the Free Command List (Dev:3x7C) to increase overall system
160
Chapter 4
32559
Rev. 3.08
July 2007
performance. Note that the optimal value needs to be determined by the developer and tested to ensure stability. Contact the chipset vendor for specific recommendations. Table 35.
Dev:0x90 LDT RspD PReqD ReqD Probe Rsp PReq Req 1 5 2 0 1 5 10
Recommended Flow Control Buffer Allocations in UMA systems

Dev:3x70 SRI-XBAR DPReq 1 DReq 1 URspD 1 DispRefReq 3 ReqD 2 URsp 1 UPReq 1 UReq 1 Dev:3x74 XBAR-SRI DPReq 1 DReq 1 DispRefReq 7 Prb 4 UPReq 2 UReq 1 Dev:3x78 MCT-XBAR RspD Prb 8 2 Dev:3x7C Free List FRspD FRsp FReq FreeCmd 2 1 1 8 /7
1 2
Rsp 10
1. Single Core Processor 2. Dual Core Processor
4.6.10.1
SRI-to-XBAR Buffer Count Register Function 3: Offset 70h

10 9 8 7 reserved 6 5 UPReq 4 3 reserved 2 1 0
SRI-to-XBAR Buffer Count Register

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 DispRefReq reserved reserved reserved DPReq URspD
DReq
ReqD
reserved
URsp
UReq
Bits 3130 2928 2726 2524 2322 21:20 1918 1716 1510 98 76 54 32 10
Mnemonic DPReq DReq reserved URspD reserved DispRefReq reserved ReqD reserved URsp reserved UPReq reserved UReq
Function Downstream Posted Request Buffer Count Downstream Request Buffer Count Upstream Response Data Buffer Count Display Refresh Request Buffer Count Request Data Buffer Count Upstream Response Buffer Count Upstream Posted Request Buffer Count Upstream Request Buffer Count
R/W R/W R/W R R/W R R/W R R/W R R/W R R/W R R/W
Reset 01b 01b 0 01b 0 0 0 10b 0 01b 0 01b 0 01b
Chapter 4
161
32559
Rev. 3.08 July 2007
Field Descriptions Upstream Request Buffer Count (UReq)Bits 10. This field defines the number of upstream request buffers available in the XBAR for use by the SRI. Upstream Posted Request Buffer Count (UPReq)Bits 54. This field defines the number of upstream posted request buffers available in the XBAR for use by the SRI. Upstream Response Buffer Count (URsp)Bits 98. This field defines the number of upstream response buffers available in the XBAR for use by the SRI. Request Data Buffer Count (ReqD)Bits 1716. This field defines the number of request data buffers available in the XBAR for use by the SRI. Display Refresh Request Buffer Count (DispRefReq)Bits 21-20. This field defines the number of display refresh request buffers available in the XBAR for use by the SRI. Upstream Response Data Buffer Count (URspD)Bits 2524. This field defines the number of upstream response data buffers available in the XBAR for use by the SRI. Downstream Request Buffer Count (DReq)Bits 2928. This field defines the number of downstream request buffers available in the XBAR for use by the SRI. Downstream Posted Request Buffer Count (DPReq)Bits 3130. This field defines the number of downstream posted request buffers available in the XBAR for use by the SRI. 4.6.10.2 MCT-to-XBAR Buffer Count Register Function 3: Offset 78h
15 14 reserved Prb 12 11 Rsp 8 7 reserved 0
MCT-to-XBAR Buffer Count Register

31 reserved 28 27 RspD 24 23
Bits 3128 2724 2315 1412 118 70
Mnemonic reserved RspD reserved Prb Rsp reserved
Function Response Data Buffer Count Probe Buffer Count Response Buffer Count
R/W R R/W R R/W R/W R
Reset 0 8h 0 011b 8h 0
Field Descriptions Response Buffer Count (Rsp)Bits 118. This field defines the number of response buffers available in the XBAR for use by the MCT.
162
Chapter 4
32559
Rev. 3.08
July 2007
Probe Buffer Count (Prb)Bits 1412. This field defines the number of probe buffers available in the XBAR for use by the MCT. Response Data Buffer Count (RspD)Bits 2724. This field defines the number of response data buffers available in the XBAR for use by the MCT. 4.6.10.3 Free List Buffer Count Register Function 3: Offset 7Ch
14 13 12 11 10 reserved reserved FRspD FRsp 8 7 reserved 6 5 4 3 FreeCmd 0
Free List Buffer Count Register

31
FReq
Bits 3114 1312 11 108 76 54 30
Mnemonic reserved FRspD reserved FRsp reserved FReq FreeCmd 1. Single Core Processor 2. Dual Core Processor
Function SRI to XBAR Free Response Data Buffer Count SRI to XBAR Free Response Buffer Count SRI to XBAR Free Request Buffer Count SRI Free Command Buffer Count
R/W R R/W R R/W R R/W R/W
Reset 0 10b 0 010b 0 01b Bh /Ah

1 2
Field Descriptions SRI Free Command Buffer Count (FreeCmd)Bits 30. This field defines the number of free list request or posted request buffers available in the SRI for use by the XBAR or CPU. SRI to XBAR Free Request Buffer Count (FReq)Bits 54. This field defines the number of free list request buffers available in the XBAR for use by the SRI. SRI to XBAR Free Response Buffer Count (FRsp)Bits 108. This field defines the number of free list response buffers available in the XBAR for use by the SRI. SRI to XBAR Free Response Data Buffer Count (FRspD)Bits 1312. This field defines the number of free list response data buffers available in the XBAR for use by the SRI.
4.6.11
XBAR-to-SRI Buffer Count Register
The Cross Bar Switch to System Request Interface (XBAR-to-SRI) Buffer Count register specifies the number of command buffers for each virtual channel available in the SRI for use by the XBAR. These buffers store commands for traffic routed to the SRI by the XBAR. The buffer counts take effect on the next warm reset. Chapter 4 Memory System Configuration 163
32559
Rev. 3.08 July 2007
Since the XBAR must route packets in both the coherent HyperTransport technology fabric and down the noncoherent HyperTransport chains, the usual virtual channels are further subdivided into upstream (from noncoherent HyperTransport devices) and downstream (to noncoherent HyperTransport devices) directions. In order to make more efficient use of the available resources, some of the buffers are allocated onto a free list pool in which case these buffers can be used by packets issued in either direction. See Free List Buffer Count Register on page 163. Any increase in the number of fixed allocated buffers in the XBAR-to-SRI Buffer Count register must be compensated by a corresponding reduction of the number of free list buffers specified in the Free List Buffer Count register. When modifying buffer counts care must be taken to allocate a minimum number of buffers to each virtual channel to avoid deadlock. Requests and Posted requests in both directions require at least one buffer and Probes require at least two buffers to avoid deadlock. Since hardware attempts to choose optimal settings, this register should not in general need to be changed. Note: When the Probe Buffer Count is changed, care must be taken to avoid generation of system management events from the time the new value is written into this register until it takes effect (on the next warm reset). XBAR-to-SRI Buffer Count Register
31 30 29 28 27 DPReq 23 22 DispRefReq 20 19 16 15 12 11

7 6 4 3 reserved 2 0
DReq
reserved
reserved
Prb
reserved
UPReq
UReq
Bits 3130 2928 2723 22:20 1916 1512 117 64 3 20
Mnemonic DPReq DReq reserved DispRefReq reserved Prb reserved UPReq reserved UReq
Function Downstream Posted Request Buffer Count Downstream Request Buffer Count Display Refresh Request Buffer Count Probe Buffer Count Upstream Posted Request Buffer Count Upstream Request Buffer Count
R/W R/W R/W R R/W R R/W R R/W R R/W
Reset 01b 01b 0 0 0 8h 0 001b 0 001b
Field Descriptions Upstream Request Buffer Count (UReq)Bits 20. This field defines the number of upstream request buffers available in the SRI for use by the XBAR. Upstream Posted Request Buffer Count (UPReq)Bits 64. This field defines the number of upstream posted request buffers available in the SRI for use by the XBAR. 164 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
Probe Buffer Count (Prb)Bits 1512. This field defines the number of probe buffers available in the SRI for use by the XBAR. Display Refresh Request Buffer Count (DispRefReq)Bits 22-20. This field defines the number of display refresh request buffers available in the SRI for use by the XBAR. Downstream Request Buffer Count (DReq)Bits 2928. This field defines the number of downstream request buffers available in the SRI for use by the XBAR. Downstream Posted Request Buffer Count (DPReq)Bits 3130. This field defines the number of downstream posted request buffers available in the SRI for use by the XBAR.
4.6.12
Display Refresh Flow Control Buffers
The Northbridge has a separate logical path from HyperTransport links to the system memory for display refresh requests. This is necessary to support the bandwidth and latency requirements of display refresh requests in systems where the display-refresh frame buffer is located in the system memory. A HyperTransport packet is defined as a display-refresh request packet when the read request has isochronous bit, PassPW bit, and RspPassPW bit set, and coherent bit and SeqID cleared. All displayrefresh requests are marked as high priority requests and will have higher priority than other requests. In order to accomplish a dedicated logical path to system memory a minimum number of buffers have to be allocated in various queues. DispRefReq (Function 3, Offset 70h) and DispRefReq (Function 3, Offset 74h) need to be set to allocate the buffers for display refresh requests. At least one buffer should be allocated for display refresh requests to avoid deadlock. Usually more than one buffer may be required to support the necessary bandwidth. The buffer count written to the registers takes effect after a warm reset. See SRI-to-XBAR Buffer Count Register on page 161 and XBAR-to-SRI Buffer Count Register on page 163 for more information on setting these registers.
4.6.13
Power Management Control Registers
These registers specify the processor response to STPCLK and HALT power management events. Each STPCLK request received contains a 3-bit System Management Action Field (SMAF) which is used as an index into the Power Management Control Low/High registers to select a Power Management Mode (PMM) value to use for the duration of the STPCLK event. A SMAF value of 000 selects PMM0, a SMAF value of 001 selects PMM1, and so on. HALT is hardwired to use PMM7. 4.6.13.1 Power Management Mode (PMM) Value
The PMM value contains the following fields:
Chapter 4
165
32559
Rev. 3.08 July 2007
Bit 64 3 2 1 0
Name ClkSel AltVidEn FidVidEn NBLowPwrEn CPULowPwrEn
Function Clock Divisor Select Alternate VID Change Enable FID/VID Change Enable Northbridge Low Power Enable CPU Low Power Enable
R/W R/W R/W R/W R/W R/W
PMM Value Field Descriptions Clock Divisor Select (ClkSel)Bits 64. This field specifies the divisor to use when ramping down the CPU clock or the Northbridge clock. 000b = Divide by 8 100b = Divide by 128 001b = Divide by 16 101b = Divide by 256 010b = Divide by 32 110b = Divide by 512 011b = Divide by 64 111b = reserved Alternate VID Change Enable (AltVidEn)Bit 3. Enables a VID change to the value programmed in the AltVid field (Function 3, Offset D8h) after CPU clocks and Northbridge clocks are ramped down. This VID change only occurs if the processor is in the minimum P-state when the STPCLK event occurred. This bit should never be set for FID/VID changes or when UMA graphics are used. 0 = Alternate VID change disabled 1 = Alternate VID change enabled FID/VID Change Enable (FidVidEn)Bit 2. Enables a change in Frequency ID (FID) and/or Voltage ID (VID). The CPU and the Northbridge clocks must also be ramped down (see NBLowPwrEn and CPULowPwrEn). 0 = FID/VID change disabled 1 = FID/VID change enabled Northbridge Low Power Enable (NBLowPwrEn)Bit 1. Causes the Northbridge clock to ramp down according to the clock divisor specified in ClkSel and puts the DRAM in self-refresh mode. The CPU clock must also be ramped down (see CPULowPwrEn) 0 = Northbridge low power disabled 1 = Northbridge low power enabled CPU Low Power Enable (CPULowPwrEn)Bit 0. Causes the CPU clock to ramp down according to the clock divisor specified in ClkSel. 0 = CPU low power disabled 1 = CPU low power enabled
166
Chapter 4
32559
Rev. 3.08
July 2007
4.6.13.2
Power Management Control Low Register Function 3: Offset 80h

8 PMM1 7 reserved 6 PMM0 0 reserved
Power Management Control Low Register

31 30 reserved PMM3 24 23 22 reserved PMM2 16 15 14
Bits 31 3024 23 2216 15 148 7 60
Mnemonic reserved PMM3 reserved PMM2 reserved PMM1 reserved PMM0
Function Power Management Mode 3 Power Management Mode 2 Power Management Mode 1 Power Management Mode 0
Reset 0 0 0 0 0 0 0 0
Field Descriptions Power Management Mode 0 (PMM0)Bits 60. See Power Management Mode (PMM) Value on page 165 for a description of these bits. Power Management Mode 1 (PMM1)Bits 148. See Power Management Mode (PMM) Value on page 165 for a description of these bits. Power Management Mode 2 (PMM2)Bits 2216. See Power Management Mode (PMM) Value on page 165 for a description of these bits. Power Management Mode 3 (PMM3)Bits 3024. See Power Management Mode (PMM) Value on page 165 for a description of these bits. 4.6.13.3 Power Management Control High Register Function 3: Offset 84h
8 PMM5 7 reserved 6 PMM4 0 reserved
Power Management Control High Register

31 30 reserved PMM7 24 23 22 reserved PMM6 16 15 14
Bits 31 3024 23
Mnemonic reserved PMM7 reserved
Function Power Management Mode 7
R/W R R/W R
Reset 0 0 0
Chapter 4
167
32559
Rev. 3.08 July 2007
Bits 2216 15 148 7 60
Mnemonic PMM6 reserved PMM5 reserved PMM4
Function Power Management Mode 6 Power Management Mode 5 Power Management Mode 4
R/W R/W R R/W R R/W
Reset 0 0 0 0 0
Field Descriptions Power Management Mode 4 (PMM4)Bits 60. See Power Management Mode (PMM) Value on page 165 for a description of these bits. Power Management Mode 5 (PMM5)Bits 148. See Power Management Mode (PMM) Value on page 165 for a description of these bits. Power Management Mode 6 (PMM6)Bits 2216. See Power Management Mode (PMM) Value on page 165 for a description of these bits. Power Management Mode 7 (PMM7)Bits 3024. See Power Management Mode (PMM) Value on page 165 for a description of these bits.
4.6.14
GART Aperture Control Register
This register contains the aperture size and enable for the graphics aperture relocation table (GART) mechanism. GART Aperture Control Register
31

7 6 DisGartTblWlkPrb 5 DisGartIo 4 DisGartCpu 3 1 0
reserved
GartSize
Bits 317 6 5 4 31 0
Mnemonic reserved DisGartTblWlkPrb DisGartIo DisGartCpu GartSize GartEn
Function Disable GART Table Walk Probes Disable GART I/O Accesses Disable GART CPU Accesses GART Size GART Enable
Reset 0 0 0 0 0 0
168
Chapter 4
GartEn
32559
Rev. 3.08
July 2007
Field Descriptions GART Enable (GartEn)Bit 0. Enables GART address translation for accesses falling within the GART aperture. GartAperBaseAddr (Function 3, Offset 94h) and other related registers should be initialized before GartEn is set. GART Size (GartSize)Bits 31. Defines the virtual address space to be allocated to the GART. 000b = 32 Mbytes 001b = 64 Mbytes 010b = 128 Mbytes 011b = 256 Mbytes 100b = 512 Mbytes 101b = 1 Gbyte 110b = 2 Gbyte 111b = reserved Disable GART CPU Accesses (DisGartCpu)Bit 4. Disables requests from CPUs from accessing the GART. Disable GART I/O Accesses (DisGartIo)Bit 5. Disables requests from I/O devices from accessing the GART. Disable GART Table Walk Probes (DisGartTblWlkPrb)Bit 6. Disables generation of probes for GART table walks. This bit may be set to improve performance in cases where the GART table entries are in address space which is marked uncacheable in processor MTRRs or page tables.
4.6.15
GART Aperture Base Register
This register contains the base address of the aperture for the graphics aperture relocation table (GART). The GART aperture base along with the GART aperture size define the GART aperture address window. BIOS can place the GART aperture below the 4-gigabyte level in address space in order to support legacy operating systems and legacy AGP cards (that do not support 64-bit address space). GART Aperture Base Register
31 reserved 15 14 GartAperBaseAddr[39:25]

0
Bits 3115 140
Mnemonic reserved
Function
R/W R R/W
Reset 0 X
GartAperBaseAddr[39:25] GART Aperture Base Address Bits 3925
Chapter 4
169
32559
Rev. 3.08 July 2007
Field Descriptions GART Aperture Base Address Bits 3925 (GartAperBaseAddr[39:25])Bits 140. These bits are used to compare to the incoming addresses to determine if they are in the aperture range. They are active based on the aperture size. The remaining address bits are assumed to be 0. [39:25] = 32 Mbytes [39:26] = 64 Mbytes [39:27] = 128 Mbytes [39:28] = 256 Mbytes [39:29] = 512 Mbytes [39:30] = 1 Gbytes [39:31] = 2 Gbytes
4.6.16
GART Table Base Register
This register contains the base address of the translation table for the graphics aperture relocation table (GART). The GART table base points to the base address of a table of 32-bit GART page table entries (PTEs) that contain the physical addresses to use for an incoming address that falls within the GART aperture. GART Table Base Register
31 GartTblBaseAddr[39:12]

4 3 reserved 0
Bits 314 30
Mnemonic GartTblBaseAddr[39:12] reserved
Function GART Table Base Address Bits 3912
R/W R/W R
Reset X 0
Field Descriptions GART Table Base Address Bits 3912 (GartTblBaseAddr[39:12])Bits 314. These bits point to the base of the table of GART PTEs to be used for GART address translation. Table 36 shows the organization of each 32-bit PTE in the GART table. Table 36.
Bits 0 1 32
GART PTE Organization

Description Valid Coherent reserved
170
Chapter 4
32559
Rev. 3.08
July 2007
Table 36.
Bits 114 3112
GART PTE Organization

Description PhysAddr[39:32] PhysAddr[31:12]
4.6.17
GART Cache Control Register
This register controls the GART cache, which is used to cache the most recently translated GART aperture accesses. The GART cache is enabled automatically whenever the GART aperture is enabled. GART Cache Control Register
31 reserved

2 1 GartPteErr R/W R R/WC R/W Reset 0 0 0 8 7 PsiVidEn 6 reserved 5 PsiVid 0 InvGart 0
Bits 312 1 0
Mnemonic reserved GartPteErr InvGart
Function GART PTE Error Invalidate GART
Field Descriptions Invalidate GART (InvGart)Bit 0. Setting this bit causes the GART cache to be invalidated. This bit is cleared by hardware when the invalidation is complete. GART PTE Error (GartPteErr)Bit 1. This bit is set when an invalid PTE is encountered during a table walk. Cleared by writing a 1.
4.6.18
Power Control Miscellaneous Register
This register specifies how the PSI_L signal is controlled. This signal may be used by the voltage regulator to improve efficiency while in reduced power states. Power Control Miscellaneous Register
31 reserved
Function 3: Offset A0h
Chapter 4
171
32559
Rev. 3.08 July 2007
Bits 318 7 6 50
Mnemonic reserved PsiVidEn reserved PsiVid
Function PsiVID Enable PsiVID
R/W R R/W R R/W
Reset 0 0 0 0
Field Descriptions PsiVID (PsiVid)Bits 50. When enabled by PsiVidEn, this field specifies the threshold value of VID code generated by the processor, which in turn determines the state of PSI_L. When the VID code generated by the processor is less than PsiVid (i.e., the VID code is specifying a higher voltage level than the PsiVid-specified voltage level), then PSI_L is high; when the VID code is greater than or equal to PsiVid, PSI_L is driven low. PsiVID Enable (PsiVidEn)Bit 7. This bit enables checking of PsiVid and the assertion of PSI_L. 0 = PSI_L disabled and drive high. 1 = PSI_L enable and controlled by PsiVid.
4.6.19
On-Line Spare Control Register
This register is used by system software to control the transition to the spare rank and to determine which rank has gone bad. On-Line Spare Control Register
31 28 27 24 23 22 21 20 19 18 EccErrCntDramChan EccErrCntDramCs EccErrCntWrEn 16 15 14 13 12 11 SwapDoneInt
Function 3: Offset B0h

7 6 4 3 2 1 0
BadDramCs
SwapDone
EccErrint
reserved
reserved
reserved
reserved
EccErrCnt
reserved
Bits 3128 2724 23 2221 20 19 1816 1514 1312
Mnemonic reserved EccErrCnt EccErCntWrEn reserved EccErrCntDramChan reserved EccErrCntDramCs EccErrInt SwapDoneInt
Function ECC Error Count ECC Error Count Write Enable ECC Error Counter DRAM Channel ECC Error Counter DRAM Chip Select ECC Error Interrupt Type Swap Done Interrupt Type
R/W R R/W R/W R R/W R R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0
172
Chapter 4
SwapEn
32559
Rev. 3.08
July 2007
Bits 117 64 32 1 0
Mnemonic reserved BadDramCs reserved SwapDone SwapEn
Function Bad DRAM Chip Select Swap Done Swap Enable
R/W R R/W R R/W R/W
Reset 0 0 0 0 0
Field Descriptions Swap Enable (SwapEn)Bit 0. Setting this bit causes the hardware to copy the contents of the DRAM chip select identified by the BadDramCs field to the spare rank. The DRAM scrubber must be enabled with a scrub address range that encompasses the address of the bad chip select for the swap to occur. The scrub rate is accelerated automatically by hardware until the copy completes at which point the scrub rate returns to normal. The DRAM scrub address registers (Function 3: Offset 5Ch and 60h) should not be modified while the copy is taking place. During the copy, DRAM accesses (including accesses to the bad CS) proceed normally. Once this bit is set it cannot be cleared by software. This bit is not cleared by a warm reset once the swap has completed. Swap Done (SwapDone)Bit 1. This bit is set when the hardware has completed copying the data to the spare rank. This bit can also be set by BIOS to immediately enable the swap to the spare rank after a suspend to RAM. Once this bit is set it cannot be cleared by software. This bit cannot be set by software if DRAM is enabled (Function 2 OffsetA0h bit 9 = 1). This bit is not cleared by a warm reset. Bad Dram Chip Select (BadDramCs)Bits 64. This field is programmed with the DRAM chip select to be replaced. This field cannot be written when SwapDone is set. This field is not cleared by a warm reset. Swap Done Interrupt Type (SwapDoneInt)Bits 1312. This field specifies what type of interrupt generated when a swap is complete. 00b = No Interrupt 01b = Reserved 10b = SMI 11b = Reserved
1
ECC Error Interrupt Type (EccErrInt)Bits 1514. This field specifies if an interrupt is generated when the EccErrCnt field for any chip select and channel transitions to 1111b. 00b = No Interrupt 01b = Reserved 10b = SMI 11b = Reserved 1. This interrupt will always be routed to cpu0 in a dual core processor.
1
Chapter 4
173
32559
Rev. 3.08 July 2007
ECC Error Counter DRAM Chip Select (EccErrCntDramCs)Bits 1816. This field specifies the DRAM chip select for which ECC error count information is returned in EccErrCnt field. ECC Error Counter DRAM Channel (EccErrCntDramChan)Bit 20. This bit specifies the DRAM channel for which ECC error count information is returned in EccErrCnt field. ECC Error Counter Write Enable (EccErrCntWrEn)Bit 23. When set, this bit enables writes to the EccErrCnt field. ECC Error Counter (EccErrCnt)Bits 2724. This field returns the number of ECC errors for the chip select selected by the EccErrCntDramCs and EccErrCntDramChan fields. This field can be written by software to clear the count. This field returns Fh if 15 or more correctable ECC errors have occurred.
4.6.20
Clock Power/Timing Low Register
This register controls the transition times between various voltage and frequency states. The value of this register is maintained through a warm reset and is initialized to 0 on a cold reset. Clock Power/Timing Low Register
31 16 15 12 11 ClkRampHyst
Function 3: Offset D4h

8 7 reserved 6 reserved 4 3 reserved 2 0
LClkPLLLock (194)
ReConDel
GPE
Bits 3116 1512 11-8 7 64 3 20
Mnemonic LClkPLLLock ReConDel ClkRampHyst reserved reserved reserved GPE
Function HyperTransport CLK PLL Lock Counter Bits 194 Link Reconnect Delay Clock Ramp Hysteresis SBZ Good Phase Error
R/W R/W R/W R/W R R/W R R/W
Reset 0 Ah 0 0 0 0 0
Field Descriptions Good Phase Error (GPE)Bits 20. This field defines the BCLK PLL time until good phase error. Counting occurs when at full frequency. 000b = reserved 001b = 200 system clocks (1us) 010b = 400 system clocks (2us) 011b = 600 system clocks (3us) 100b = 800 system clocks (4us) 174 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
101b = 1600 system clocks (8us) 110b = 3000 system clocks (15us) 111b = 20000 system clocks (100us) Clock Ramp Hysteresis (ClkRampHyst)Bits 11-8. A non-zero value in this field enables a hysteresis time which prevents the CPU clock grid from being ramped down after processing a probe. It avoids unnecessary changes of the CPU clock grid when the probe arrival rate is relatively low. 0000b= 0 0001b= 125 ns 0010b= 250 ns 0011b= 375 ns 0100b= 500 ns 0101b= 750 ns 0110b= 1000 ns 0111b= 2000 ns 1000b= 4 s 1001b= 8 s 1010b= 16 s 1011b= 32 s 1100b= Reserved 1101b= Reserved 1110b= Reserved 1111b= Reserved Link Reconnect Delay (ReConDel)Bits 12-15. A non-zero value in this field specifies the approximate delay, in microseconds, from the deassertion of LDTSTOP_L until the link initialization process is allowed to start. This delay is only applied if LdtStopTriEn = 1 and LdtStopTriClkOvr = 0 (Function 0, Offsets 84h, Ah4, C4h). 0000b= 0 0001b= 1 s 0010b= 2 s 0011b= 3 s 0100b= 4 s 0101b= 5 s 0110b= 6 s 0111b= 7 s 1000b= 8 s 1001b= 9 s 1010b= 10 s 1011b= Reserved 1100b= Reserved 1101b= Reserved 1110b= Reserved Chapter 4 Memory System Configuration 175
32559
Rev. 3.08 July 2007
1111b= Reserved HyperTransport CLK PLL Lock Counter Bits 194 (LClkPLLLock)Bits 3116. This bit field indicates how long it takes for the slowest HyperTransport technology clock PLL to ramp to its new frequency and lock. The 16 bits represent the most significant 16 bits of a 20-bit value. The number reflects the number of system bus clocks (5 ns) to wait. It is only necessary to wait a short amount of time in the event the frequency change is one that maintains the VCO frequency. In this case, the PLL will be re-locked quickly.
4.6.21
Clock Power/Timing High Register
This register controls the transition times between various voltage and frequency states. The value of this register is maintained through a warm reset and is initialized to 0 on a cold reset. Clock Power/Timing High Register
31 30 RampVIDOff reserved 28 27 26 25 AltVidTriEn GfxMode 20 19
Function 3: Offset D8h

0
AtlVid
VSTime
Bits 31 3028 27 26 2520 190
Mnemonic reserved RampVIDOff GfxMode AltVidTriEn AltVid VSTime
Function Ramp VID Offset Graphics Mode AltVID HyperTransport Clock Tristate Enable Alternate VID Voltage Regulator Stabilization Time
Reset 0 0 0 0 0 0
Field Descriptions Voltage Regulator Stabilization Time (VSTime)Bits 190. This field indicates when the voltage regulator is stable at the Ramp VID before ramping up the clocks. The count is the number of 5-ns system clock cycles. 00000h = 0 ns 00001h = 5 ns 00002h = 10 ns ... FFF36h = 5,241,870 ns FFF37h = 5,241,875 ns FFF38h - FFFFFh = Reserved 176 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
Alternate VID (AltVid)Bits 2520. This field specifies the alternate VID while in low power states when enabled through PMM fields in the Power Management Control registers. Switching to AltVID can only be initiated by a StpClk message from the I/O Hub. If an attempt is made to write a VID value that corresponds to a voltage greater than the voltage that MaxVID corresponds to in the FIDVID_STATUS register then the MaxVID value is written instead. If an attempt is made to write a VID value that corresponds to a voltage lower than the voltage that MinVID corresponds to in the FIDVID_STATUS register then the MinVID value is written instead. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. AltVID HyperTransport Clock Tristate Enable (AltVidTriEn)Bit 26. This bit enables the HyperTransport clock to be tristated when AltVID is active. Graphics Mode (GfxMode)Bit 27. This bit is set by BIOS to indicate the graphics mode used by the system. This bit must be set to 0 if LDTSTOP_L with SMAF 001b is used to change link width/freq at boot.See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. 0b = Discrete graphics mode 1b = UMA graphics mode Ramp VID Offset (RampVIDOff)Bits 3028. Defines the amount of extra voltage required while the PLL is ramping for low power states. This over-voltage is applied only until the PLL has phase locked to within specifications. 000b = 0 mV 001b = 25 mV 010b = 50 mV 011b = 75 mV 100b = 100 mV 101b = 125 mV 110b = 150 mV 111b = 175 mV
4.6.22
HyperTransport FIFO Read Pointer Optimization Register
This register allows the separation of read/write pointers in the HyperTransport technology receive/transmit FIFOs to be changed from their default settings. The pointer separation written to this register takes effect after a warm reset. The value of this register is maintained through a warm reset and is initialized to 0 on a cold reset.
Chapter 4
177
32559
Rev. 3.08 July 2007
HyperTransport FIFO Read Pointer Optimization Register

31 22 21 20 19 18 XmtRdPtrLdt2 RcvRdPtrLdt2 reserved 16 15 14 13 12 11 10 XmtRdPtrLdt1 RcvRdPtrLdt1 reserved reserved
Function 3: Offset DCh

8 7 reserved 6 5 XmtRdPtrLdt0 4 3 reserved 2 RcvRdPtrLdt0 0
reserved
Bits 3122 2120 19 1816 1514 1312 11 108 76 54 3 20
Mnemonic reserved XmtRdPtrLdt2 reserved RcvRdPtrLdt2 reserved XmtRdPtrLdt1 reserved RcvRdPtrLdt1 reserved XmtRdPtrLdt0 reserved RcvRdPtrLdt0
Function Change Read Pointer For HyperTransport Link 2 Transmitter Change Read Pointer For HyperTransport Link 2 Receiver Change Read Pointer For HyperTransport Link 1 Transmitter Change Read Pointer For HyperTransport Link 1 Receiver Change Read Pointer For HyperTransport Link 0 Transmitter Change Read Pointer For HyperTransport Link 0 Receiver
R/W R R/W R R/W R R/W R R/W R R/W R R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0
Field Descriptions Change Read Pointer For HyperTransport Link 0 Receiver (RcvRdPtrLdt0)Bits 20. See RcvRdPtrLdt2 for values. Change Read Pointer For HyperTransport Link 0 Transmitter (XmtRdPtrLdt0)Bits 54. See XmtRdPtrLdt2 for values. Change Read Pointer For HyperTransport Link 1 Receiver (RcvRdPtrLdt1)Bits 108. See RcvRdPtrLdt2 for values. Change Read Pointer For HyperTransport Link 1 Transmitter (XmtRdPtrLdt1)Bits 1312. See XmtRdPtrLdt2 for values. Change Read Pointer For HyperTransport Link 2 Receiver (RcvRdPtrLdt2)Bits 1816. Moves the read pointer for the HyperTransport receive FIFO closer to the write pointer thereby reducing latency through the receiver. 000b = RdPtr assigned by hardware 001b = Move RdPtr closer to WrPtr by 1 HyperTransport clock period 178 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
010b 011b 100b 101b 110b 111b
= Move RdPtr closer to WrPtr by 2 HyperTransport clock periods = Move RdPtr closer to WrPtr by 3 HyperTransport clock periods = Move RdPtr closer to WrPtr by 4 HyperTransport clock periods = Move RdPtr closer to WrPtr by 5 HyperTransport clock periods = Move RdPtr closer to WrPtr by 6 HyperTransport clock periods = Move RdPtr closer to WrPtr by 7 HyperTransport clock periods
The BIOS must set this field to 5 for all noncoherent HyperTransport links and to 5 or more for all coherent HyperTransport links. The BIOS must issue a warm reset for the settings to take affect and the warm reset must occur prior to an LDTSTOP_L initiated link frequency change. AMD recommends setting this field to 6 for all coherent HyperTransport links for optimal performance, however the optimal value needs to be determined by the developer and tested to ensure stability. Change Read Pointer For HyperTransport Link 2 Transmitter (XmtRdPtrLdt2)Bits 2120. Moves the read pointer for the HyperTransport technology transmit FIFO closer to the write pointer thereby reducing latency through the transmitter. 00b = RdPtr assigned by hardware 01b = Move RdPtr closer to WrPtr by 1 HyperTransport clock period 10b = Move RdPtr closer to WrPtr by 2 HyperTransport clock periods 11b = Move RdPtr closer to WrPtr by 3 HyperTransport clock periods The BIOS must set this field to 2 for all coherent and noncoherent HyperTransport links and issue a warm reset prior to an LDTSTOP_L initiated link frequency change.
4.6.23
Thermtrip Status Register
The Thermtrip Status register provides status information regarding the THERMTRIP thermal sensor. Thermtrip Status Register
31 30 29 28 24 23 14 13
Function 3: Offset E4h

8 7 6 ThermSenseSel 5 4 ThermtpSense1 3 ThermtpSense0 2 ThermSenseCoreSel 1 0
SwThermtp
ThermtpEn
reserved
reserved
TjOffset
CurTmp
DiodeOffset
Bits 31 3029 28-24 23-14 138
Mnemonic SwThermtp reserved TjOffset CurTmp DiodeOffset
Function Software Thermtrip Tj Offset Current Temperature Diode Offset
R/W W R R R R
Reset 0 0 0
Chapter 4
179
reserved
Thermtp
32559
Rev. 3.08 July 2007
Bits 7 6 5 4 3 2 1 0
Mnemonic reserved ThermSenseSel ThermtpEn ThermtpSense1 ThermtpSense0 ThermSenseCoreSel Thermtp reserved
Function Thermal Sensor Select Thermtrip Enabled Thermtrip Sense 1 Thermtrip Sense 0 Thermal Sensor Core Select Thermtrip
R/W R R/W R R R R/W R R
Reset 0 0
0 0
Field Descriptions Thermtrip (Thermtp)Bit 1. Set to 1 if a temperature sensor trip occurs and was enabled. Thermal Sensor Core Select (ThermSenseCoreSel)Bit 2. This bit selects the CPU whose temperature is reported in the CurTemp field. This bit only applies to dual core processors. For single core processors CPU0 Thermal Sensor is always selected. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. 0b = CPU1 Thermal Sensor. 1b = CPU0 Thermal Sensor. Thermtrip Sense 0 (ThermtpSense0)Bit 3. Set to 1 if a temperature sensor trip occurs for CPU0 on a single core processor, or for CPU1 on a dual core processor. The value of this bit is maintained through warm reset. Thermtrip Sense 1 (ThermtpSense1)Bit 4. Set to 1 if a temperature sensor trip occurs for CPU0 on a dual core processor. The value of this bit is maintained through warm reset. Thermtrip Enabled (ThermtpEn)Bit 5. Indicates that the thermtrip temperature sensor is enabled. When this bit is set to 1, a THERMTRIP High event causes the hardware to shut down the PLL, assert the THERMTRIP output pin and set the ThermtpHi bit. The ThermtpSense bit is set for a THERMTRIP High event, irrespective of the state of ThermtpEn. Thermal Sensor Select (ThermSenseSel)Bit 6. This bit selects which sensor in the CPU whose temperature is reported in the CurTemp field. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. 0b = CPU[0,1] Thermal Sensor 0. 1b = CPU[0,1] Thermal Sensor 1. Diode Offset (DiodeOffset[5:0])Bits 138. Thermal diode offset is used to correct the measurement made by an external temperature sensor. This diode offset supports temperature sensors using two sourcing currents only. Other sourcing current implementations are not compatible with the diode offset and are not supported by AMD. The allowable offset range is provided in the appropriate processor functional data sheet, and the maximum offset can vary 180 Memory System Configuration Chapter 4
32559
Rev. 3.08
July 2007
for different processors. A correction to the offset may be needed for some temperature sensors. Contact the temperature sensor vendor to determine whether an offset correction is needed. With the diode offset, the thermal diode can be used to ensure the processor is within its functional temperature limits. When the thermal diode measurement plus diode offset equals the maximum control temperature (TCONTROL max), the processor has reached its case temperature specification (TCASE max). The maximum control temperature is provided in the appropriate power and thermal data sheet. The relationship between TCASE max and TCONTROL max is described in more detail in section Section 3.7.3.1 on page 41. This field defines a 6 bit signed value between the range of +10 C to -52 C. The diode offset value (in degrees celcius) should be added to the external temperature sensor reading. 000000b = undefined. 000001b = +10C 000010b = +9C ... 111110b = -51C 111111b = -52C Current Temperature (CurTmp)Bits 23-14. This field returns the current value of the internal thermal sensor. The value returned in this field is selected by ThermSenseCoreSel and ThermSenseSel. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. Revision F encodings bits 23-16 (ignore bits 15-14) 00h = -49C 01h = -48C ... ffh = 206C Revision G encodings bits 23-14 000h = -49.00C 001h = -48.75C 002h = -48.50C 003h = -48.25C 004h = -48.00C ... 0C4h = 0.00C ... 3ffh = 206.75C Tj Offset (TjOffset)Bits 28-24. This field is the offset from CurTmp used to normalize to Tcontrol. Tcontrol = CurTmp - TjOffset*2 - 49. If CurTmp - (TjOffset*2) - 49 < 0 Tcontrol = 0. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. Chapter 4 Memory System Configuration 181
32559
Rev. 3.08 July 2007
Software Thermtrip (SwThermtp)Bit 31. Writing a 1 to this bit position induces a THERMTRIP event. This bit is write-only and returns 0 when read. This is a diagnostic bit, and it should be used for testing purposes only.
4.6.24
Northbridge Capabilities Register
The Northbridge Capabilities register indicates whether or not this Northbridge is capable of certain behavior. Northbridge Capabilities Register
31 14 13 12 11 10 9 CmpCap reserved reserved HtcCap
Function 3: Offset E8h

8 MemCntCap 7 reserved 6 DramFreq 5 4 ChipKillEccCap 3 EccCap 2 BigMPCap 1 MPCap 0 128BitCap
reserved
Bits 3114 1312 11 10 9 8 7 65 4 3 2 1 0
Mnemonic reserved CmpCap reserved HtcCap reserved MemCntCap reserved DramFreq ChipKillEccCap EccCap BigMPCap MPCap 128BitCap
Function Dual-Core Capability HTC Capable Memory Controller Capable Maximum DRAM Frequency Chip-Kill ECC Capable ECC Capable Big MP Capable MP Capable 128-Bit DRAM Capable
R/W R R R R R R R R R R R R R
Reset 0 0 0 0
Field Descriptions 128-Bit DRAM Capable (128BitCap)Bit 0. This bit is set to 1 if the Northbridge is capable of supporting a 128-bit DRAM interface. MP Capable (MPCap)Bit 1. This bit is set to 1 if the Northbridge is capable of supporting multiprocessor systems. Big MP Capable (BigMPCap)Bit 2. This bit is set to 1 if the Northbridge is capable of supporting multiprocessor systems greater than DP. ECC Capable (EccCap)Bit 3. This bit is set to 1 if the Northbridge is capable of supporting ECC.
182
Chapter 4
32559
Rev. 3.08
July 2007
Chip-Kill ECC Capable (ChipKillEccCap)Bit 4. This bit is set to 1 if the Northbridge is capable of supporting chip-kill ECC. Maximum DRAM Frequency (DramFreq)Bits 65. Indicates the maximum DRAM frequency supported. 00b = No limit 01b = 333 MHz 10b = 266 MHz 11b = 200 MHz Memory Controller Capable (MemCntCap)Bit 8. This bit is set to 1 if the Northbridge is capable of supporting an on-chip memory controller. HTC Capable (HtcCap)Bit 10. This bit is set to 1 if the Northbridge is capable of supporting HTC. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. Dual-Core Capability (CmpCap)Bits 1312. Indicates the number of CPU cores present in the processor. 00b = Single core processor 01b = Dual core processor 10b = Reserved 11b = Reserved
4.6.25
CPUID Family Model Register
This register returns the same information that is returned in EAX by CPU function 0000_0001h. CPUID Family Model Register
31 reserved 28 27 ExtFamily 20 19 16 15 reserved 12 11 Family
Function 3: Offset FCh

8 7 Model 4 3 Stepping 0
ExtModel
Bits 3128 2720 1916 1512 118 74 30
Mnemonic reserved ExtFamily ExtModel reserved Family Model Stepping
Function Extended Family Extended Model Family Model Stepping
R/W R R R R R R R
Reset 0 X X 0 X X X
Chapter 4
183
32559
Rev. 3.08 July 2007
Field Descriptions Stepping (Stepping)Bits 30. These bits identify the processor revision. Model (Model)Bits 74. These bits identify the processor model number. Family (Family)Bits 118. These bits identify the processor instruction family. Extended Model (ExtModel)Bits 1916. These bits provide additional model information if Family is F0h. Extended Family (ExtFamily)Bits 2720. These bits provide additional family information if Family is F0h.
184
Chapter 4
32559
Rev. 3.08
July 2007
5
5.1
DRAM Configuration
Programming Interface
This section describes how to program various DRAM controller registers. There are two principal areas of configuration: Configuration state information obtained via the DIMM Serial Presence Detect (SPD) ROMs. Non-SPD memory sizing is not supported. All other configuration state information.
5.1.1
SPD ROM-Based Configuration
The SPD device is an EEPROM on the DIMM encoded by the DIMM manufacturer. The description of the EEPROM is usually provided on a data sheet for the DIMM itself along with data describing the memory devices used. The data describes configuration and speed characteristics of the DIMM and the SDRAM components mounted on the DIMM. The data sheet also contains the DIMM byte values that are encoded in the SPD on the DIMM. BIOS acquires the values encoded in the SPD ROM through the I/O hub, which obtains the information through a secondary device connected to the I/O hub through the SMBus. This secondary device communicates with the DIMM by means of the I C bus.
2
The SPD ROM provides values for several DRAM timing parameters that are required by the DRAM controller. These parameters are: tCL: (CAS latency) tRC: Active-to-Active/Auto Refresh command period tRFC: Auto-Refresh-to-Active/Auto Refresh command period tRCD: Active-to-Read-or-Write delay tRRD: Active-Bank-A to-Active-Bank-B delay tRAS: Active-to-Precharge delay tRP: Precharge time tREF: Refresh interval tRTP: Internal Read to Precharge command delay tWTR: Internal Write to Read command delay tWR: Write recovery time DRAM Configuration 185
Chapter 5
32559
Rev. 3.08 July 2007
5.1.1.1
tCL (CAS Latency)
The number of memory clocks it takes a DRAM to return data after the read CAS_L is asserted depends on the memory clock frequency. The value that BIOS programs into the memory controller is a function of the target clock frequency. The target clock frequency is determined from the supported CAS latencies at given clock frequencies of each DIMM. A suggested algorithm is as follows: 1. Determine all CAS latencies supported by each installed DIMM, defined in SPD byte 18. One bit corresponds to each supported CAS latency. SPD byte 18 specifies CAS latencies with which devices on the DIMM can reliably operate. BIOS should not assume that a device can operate with either slower or faster CAS latency than those specified by the SPD. The DRAM controller is designed to support CAS latencies 2, 3,4, 5 and 6. 2. Determine the maximum clock frequency at each supported CAS latency for each DIMM. The minimum cycle time for the highest, the second highest, and the third highest supported CAS latencies are defined in SPD byte 9, 23, and 25, respectively. The minimum cycle time has 1/10ns granularity. 3. Determine the best CAS latency and clock frequency combination. Find the highest clock frequency supported by the slowest DIMM and determine the CAS latency at that operating frequency. It is necessary to choose the highest CAS latency supported by all the DIMMs at the target frequency. 5.1.1.2 tRCD (RAS-to-CAS Delay)
This parameter is defined in SPD byte 29 and it has 1/4ns granularity. BIOS should read and convert this value into a number of DRAM clocks using the target frequency obtained in tCL (CAS Latency) on page 186. 5.1.1.3 tRAS (Active-to-Precharge Delay)
This parameter is defined in SPD byte 30 and it has 1-ns granularity. BIOS should read and convert this value into a number of DRAM clocks using the target frequency obtained in tCL (CAS Latency) on page 186. 5.1.1.4 tRP (Precharge Command Period)
This parameter is defined in SPD byte 27 and it has 1/4-ns granularity. BIOS should read and convert this value into a number of DRAM clocks using the target frequency obtained in tCL (CAS Latency) on page 186. 5.1.1.5 tRC (Active-to-Active/Auto-Refresh Command Period)
This parameter is defined in SPD bytes 40 and 41 and it has 1/4-ns granularity. BIOS should read and convert this value into a number of DRAM clocks using the target frequency obtained in tCL (CAS Latency) on page 186.
186
DRAM Configuration
Chapter 5
32559
Rev. 3.08
July 2007
5.1.1.6
tRRD (Active-to-Active of a Different Bank)
This parameter is defined in SPD byte 28 and it has 1/4-ns granularity. BIOS should read this value and convert into a number of DRAM clocks using the target frequency obtained in tCL (CAS Latency) on page 186. 5.1.1.7 tRFC (Auto-Refresh-to-Active/Auto-Refresh Command Period)
This parameter is defined in SPD bytes 40 and 42 and it has 1/4-ns granularity. BIOS should read and convert this value into a number of DRAM clocks using the target frequency obtained in tCL (CAS Latency) on page 186. 5.1.1.8 tREF (Refresh Rate)
This parameter is defined in SPD byte 12. 5.1.1.9 tRTP (Internal Read to Precharge Command Delay)
This parameter is defined in SPD byte 38 and it has 1/4-ns granularity. BIOS should read and convert this value into a number of DRAM clocks using the target frequency obtained in tCL (CAS Latency) on page 186. 5.1.1.10 tWR (Write Recovery)
This parameter is defined in SPD byte 36 and it has 1/4-ns granularity. BIOS should read and convert this value into a number of DRAM clocks using the target frequency obtained in tCL (CAS Latency) on page 186. 5.1.1.11 tWTR (Internal Write to Read Command Delay)
This parameter is defined in SPD byte 37 and it has 1/4-ns granularity. BIOS should read and convert this value into a number of DRAM clocks using the target frequency obtained in tCL (CAS Latency) on page 186. 5.1.1.12 Registered or Unbuffered DIMMs
SPD byte 20, bits 0 and 4 indicates whether the DIMM registers the address and commands or not. If either of these bits is set the DIMM registers the address and commands. Refer to the processor data sheet to determine the type of DIMMs supported. Systems with mixed unbuffered and registered DIMMs are not supported. 5.1.1.13 Four Rank SO-DIMMs
SPD byte 5, bits 20 indicate the number of ranks present on the DIMM. If these bits are 011b and SPD byte 20, bits 60 are 0001000b the DIMM is a four rank SO-DIMM.
Chapter 5
DRAM Configuration
187
32559
Rev. 3.08 July 2007
5.1.1.14
Four Rank RDIMMs
SPD byte 5, bits 20 indicate the number of ranks present on the DIMM. If these bits are 011b and SPD byte 20, bits 60 are 000001b or 010000b the DIMM is a four rank RDIMM. 5.1.1.15 DIMM Chip Select Density
The size of a DIMM module chip-select range can be obtained directly from the SPD ROM. Byte 31 indicates the size of the chip-select range or ranges on the DIMM (from 128 Mbyte to 8 Gbyte). See the SPD ROM specification for the encoding. Note: If the DIMM uses more than one chip select, then each chip-select range is the same size. 5.1.1.16 DIMM ECC Enable
SPD byte 11 indicates whether the DIMM supports ECC bits. Bit 1 indicates that ECC is supported for all chip-select banks on the DIMM. 5.1.1.17 DIMM Address/Command Parity Enable
SPD byte 11 indicates whether the DIMM supports Address/Command parity. Bit 2 indicates that Address/Command parity is supported for all chip-select banks on the DIMM. 5.1.1.18 x4 DIMMs
The DRAM device width is largely inconsequential to the controller except when they are x4 devices. The width of the devices on the DIMM is determined by the value of SPD byte 13. If the DIMM implements two chip-select banks, then each chip-select bank uses devices of the same width. See the SPD ROM specification for the encoding. 5.1.1.19 DIMM ODT
SPD byte 22, bit 1 indicates support for 50 Ohm ODT. All DIMMs must support 75 Ohms and 150 Ohms. The following tables specify the recommended ODT settings for both the processor and the DIMMs on a per-channel basis. Table 37. RDIMM and Unbuffered DIMM ODT Settings
Processor ODT 75 Ohms 75 Ohms 150 Ohms 150 Ohms DIMM ODT 75 Ohms 75 Ohms 75 Ohms 50 Ohms 75 Ohms
Number of DIMMS 1 at 200-333 MHz 1 at 400 MHz 2 at 400 MHz 3 or more at 400 MHz
2 or more at 200-333 MHz 150 Ohms
188
DRAM Configuration
Chapter 5
32559
Rev. 3.08
July 2007
Table 37.
RDIMM and Unbuffered DIMM ODT Settings

Processor ODT DIMM ODT
Number of DIMMS
Note: DIMMs on channel A and B must have same DIMM ODT value. The DIMM ODT recommendation for a system with 3 total unbuffered DIMMs at 400 MHz is 50 Ohms.
Table 38.
SO-DIMM ODT Settings

Processor ODT 150 Ohms 150 Ohms DIMM ODT 150 Ohms 150 Ohms
MEMCLK Frequency 200-333 MHz 400 MHz
5.1.2
Non-SPD ROM-Based Configuration
Many other bit fields are also required by the DRAM controller configuration registers, but these values cannot be obtained from the SPD ROM. In many cases, these values are hardcoded into the BIOS, but in others they are functions of other bit values. This section describes how BIOS programs each non-SPD related field that is not hardcoded. 5.1.2.1 TrwtTO (Read-to-Write Turnaround for Data, DQS Contention)
This timing parameter ensures read-to-write data-bus turnaround. The setting for this parameter is a function of the asynchronous round-trip loop delay. This parameter should be programmed to 010b for all DIMM types and loadings. 5.1.2.2 Twrrd (Write to Read DIMM Termination Turn-around)
This timing parameter accounts for termination timing when a write is followed by a read to a different DIMM. The setting for this parameter is a function of the asynchronous round-trip loop delay. This parameter should be programmed to 00b for all DIMM types and loadings. 5.1.2.3 Trdrd (Read to Read Timing)
This timing parameter accounts for turn-around and termination timing when a read is followed by a read to a different chip select. The setting for this parameter is a function of the asynchronous roundtrip loop delay. This parameter should be programmed to 00b for all DIMM types and loadings. 5.1.2.4 Twrwr (Write to Write Timing)
This timing parameter accounts for turn-around timing when a write is followed by a write to a different DIMM. This parameter should be programmed to 01b for all DIMM types and loadings.
Chapter 5
DRAM Configuration
189
32559
Rev. 3.08 July 2007
5.1.2.5
Maximum Asynchronous Latency (MaxAsyncLat)
The maximum asynchronous latency is the maximum round-trip latency in the system from the processor to the DRAM devices and back. MaxAsyncLat is the sum of two components: 1. The worst case delay setting for DqsRcvEnDelay 2. A constant of 2ns associated with the delay internal to the processor. If the sum of these two components is less than 4ns, MaxAsyncLat should be programmed to 4ns. 5.1.2.6 tFAW (Four Bank Activate Window)
No more than 4 banks may be activated in a rolling tFAW window. tFAW can be converted to clocks by dividing tFAW(ns) by tCK(ns) and rounding up to next integer value. As an example of the rolling window, if (tFAW/tCK) rounds up to 10 clocks, and an activate command is issued in clock N, no more than three further activate commands may be issued in clock N+1 through N+9. Table 39.
1K 2K
Four Bank Activate Window Values

333 MHz 13 Clocks 17 Clocks 266 MHz 10 Clocks 14 Clocks 200 MHz 8 Clocks 10 Clocks 14 Clocks 18 Clocks
Page Size 400 MHz
5.1.2.7
Pad Receive FIFO Delay
The RdPadRcvFifoDly (Function 2, Offset 78h) specifies how long after the DQS receiver is enabled before the read data is read out of the pad receive FIFO by the DRAM controller. Table 40. RdPadRcvFifoDly Values
4 4 4 4
DRAM Frequency Initial Value 400 MHz 333 MHz 266 MHz 200 MHz
5.1.2.8
Read Pointer Initialization (RdPtrInit)
The RdPtrInit (Function 2, Offset 78h) specifies the initial value of the DRAM FIFO read pointer. This field should be programmed to 6h for all DRAM frequencies. 5.1.2.9 Address Timing
The DRAM controller provides controls for programming the setup on the address pins, the CS/ODT pins and the CKE pins. Table 41, Table 42, Table 43, and Table 44 document the address timing 190 DRAM Configuration Chapter 5
32559
Rev. 3.08
July 2007
settings on a per channel basis. The DIMMs on each channel are numbered from 1 to n where DIMM1 is the DIMM closest to the processor on that channel and DIMMn is the DIMM farthest from the processor on that channel. DIMMs must be populated from farthest slot to closest slot to the processor on a per channel basis. Populations that are not shown in these tables are not supported. Table 41.
DRAM Speed DDR2-400 DDR2-400 DDR2-533 DDR2-533 DDR2-533 DDR2-533 DDR2-533 DDR2-533 DDR2-533 DDR2-667 DDR2-667 DDR2-667 DDR2-667 DDR2-667 DDR2-667 DDR2-800 DDR2-800
Unbuffered DIMM Address Timings and Drive Strengths for AM2 Package
Address Timing Control Register 002F_2F00h 002F_2F00h 002B_2F00h 002F_2F00h 002F_2F00h 0000_2F00h 0034_2F00h 0038_2F00h 0037_2F00h 0020_2220h 0020_2220h 0030_2220h 002B_2220h 002C_2220h 002A_2220h 0020_2520h 0020_2520h Output Driver Compensation Control Register X011_1222h X011_1322h X011_1222h X011_1222h X011_1322h X011_1322h X011_1322h X011_1322h X011_1322h X011_1222h X011_1322h X011_1322h X011_1322h X011_1322h X011_1322h X011_3222h X011_3322h
DIMM1 Any -
DIMM2 Any Any
Timing Mode 1T 2T 1T 1T 2T 2T 2T 2T 2T 1T 2T 2T 2T 2T 2T 2T 2T
SRx16 DRx8 SRx8 SRx16 SRx8 SRx16 SRx8 DRx8 SRx16 DRx8 SRx8 DRx8 Any SRx16 SRx8 SRx16 SRx8 DRx8 SRx16 DRx8 SRx8 DRx8 Any Any
SRx16 SRx16 SRx8 SRx8 DRx8 DRx8 SRx16 DRx8 SRx8 SRx16 SRx16 SRx8 SRx8 DRx8 DRx8 SRx16 DRx8 SRx8 Any
1. SRx16 = Single Rank x16 DIMM SRx8 = Single Rank x8 DIMM DRx16 = Dual Rank x16 DIMM DRx8 = Dual Rank x8 DIMM 2. See Table 37 on page 188 for ProcOdt settings.
Chapter 5
DRAM Configuration
191
32559
Rev. 3.08 July 2007
Table 42.
DRAM Speed DDR2-400 DDR2-400
SO-DIMM Address Timings and Drive Strengths for S1g1 Package

Address Timing Control Register 002F_2F00h 002F_2F00h Output Driver Compensation Control Register X011_1222h X011_1222h
DIMM SRx8
Timing Mode 1T 1T
SRx16 DRx16 DRx8 SRx16
DDR2-533 DDR2-533 DDR2-667 DDR2-667 DDR2-667 DDR2-800 DDR2-800 DDR2-800
SRx8 DRx16 DRx8 SRx16 SRx8 DRx16 DRx8 SRx16 SRx8 DRx16 DRx8
1T 1T 1T 1T 2T 1T 1T 2T
002F_2F00h 002C_2C00h 0027_2700h 002A_2A00h 0000_2800h 0029_2900h 002A_2A00h 0000_2A00h
X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h
1. SRx16 = Single Rank x16 DIMM SRx8 = Single Rank x8 DIMM DRx16 = Dual Rank x16 DIMM DRx8 = Dual Rank x8 DIMM 2. See Table 38 on page 189 for ProcOdt settings.
Table 43.
Registered DIMM Address Timings and Drive Strengths for F(1207) Package 4-DIMM System
Address Timing Control Register 002F_0000h 002F_2700h Output Driver Compensation Control Register X011_1222h X011_1222h
DRAM Speed DDR2-400 DDR2-400
DIMM1 -
DIMM2 SR or DR QR
Timing Mode 1T 1T
1. SR = Single Rank DIMM DR = Dual Rank DIMM QR = Quad Rank DIMM any = SR, DR, or QR 2. See Table 37 on page 188 for ProcOdt settings. 3. Only SR and DR supported.
192
DRAM Configuration
Chapter 5
32559
Rev. 3.08
July 2007
Table 43.
Address Timing Control Register 002F_0000h 002F_2700h 002F_2700h 002F_2700h 002F_0000h 002F_2F00h 002F_0000h 002F_2F00h 002F_2F00h 002F_2F00h Output Driver Compensation Control Register X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h
DRAM Speed DDR2-400 DDR2-400 DDR2-533 DDR2-533 DDR2-667 DDR2-667 DDR2-667 DDR2-667 DDR2-667 DDR2-667
3 3
DIMM1 SR or DR QR any SR SR DR DR
DIMM2 SR or DR QR any any SR DR SR DR SR DR
Timing Mode 1T 1T 1T 1T 1T 1T 1T 1T 1T 1T
1. SR = Single Rank DIMM DR = Dual Rank DIMM QR = Quad Rank DIMM any = SR, DR, or QR 2. See Table 37 on page 188 for ProcOdt settings. 3. Only SR and DR supported.
Table 44.
Address Timing Control Register 002F_0000h 002F_0000h 002F_0000h 002F_0000h 0032_0000h 002F_0000h 0032_0000h 0032_0000h 0032_0000h Output Driver Compensation Control Register X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h
DRAM Speed DDR2-400 DDR2-400 DDR2-400 DDR2-400 DDR2-533 DDR2-533 DDR2-533 DDR2-533 DDR2-533
DIMM1 any any
DIMM2 any any any any
DIMM3 any any any any any any
DIMM4 any any any any SR DR any any any
Timing Mode 1T 1T 1T 1T 1T 1T 1T 1T 1T
1. any = SR or DR. 2. See Table 37 on page 188 for ProcOdt settings.
Chapter 5
DRAM Configuration
193
32559
Rev. 3.08 July 2007
Table 44.
Address Timing Control Register 0000_0000h 002F_0000h 002F_2F00h 002F_0000h 002F_2F00h 002F_2F00h Output Driver Compensation Control Register X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h X011_1222h
DRAM Speed DDR2-667 DDR2-667 DDR2-667 DDR2-667 DDR2-667 DDR2-667
DIMM1 -
DIMM2 -
DIMM3 SR or DRx8 any DRx4
DIMM4 SR DRx8 DRx4 SR or DRx8 DRx4 any
Timing Mode 1T 1T 1T 1T 1T 1T
1. any = SR or DR. 2. See Table 37 on page 188 for ProcOdt settings.
5.1.3
Mismatched DIMM Support
The Mod64BitMux bit (Function 2: Offset A0h), can be used to configure the 128-bit interface as two 64-bit interfaces for AM2 and S1g1 packages. This mode allows mismatched DIMMs to be installed in the upper and lower DIMM slots without reducing any of the DRAM capacity. When operating in this mode the DRAM interface never operates as a 128-bit interface and the two 64-bit interfaces are not accessed simultaneously.
5.1.4
DRAM Initialization
There are two mechanisms that BIOS can use to initialize DRAM hardware controlled initialization and BIOS controlled initialization. Hardware controlled initialization is initiated by BIOS, but hardware controls the initialization sequence. BIOS controlled initialization allows the BIOS to control the initialization sequence. 5.1.4.1 Hardware Controlled DRAM Initialization
Some DRAM configuration fields must be programmed in a specific order. BIOS must set MemClkFreqVal (Function 2 Offset 94h) and wait a minimum of 20 MEMCLKs before setting InitDram (Function 2 Offset 90h). BIOS must set MemClkFreqVal (Function 2 Offset 94h) and wait a minimum of 10us for unbuffered DIMMs and 100us for registered Dimms before setting ExitSelfRef (Function 2 Offset 90h). BIOS must set InitDram last to start DRAM initialization after a cold reset not associated with suspend to RAM mode. BIOS must set ExitSelfRef last to exit self-refresh mode after a cold reset associated with suspend to RAM mode. BIOS should not assert LDTSTOP_L to change HyperTransport link width and frequency while DramConfigRegLo[0] or DramConfigRegLo[1] are set.
194
DRAM Configuration
Chapter 5
32559
Rev. 3.08
July 2007
5.1.4.2
BIOS Controlled DRAM Initialization
BIOS can directly control the DRAM initialization sequence using the DRAM Initialization register (Function 2 Offset 7Ch). BIOS must set MemClkFreqVal (Function 2 Offset 94h) and wait a minimum of 200us before setting EnDramInit (Function 2 Offset 7Ch). BIOS must set the EnDramInit bit last to start DRAM initialization after a cold reset not associated with suspend to RAM mode. BIOS should then complete the initialization sequence specified in the JEDEC DDR2 specification. For registered DIMMs, BIOS should follow the recommendations for reset usage in the JEDEC RDIMM specification during the initialization sequence. After completing the initialization sequence, BIOS must clear the EnDramInit (Function 2 Offset 7Ch) bit. BIOS should not assert LDTSTOP_L to change HyperTransport link width and frequency while EnDramInit (Function 2 Offset 7Ch) is set.
5.1.5
DRAM Training
DRAM training is used to determine when to enable the DQS receivers on reads and to position the DQS strobe in the center of the data eye for both reads and writes. DRAM training must be performed by BIOS after initializing the DRAM controller to ensure proper DRAM operation. Prior to performing DRAM training, the BIOS must transition the processor frequency to 2 GHz and the processor voltage to MaxVID - RVO. If the maximum frequency of the processor is less than 2 GHz, the BIOS must transition the processor to the maximum operating frequency and processor voltage to MaxVID - RVO before performing DRAM training. After DRAM training is complete, the BIOS should transition server and desktop processors the maximum P-state and mobile parts to the minimum P-state. 5.1.5.1 DQS Receiver Enable Training
DQS receiver enable training determines when to enable the DQS receiver for each chip select pair in use. DQS receiver enable training is done in two phases. The first phase is used to detect when the DIMM starts driving the read preamble. The second phase is used to determine effective width of the read preamble. The first phase of DQS receiver enable training is performed using the following procedure: Program the DQS Write Timing Control Registers (Function 2: Offset 9Ch, indexes 01h-03h and 21h-23h) to 00h for all bytes. Program the Read DQS Timing Control Registers (Function 2: Offset 9Ch, indexes 05h-07h and 25h-27h) to 2Fh for all bytes. Set the DqsRcvEnTrain bit. Select two non-consecutive cacheline aligned test addresses on each chip select present. If BASE is the base address of the target chip select, then assigning the first test address to BASE and the second test address to BASE+2MB, targets two non-consecutive addresses on the same chip select regardless of the logical DIMM width, DRAM device configuration, and DRAM bank interleaving mode. DRAM Configuration 195
Chapter 5
32559
Rev. 3.08 July 2007
Write a cache line where each byte is 55h to the first test address for each rank. Write a cache line where each byte is AAh to the second test address for each rank. For each channel For each rank For each DQS receiver enable setting Write the current DQS receiver enable setting to the DQS Receiver Enable Timing Control Register for the current rank. Program the MaxAsyncLat field with the current DQS receiver enable setting plus 2ns. See Section 5.1.2.5 on page 190 for further information on MaxAsyncLat requirements. Read first test address for the rank and compare the first data beat (either 64 or 128 bit depending on the DRAM interface width) with the value written. Reset the read pointer in the DRAM controller FIFO by writing the current DQS receiver enable setting to the DQS Receiver Enable Timing Control Register. Read second test address for the rank and compare the first data beat with the value written. Mark the DQS receiver enable setting as a pass if both reads passed.
For each chip select pair Program the DQS Receiver Enable Timing Control Register for the chip select pair with the first DQS receiver enable setting that passes for both ranks plus 300ps. Save the first DQS receiver enable setting that passes for both ranks. Program the MaxAsyncLat field with the largest DQS receiver enable setting plus 2ns. See Section 5.1.2.5 on page 190 for further information on MaxAsyncLat requirements. Clear the DqsRcvEnTrain bit.
The second phase of DQS receiver enable training is performed after DQS position training is complete using the following procedure: Select two test addresses on each chip select present. Program the starting DQS Receiver Enable Timing Control Register for the chip select pair with the first DQS receiver enable setting that passes for both ranks in phase one plus 300ps. Write a cache line with the following data pattern to the first test address for each rank: 1234_5678_8765_4321h 2345_6789_9876_5432h 5938_5824_3049_6724h 2449_0795_9993_8733h 4038_5642_3846_5245h 2943_2163_0506_7894h
196
DRAM Configuration
Chapter 5
32559
Rev. 3.08
July 2007
1234_9045_9872_3467h 1238_7634_3458_7623h Write a cache line with the following data pattern to the second test address for each rank: 1234_5678_8765_4321h 2345_6789_9876_5432h 5938_5824_3049_6724h 2449_0795_9993_8733h 4038_5642_3846_5245h 2943_2163_0506_7894h 1234_9045_9872_3467h 1238_7634_3458_7623h For each channel For each rank For each DQS receiver enable setting Write the current DQS receiver enable setting to the DQS Receiver Enable Timing Control Register for the current rank. Program the MaxAsyncLat field with the current DQS receiver enable setting plus 2ns. See Section 5.1.2.5 on page 190 for further information on MaxAsyncLat requirements. Read first test address for the rank and compare the data read with the expected value. Reset the read pointer in the DRAM controller FIFO by writing the current DQS receiver enable setting to the DQS Receiver Enable Timing Control Register. Read second test address for the rank and compare the data read with the expected value. Mark the DQS receiver enable setting as a pass if all three reads passed. For each chip select pair Center the DQS receiver enable setting between the first DQS receiver enable setting that passed the first phase of training for both ranks and the last DQS receiver enable setting that passed the second phase of training for both ranks. Program the MaxAsyncLat field with the largest DQS Receiver Enable setting plus 2ns. See Section 5.1.2.5 on page 190 for further information on MaxAsyncLat requirements. DQS Position Training
5.1.5.2
DQS position training is used to place the DQS strobe in the center of the data eye. Determining the correct DQS position for both reads and writes must be performed using a two dimensional search of the read and write position settings. Select three test addresses on each rank present.
Chapter 5
DRAM Configuration
197
32559
Rev. 3.08 July 2007
For each channel For each byte For each rank For each write DQS position (Write DQS Position Loop) Write the current write DQS position to the DQS Write Timing Control Register byte for the current byte. Write the DRAM training pattern to the first test address for the rank. For each read DQS position Write the current read DQS position to the Read DQS Timing Control Register byte for the current byte. Read the DRAM training pattern from the first test address three times. If the training pattern is read correctly all three times mark read position for the byte as a pass. If the read position contains three or more consecutive passing positions exit write DQS position loop. Write the Read DQS Timing Control Register byte for the current byte with the middle position of the passing region. For each write DQS position Write the current write DQS position to the DQS Write Timing Control Register byte for the current byte. Write 0s to the three test addresses for the rank. Write the DRAM training pattern to the three test addresses for the rank. Read the DRAM training pattern from the three test addresses. If the training pattern is read correctly from each test address mark read position for the byte as a pass. Compare the passing regions for the current byte for each rank to determine a mutual passing region for all ranks. Write the Read DQS Timing Control Register byte for the current byte with the middle position of the mutual passing region for reads. Write the DQS Write Timing Control Register byte for the current byte with the middle position of the mutual passing region for writes.
5.2
DDR2 DRAM Configuration
This section shows in a quick reference format how each DRAM controller configuration register should be programmed. BaseAddrReg[7:0][31:0] = See DRAM CS Base Address Registers on page 90. 198 DRAM Configuration Chapter 5
32559
Rev. 3.08
July 2007
BaseMaskReg[7:0][31:0] = See DRAM CS Mask Registers on page 95. BankAddrReg[31:0] = See DRAM Bank Address Mapping Register on page 98. DramTimingRegLo[31:24] = MemClkDis[7:0] = See DRAM Timing Low Register on page 104. DramTimingRegLo[23:22] = tRRD[1:0] = See tRRD (Active-to-Active of a Different Bank) on page 187. DramTimingRegLo[21:20] = tWR[1:0] = See tWR (Write Recovery) on page 187. DramTimingRegLo[19:16] = tRC[3:0] = See tRC (Active-to-Active/Auto-Refresh Command Period) on page 186. DramTimingRegLo[15:12] = tRAS[3:0] = See tRAS (Active-to-Precharge Delay) on page 186. DramTimingRegLo[11] = tRTP= See tRTP (Internal Read to Precharge Command Delay) on page 187. DramTimingRegLo[9:8] = tRP[1:0] = See tRP (Precharge Command Period) on page 186. DramTimingRegLo[5:4] = tRCD[1:0] = See tRCD (RAS-to-CAS Delay) on page 186. DramTimingRegLo[2:0] = tCL[2:0] = See tCL (CAS Latency) on page 186. DramTimingRegHi[31:29] = tRFC3[3:0] = See tRFC (Auto-Refresh-to-Active/Auto-Refresh Command Period) on page 187. DramTimingRegHi[28:26] = tRFC2[3:0] = See tRFC (Auto-Refresh-to-Active/Auto-Refresh Command Period) on page 187. DramTimingRegHi[25:23] = tRFC1[3:0] = See tRFC (Auto-Refresh-to-Active/Auto-Refresh Command Period) on page 187. DramTimingRegHi[22:20] = tRFC0[3:0] = See tRFC (Auto-Refresh-to-Active/Auto-Refresh Command Period) on page 187. DramTimingRegHi[17:16] = tREF[1:0] = See tREF (Refresh Rate) on page 187. DramTimingRegHi[15:14] = Trdrd[1:0] = See Trdrd (Read to Read Timing) on page 189. DramTimingRegHi[13:12] = Twrwr[1:0] = See Twrwr (Write to Write Timing) on page 189. DramTimingRegHi[11:10] = Twrrd[1:0] = See Twrrd (Write to Read DIMM Termination Turnaround) on page 189. DramTimingRegHi[9:8] = tWTR[1:0] = See tWTR (Internal Write to Read Command Delay) on page 187. DramTimingRegHi[0] = TrwtTO = See TrwtTO (Read-to-Write Turnaround for Data, DQS Contention) on page 189. DramConfigRegLo[19] = DimmEccEn = See DIMM ECC Enable,. Chapter 5 DRAM Configuration 199
32559
Rev. 3.08 July 2007
DramConfigRegLo[16] = UnBuffDimm = See Registered or Unbuffered DIMMs on page 187. DramConfigRegLo[15:12] = x4DIMMs[3:0] = See x4 DIMMs on page 188. DramConfigRegLo[11] = Width128 = 0h
1
DramConfigRegLo[10] = BurstLength32 = 0h DramConfigRegLo[9] = SelRefRateEn = 0h
DramConfigRegLo[8] = ParEn = See DIMM Address/Command Parity Enable on page 188 DramConfigRegLo[7] = DramDrvWeak= 0 DramConfigRegLo[5:4] = DramTerm[1:0] = See DIMM ODT on page 188 DramConfigRegLo[1] = ExitSelfRef = 0h DramConfigRegLo[0] = InitDram = 1h DramConfigRegHi[31:28] = FourActWindow[3:0] = See tFAW (Four Bank Activate Window) on page 190. DramConfigRegHi[27:24] = DqsBypassMax[3:0] = 4h DramConfigRegHi[22] = BankSwizzleMode = 0h DramConfigRegHi[20] = SlowAccessMode = 0h DramConfigRegHi[16] = PowerDownMode = 0/1h DramConfigRegHi[15] = PowerDownEn = 1h DramConfigRegHi[14] = DisDramInterface = 0h
4 3
DramConfigRegHi[6:4] = MaxAsyncLat[2:0] = See Maximum Asynchronous Latency (MaxAsyncLat) on page 190. DramConfigRegHi[3] = MemClkFreqVal = 1h DramConfigRegHi[2:0] = MemClkFreq[2:0] = See tCL (CAS Latency) on page 186. DramMisc[8:6] = ILD_lmt[2:0] = 3h DramMisc[5] = DCC_EN = 1h DramMisc[3:2] = RdWrQByp = 2h ScrubControl[31:0] = DramScrubAddrHi[31:0] = DramScrubAddrLo[31:0] = 0000_0000h Notes: 1. This bit should be cleared for 64-bit interfaces and set for 128-bit interfaces. The value of this bit is a function of the system design and cannot be determined via the SPD ROM.
200
DRAM Configuration
Chapter 5
32559
Rev. 3.08
July 2007
2. This bit should be set for 64-bit interfaces (when DramConfigRegLo[11] is cleared) on a platform with a graphics core (AGP or UMA) that issues 32-byte requests. This bit is ignored for 128-bit interfaces (when DramConfigRegLo[11] is set). 3. This bit should be set for mobile systems and cleared for server and desktop systems. 4. In systems that do not use the DRAM controller, it is preferable to disable DRAM input receivers and output drivers and to leave input receivers unconnected. This bit should be set in such systems.
Chapter 5
DRAM Configuration
201
32559
Rev. 3.08 July 2007
202
DRAM Configuration
Chapter 5
32559
Rev. 3.08
July 2007
Machine Check Architecture
The AMD NPT Family 0Fh Processor machine check mechanism allows the processor to detect and report a variety of hardware (or machine) errors found when reading and writing data, probing, cacheline fills and writebacks. These include parity errors associated with caches and TLBs, ECC errors associated with caches and DRAM, as well as system bus errors associated with reading and writing to the external bus. Software can enable the processor to report machine check errors through the machine check exception (See #MCMachine Check Exception in AMD64 Architecture Programmers Manual, Volume 2: System Programming). Most machine check exceptions do not allow reliable restarting of the interrupted programs. However, error conditions are logged in a set of model-specific registers (MSRs) that can be used by system software to determine the possible source of a hardware problem.
6.1
Determining Machine Check Support
The availability of machine check registers and support of the machine check exception is implementation dependent. System software executes the CPUID instruction to determine whether a processor implements these features. After CPUID is executed, the values of the machine check architecture (MCA) bit and the machine check exception (MCE) bit loaded in the EDX register indicate whether the processor implements the machine check registers and the machine check exception, respectively. See Processor Feature Identification in AMD64 Architecture Programmers Manual, Volume 2: System Programming, and CPUID in AMD64 Architecture Programmers Manual, Volume 3: General Purpose and System Instructions, for further information on the level of machine check support. Once system software determines that the machine check registers are available, it must determine the extent of processor support for the machine check mechanism. This is accomplished by reading the machine check capabilities register (MCG_CAP). See Machine Check Global Capabilities Register in AMD64 Architecture Programmers Manual, Volume 2: System Programming, for more information on the interpretation of the MCG_CAP contents.
6.2
Machine Check Errors
Machine check errors are either recoverable or irrecoverable. Recoverable errors are those that the processor can correct and, thus, do not raise the machine check exception (#MC). However, the error is still logged in the machine check MSRs and it is the responsibility of the system software to periodically poll the machine check MSRs to determine if recoverable errors have occurred. If a recoverable error has been logged in the machine check MSRs, a second recoverable error can overwrite it.
Chapter 6
203
32559
Rev. 3.08 July 2007
Irrecoverable, or fatal, machine check errors cannot be corrected by the processor. If machine check is enabled, they raise the machine check exception (#MC). If an irrecoverable error is logged in the machine check MSRs, a second recoverable or irrecoverable MCA error does not overwrite it but does set a bit that indicates overflow. In the case of both recoverable and irrecoverable MCA errors, the contents of the machine check MSRs are maintained though a warm reset. This is helpful since in some cases it may not be possible to invoke the machine check handler due to the error and this allows the BIOS or other system boot software to recover and report the information associated with the error.
6.2.1
Sources of Machine Check Errors
The processor can detect errors from the following hardware blocks within the processor. Each block forms an error reporting bank for the purpose of reporting machine check errors. Data cache unit (DC)Includes the cache structures that hold data and tags, the data TLBs, and the data cache probing logic. Instruction cache unit (IC)Includes the instruction cache structures that hold instructions and tags, the instruction TLBs, and the instruction cache probing logic. Bus unit (BU)Includes the system bus interface to the Northbridge and the level 2 cache. Load/store unit (LS)Includes logic used to manage loads and stores. Northbridge unit (NB)Includes the Northbridge and DRAM controller.
A scrubber is associated with the data cache in DC, the L2 cache tag array in bus unit, and the DRAM in the Northbridge. A scrubber is a hardware widget that periodically wakes up during idle cache cycles and inspects the next line of the array with which it is associated to look for errors. If it finds a single bit ECC error, the scrubber corrects the error and prevents a regular access from encountering the same error. This is important in the data cache, since all ECC errors encountered during regular operation in the data cache are fatal. In the bus unit and the Northbridge, this scrubber function also saves the additional time it would take to fix single-bit ECC errors when they are encountered during normal operation. Even though 64-bits are read from the data cache during normal operation, a byte load uses one byte, a word load uses two bytes, and a double word load uses 4 bytes. If a data cache ECC error is detected in a byte that is not being used by a load, it is corrected like an error detected by the data cache scrubber. This feature is referred to as piggyback scrubbing. Table 45 on page 205 shows the various sources of machine check errors that can be encountered in the processor. ECC check on the linefill data for IC data is done by the DC unit. However, the ECC error is reported by IC which is the unit responsible for receiving the data. The data cache is protected by ECC; however single bit ECC errors encountered by normal load and store accesses are not corrected. When the data cache scrubber encounters a single-bit ECC error, it corrects it.
204
Chapter 6
32559
Rev. 3.08
July 2007
Table 45.
Unit DC
Sources of Machine Check Errors

Type of Check ECC ECC Recoverable Yessingle bit error Nomultiple bit error Yesif single bit ECC error is detected by the scrubber Noany other case different than single bit ECC error detected by the scrubber No No No No Yessingle bit error ECC Nomultiple bit error Yes No Parity Yes Yes
1 1
Error Source System line fill into the data cache L2 cache line fill into the data cache Cache datadata array
Cache datasnoop tag array Cache datatag array L1 Data TLBphysical and virtual arrays L2 Data TLBphysical and virtual arrays IC System line fill into the instruction cache L2 cache line fill into the instruction cache Instruction cachedata array Instruction cachetag array Instruction cachesnoop tag array L1 Instruction TLBphysical and virtual arrays L2 Instruction TLBphysical and virtual arrays System address out of range BU L2 cache data array L2 cachetag array (during scrubs) L2 cachetag array
Parity
Yes
No, but detection is precise ECC Parity Yessingle bit error Nomultiple bit error No. Single and multiple bit errors in an L2 cache tag can also be detected, but not corrected. No No. Loads are detected precisely and stores are detected imprecisely.
System Address Out-of-Range LS NB System Address Out-of-Range See Machine Check Architecture (MCA) Registers on page 132 for information on Northbridge machine check errors.
Read Data Read Data
Note: Instruction cache lines are invalidated and refetched.
Chapter 6
205
32559
Rev. 3.08 July 2007
6.3
Machine Check Architecture Registers
The processor architecture defines a set of model-specific registers (MSRs) to support the MCA mechanism. They are used to configure the MCA functions and provide a way for the hardware to report errors in a manner compatible with the machine check architecture. These registers include the global MCA registers used to set up machine checks and additional banks of MSRs for recording errors that are detected by the hardware blocks listed in Sources of Machine Check Errors on page 204. They can be read and written using the RDMSR and WRMSR instructions. The following is a complete list of MCA MSRs. Global status and control registers Machine check capabilities MSR (MCG_CAP) Processor status MSR (MCG_STATUS) Exception reporting control MSR (MCG_CTL) Each error reporting bank (associated with a specific hardware block listed in Sources of Machine Check Errors on page 204) contains the following registers: Error reporting control register (MCi_CTL) Error reporting control register mask (MCi_CTL_MASK) Error reporting status register (MCi_STATUS) Error reporting address register (MCi_ADDR) Machine check miscellaneous error information register (MCi_MISC)
The i in each register name corresponds to the number of a supported register bank. Each errorreporting register bank is associated with a specific processor unit (or group of processor units). The number of error-reporting register banks is implementation-specific. Software reads the MCG_CAP register to determine the number of supported register banks. AMD NPT Family 0Fh Processors support five banks. The first error-reporting register (MC0_CTL) always starts with MSR address 400h, followed by MC0_STATUS (401h), MC0_ADDR (402h), and MC0_MISC (403h). Error-reporting-register MSR addresses are assigned sequentially through the remaining supported register banks. Using this information, software can access all error-reporting registers in an implementation-independent manner. The global machine check registers as well as the generic form of each register in the error reporting banks are now described. The specifics of each error reporting bank are described in Error Reporting Banks on page 213. The MCi_STATUS_WREN bit in MSR C001_0015h can be used to help debug Machine Check exception handlers. When the MCi_STATUS_WREN bit is set, privileged software can write nonzero values to the MCG_STATUS, MCi_STATUS, MCi_ADDR, and if implemented, MCi_MISC MSRs without generating exceptions, and then simulate a machine check using the INT 18 instruction. Setting a reserved bit in these MSRs does not generate an exception when this mode is enabled. However, setting a reserved bit may result in undefined behavior. 206 Machine Check Architecture Chapter 6
32559
Rev. 3.08
July 2007
6.3.1
Global Machine Check Model-Specific Registers (MSRs)
The global MSRs supported by the machine check mechanism include the MCG_CAP, the MCG_STATUS and the MCG_CTL registers. 6.3.1.1 MCG_CAPGlobal Machine Check Capabilities Register
This read-only register indicates the capabilities of the processor machine check architecture. Attempting to write to this register results in a #GP exception. See AMD64 Architecture Programmers Manual, Volume 2: System Programming, for more details. MCG_CAP Register
63 reserved
MSR 0179h
32
31
8 MCG_CTL_P
reserved
Count
Bit 639 8 70
Name reserved MCG_CTL_P Count
Function MCG_CTL Register Present Count
R/W R R
Reset 0 1 05h
Field Descriptions Count (Count)Bits 70. Number of error-reporting banks supported by the processor implementation. MCG_CTL Register Present (MCG_CTL_P)Bit 8. Indicates if the MCG_CTL register is present. When the bit is set to 1, the register is supported. When the bit is cleared to 0, the register is unsupported. 6.3.1.2 MCG_STATUSGlobal Machine Check Processor Status Register
This register contains basic information about the processor state after a machine check error is detected. See AMD64 Architecture Programmers Manual, Volume 2: System Programming, for more details.
Chapter 6
207
32559
Rev. 3.08 July 2007
MCG_STATUS Register
63 reserved
MSR 017Ah
32
31 reserved
2 MCIP
1 EIPV 1 ICE
0 RIPV 32 0 DCE
Bit 633 2 1 0
Name reserved MCIP EIPV RIPV
Function Machine Check in Progress Error IP Valid Flag Restart IP Valid Flag
R/W R/W R/W R/W
Reset 0 0 0 0
Field Descriptions Restart IP Valid Flag (RIPV)Bit 0. When set, indicates if program execution can be restarted at EIP pushed on stack. Error IP Valid Flag (EIPV)Bit 1. When set, indicates if the EIP pushed on stack is that of the instruction which caused the detection of the machine check error. Machine Check in Progress (MCIP)Bit 2. When set, indicates that a machine check is in progress. 6.3.1.3 MCG_CTLGlobal Machine Check Exception Reporting Control Register
This register contains a global bit for each error-checking unit in the processor. Each bit enables the reporting, through the MCA interface, of errors detected by that particular unit. See AMD64 Architecture Programmers Manual, Volume 2: System Programming, for more details. MCG_CTL Register
63 reserved
MSR 017Bh
31 reserved
4 NBE
3 LSE
2 BUE
208
Chapter 6
32559
Rev. 3.08
July 2007
Bit 635 4 3 2 1 0
Name reserved NBE LSE BUE ICE DCE
Function NB Register Bank Enable LS Register Bank Enable BU Register Bank Enable IC Register Bank Enable DC Register Bank Enable
Reset 0 0 0 0 0 0
Field Descriptions DC Register Bank Enable (DCE)Bit 0. When set, indicates that the Data Cache register bank is enabled. IC Register Bank Enable (ICE)Bit 1. When set, indicates that the Instruction Cache register bank is enabled. BU Register Bank Enable (BUE)Bit 2. When set, indicates that the Bus Unit register bank is enabled. LS Register Bank Enable (LSE)Bit 3. When set, indicates that the Load/Store register bank is enabled. NB Register Bank Enable (NBE)Bit 4. When set, indicates that the Northbridge register bank is enabled.
6.3.2
Error Reporting Bank Machine Check MSRs
The registers in each error reporting bank include MCi_CTL, MCi_CTL_MASK, MCi_STATUS and MCi_ADDR. 6.3.2.1 MCi_CTLMachine Check Control Registers
The machine check control registers (MCi_CTL) contain an enable bit for each error source within an error-reporting register bank. Setting an enable bit to 1 enables error-reporting for the specific feature controlled by the bit, and clearing the bit to 0 disables error reporting for the feature. These registers should generally be set to either all zeros or all ones. As described in Section 5.4.2.2, each enable bit can be masked by the MCi_CTL_MASK register. MCi_CTL Registers
63 EN63 Error-Reporting Register-Bank Enable Bits
MSRs 0400h, 0404h, 0408h, 040Ch, 0410h

2 EN2 1 EN1 0 EN0
Chapter 6
209
32559
Rev. 3.08 July 2007
Bit 630
Name EN63EN0
Function Enables
R/W R/W
Reset 0
Field Descriptions Enables (EN63EN0)Bits 630. If set, error reporting for the specific feature controlled by the bit is enabled. Not all these bits may be implemented in each bank. 6.3.2.2 MCi_CTL_MASKMachine Check Control Mask Registers
The machine check control register masks (MCi_CTL_MASK) provide another level of control in enabling and disabling specific error reporting features. Each bit set to 1 in the MCi_CTL_MASK register inhibits the setting of the corresponding enable bit in the MCi_CTL register that it is associated with. This register is typically programmed by BIOS and not by the Kernel software. MCi_CTL_MASK Registers
63 MSK63 Error-Reporting Register-Bank Mask Bits
MSR C001_0044h, C001_0045h, C001_0046h, C001_0047h, C001_0048h

2 MSK2 1 MSK1 0 MSK0
Bit 630
Name MSK63MSK0
Function Control Register Masks
R/W R/W
Reset 0
Field Descriptions Control Register Masks (MSK63MSK0)Bits 630. If set, the corresponding bit in the MCi_CTL register is forced to 0 when written. 0 = Masking not enabled 1 = Masking enabled 6.3.2.3 MCi_STATUSMachine Check Status Registers
The machine check status register (MCi_STATUS) describes the error detected by its unit. It is written by the processor and should be cleared to 0 by software; writing any other value to the register causes a general protection (GP#) exception. When a machine check error occurs, the processor loads an error code into bits [15:0] of the appropriate MCi_STATUS register. Errors are classified as either system bus, cache, or TLB errors. All the status registers in each bank use the same general logging format shown below.
210
Chapter 6
32559
Rev. 3.08
July 2007
MCi_STATUS Registers
63 62 61 60 59 58 57 56 ADDRV MISCV OVER PCC VAL UC EN
MSRs 0401h, 0405h, 0409h, 040Dh, 0411h

32 Other Information
31 Model-Specific Error Code
16 15 MCA Error Code
Bit 63 62 61 60 59 58 57 5632 3116 150
Name VAL OVER UC EN MISCV ADDRV PCC
Function Valid Status Register Overflow Uncorrected Error indication Error Condition Enabled Miscellaneous-Error Register Valid Error-Address Register Valid Processor-Context Corrupt Other Information Model-Specific Error Code MCA Error Code
Reset X X X X X X X X X X
Field Descriptions MCA Error CodeBits 150. This field holds an error code when an error is detected. Model-Specific Error CodeBits 3116. This field encodes model-specific information about the error. Other InformationBits 5632. This field holds model-specific error information. Processor-Context Corrupt (PCC)Bit 57. If set to 1, this bit indicates that the state of the processor may be corrupted by the error condition. Reliable restarting might not be possible. 0 = Processor not corrupted 1 = Processor may be corrupted Error-Address Register Valid (ADDRV)Bit 58. If set to 1, this bit indicates that the address saved in the address register is the address where the error occurred. 0 = Address register not valid 1 = Address register valid Miscellaneous-Error Register Valid (MISCV)Bit 59. If set to 1, this bit indicates whether the Miscellaneous Error register contains valid information for this error. Error Condition Enabled (EN)Bit 60. If set to 1, this bit indicates that MCA error reporting is enabled for this error in the MCA Control register. Chapter 6 Machine Check Architecture 211
32559
Rev. 3.08 July 2007
0 = Error checking not enabled 1 = Error checking enabled Uncorrected Error indication (UC)Bit 61. If set to 1, this bit indicates that the error was not corrected by hardware. 0 = Error corrected 1 = Error not corrected Status Register Overflow (OVER)Bit 62. Set to 1 if the unit detects an error but the valid bit of this register already set. Enabled errors are written over disabled errors, uncorrectable errors are written over correctable errors. Uncorrectable errors are not overwritten. 0 = No error overflow 1 = Error overflow Valid (VAL)Bit 63. If set to 1, this bit indicates that a valid error has been detected by the unit. This bit should be cleared to 0 by software after the register is read. 0 = No valid error detected 1 = Valid error detected 6.3.2.4 MCi_ADDRMachine Check Address Registers
Each error-reporting register bank includes a machine check address register (MCi_ADDR) that the processor uses to report the instruction memory address or data memory address responsible for the machine check error. The contents of this register are valid only if the ADDRV bit in the corresponding MCi_STATUS register is set to 1. The address can be up to 48 bits wide. In reality, not all address bits may be valid and the address field of the MCi_ADDR varies in width. The address field can hold either a virtual (linear) or physical address, depending on the type or error. Some error types cause the processor to load valid values into a subset of the address bits. The bit ranges depend on the values in the MCi_STATUS register, i.e., on the type of error detected by the unit, which can be determined by examining the MCA error code contained in the MCi_STATUS register. MCi_ADDR Registers
63 reserved
MSRs 0402h, 0406h, 040Ah, 040Eh, 0412h

48 47 ADDR (4732) 32
31 ADDR (310)
212
Chapter 6
32559
Rev. 3.08
July 2007
Bit 6348 470
Name reserved ADDR
Function Address
R/W R/W
Reset 0 0
Field Descriptions Address (ADDR)Bits 470. Not all these bits may be valid. The valid bits depend on the type of error registered in the corresponding MCi_STATUS register. 6.3.2.5 MCi_MISCMachine Check Miscellaneous Registers
Each error-reporting register bank includes a machine check address register (MCi_MISC) that the processor uses to report additional information about specific errors. MC0_MISC, MC1_MISC, MC2_MISC, and MC3_MISC are read only zero and provide no additional information. MC4_MISC is described in MC4_MISCDRAM Errors Threshold Register on page 229. MCi_MISC Registers
63 reserved
MSRs 0403h, 0407h, 040Ah, 040Fh, 0413h

32
31 reserved
Bit 630
Name reserved
Function
R/W
Reset 0
6.4
Error Reporting Banks
AMD NPT Family 0Fh Processors have five error-reporting banksDC, IC, BU, LS, and NB. Each error reporting bank includes the following registers: Machine check control register (MCi_CTL). Error reporting control register mask (MCi_CTL_MASK). Machine check status register (MCi_STATUS). Machine check address register (MCi_ADDR).
The general format of these registers was described in Error Reporting Bank Machine Check MSRs on page 209. This section describes the specifics of each register as it relates to each error reporting bank. Chapter 6 Machine Check Architecture 213
32559
Rev. 3.08 July 2007
6.4.1
Data Cache (DC)
The data cache unit includes the level 1 data cache that holds data and tags, as well as two levels of TLBs and data cache probing logic. 6.4.1.1 MC0_CTLDC Machine Check Control Register
This register enables the reporting, via the MCA interface, of a variety of errors detected by the Data Cache (DC) processor unit. For an error to be reported, both the global DC enable, MCG_CTL[DCE], and the corresponding local enable shown below must be set. If the local enable is off, the error is logged in MC0_STATUS, but not reported via the machine check exception. If the global DC enable is off, no error is logged or reported. MC0_CTL Register
63 reserved
MSR 0400h
32
31 reserved
6 L2TP
5 L1TP
4 DSTP
3 DMTP
2 DECC
1 ECCM
0 ECCI
Bit 637 6 5 4 3 2 1 0
Name reserved L2TP L1TP DSTP DMTP DECC ECCM ECCI
Function L2 TLB Parity Errors L1 TLB Parity Errors Snoop Tag Array Parity Errors Main Tag Array Parity Errors Data Array ECC Errors Multi-bit ECC Data Errors Single-bit ECC Data Errors
R/W R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0
Field Descriptions Single-bit ECC Data Errors (ECCI)Bit 0. Report single-bit ECC data errors during data cache line fills or TLB reloads from the internal L2 or the system. Multi-bit ECC Data Errors (ECCM)Bit 1. Report multi-bit ECC data errors during data cache line fills or TLB reloads from the internal L2 or the system. Data Array ECC Errors (DECC)Bit 2. Report data cache data array ECC errors. Main Tag Array Parity Errors (DMTP)Bit 3. Report data cache main tag array parity errors. Snoop Tag Array Parity Errors (DSTP)Bit 4. Report data cache snoop tag array parity errors. 214 Machine Check Architecture Chapter 6
32559
Rev. 3.08
July 2007
L1 TLB Parity Errors (L1TP)Bit 5. Report data cache L1 TLB parity errors. L2 TLB Parity Errors (L2TP)Bit 6. Report data cache L2 TLB parity errors. 6.4.1.2 MC0_CTL_MASKDC Machine Check Control Mask Register
The MC0_CTL_MASK register, at address MSR C001_0044h, can be used to mask the enabling of individual error reporting features in DC. Any bit set to 1 in this register inhibits writing 1s to the corresponding bits of the MC0_CTL register. See MCi_CTL_MASKMachine Check Control Mask Registers on page 210 for a description of this register. 6.4.1.3 MC0_STATUSDC Machine Check Status Register
The MC0_STATUS MSR describes the error that was detected by DC. The machine check mechanism writes the status register bits when an error is detected and sets the register valid bit (bit 63) to 1 to indicate that the status information is valid. This register is written even if error reporting for the detected error is not enabled. However, error reporting must be enabled for the error to result in a machine check exception. Software is responsible for clearing this register. MC0_STATUS Register
63 62 61 60 59 58 57 56 55 54 reserved ADDRV MISCV OVER PCC VAL UC EN SYND 47 46 45 44 CECC UECC reserved 41 40 39 SCRUB reserved
MSR 0401h
32
31
20 19 EXT_ERR_CODE
16 15
reserved
ERR_CODE
Bit 63 62 61 60 59 58 57 5655 5447 46 45
Name VAL OVER UC EN MISCV ADDRV PCC reserved SYND CECC UECC
Function Valid. Second error detected Error not corrected Error reporting enabled Additional info in MCi_MISC Error address in MCi_ADDR Processor state corrupted by error ECC Syndrome (70) Correctable ECC error Uncorrectable ECC error
Reset X X X X X X X 0 X X X
Chapter 6
215
32559
Rev. 3.08 July 2007
Bit 4441 40 3920 1916 150
Name reserved SCRUB reserved EXT_ERR_CODE ERR_CODE
Function Error detected on a scrub Extended error code Error subsection
R/W R/W R/W R/W R/W
Reset 0 X 0 X X
Field Descriptions Error Subsection (ERR_CODE)Bits 150. Indicates in which subsection the error was detected and what transaction initiated it. See Table 16 on page 140 for the format of the error code and Tables 17 through 22 on pages 140141 for descriptions of the subfields. Extended Error Code (EXT_ERR_CODE)Bits 1916. Contains an extended error code. DC/IC: 0000b = TLB parity error in physical array 0001b = TLB parity error in virtual array (multi-match error) BU: 0000b = Bus or cache data array error 0010b = Cache tag array error LS: Reserved Table 46.
Error Type System Line Fill L2 Cache Line Fill Data Load/ Store/ Victim/ Snoop Data Scrub Tag Snoop/ Victim Tag Load/ Store L1 TLB
DC Error Codes
Extended Error Code 0000 0000 0000 Error Code Type BUS Memory Memory PP SRC T 0 RRRR DRD DRD DRD/D WR/Evi ct/Snoo p GEN Snoop/ Evict DRD/D WR II MEM/IO LL LG L2 L1 TT Data Data
0000 0000 0000 0000
Memory Memory Memory TLB
L1 L1 L1 L1
Data Data Data Data
216
Chapter 6
32559
Rev. 3.08
July 2007
Table 46.
Error Type L1 TLB Multimatch L2 TLB L2 TLB Multimatch
DC Error Codes
Extended Error Code 0001 0000 0001 Error Code Type TLB TLB TLB PP T RRRR II LL L1 L2 L2 TT Data Data Data
Error Detected on a Scrub (SCRUB)Bit 40. DC: If set, indicates that the error was detected on a scrub. IC/BU/LS: Reserved. Uncorrectable ECC Error (UECC)Bit 45. If set, indicates that the error is an uncorrectable ECC error. Correctable ECC Error (CECC)Bit 46. If set, indicates that the error is a correctable ECC error. ECC Syndrome Bits 70 (SYND)Bits 5447. DC: The lower 8 syndrome bits when an ECC error is detected. IC/BU/LS: Reserved. Processor State Corrupted By Error (PCC)Bit 57. If set, indicates that the processor state may have been corrupted by the error condition. Error address in MCi_ADDR (ADDRV)Bit 58. If set, indicates that the address saved in the corresponding MCi_ADDR register is the address where the error occurred. Additional Info in MCi_MISC (MISCV)Bit 59. If set, indicates that the MCi_MISC register contains additional info. This bit is always set to 0 on an AMD NPT Family 0Fh Processor. Error Reporting Enabled (EN)Bit 60. If set, indicates that error reporting was enabled for this error in the corresponding MCi_CTL register. Error Not Corrected (UC)61. If set, it indicates that the error was not corrected. Second Error Detected (OVER)Bit 62. Set if a second error is detected while the VAL bit of this register is already set. Valid (VAL)Bit 63. Set if the information in this register is valid.
Chapter 6
217
32559
Rev. 3.08 July 2007
Table 47.
Error Type System Line Fill L2 Cache Line Fill Data Load/ Store/ Victim/ Snoop Data Scrub Tag Snoop/ Victim Tag Load/ Store L1 TLB L1 TLB Multimatch L2 TLB L2 TLB Multimatch
DC Error Status Bit Settings

UC If multi-bit If multi-bit 1 ADDRV 1 1 1/0 PCC If multi-bit If multi-bit 1 ECC_Synd Valid Y Y Y CECC UECC SCRUB 0 0 0
If single- If multibit bit If single- If multibit bit If single- If multibit bit
If multi-bit 1
1 1/0
0 1
Y N
If single- If multibit bit 0 0
1 0
1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
N N N N N
0 0 0 0 0
0 0 0 0
0 0 0 0 0
6.4.1.4
MC0_ADDRDC Machine Check Address Register (MSR 0402h)
The contents of this register are valid only if the ADDRV bit in the MC0_STATUS register is set to 1. As shown in MCi_ADDRMachine Check Address Registers on page 212, the address can be up to 48 bits wide. In reality, not all address bits may be valid and the address field of the MC0_ADDR varies in width. The bit ranges depend on the values in the MC0_STATUS register (i.e., the type of error detected) as shown in Table 48. Table 48. Valid MC0_ADDR Bits
Access Type snoop victim Data Tag Array load store physical[393] Valid Address Bits physical[116]
Error Source Snoop Tag Array
218
Chapter 6
32559
Rev. 3.08
July 2007
Table 48.
Data Array
Valid MC0_ADDR Bits

Access Type victim snoop load store scrub physical[113] linear[4712] physical[396] physical[393] Valid Address Bits physical[116]
Error Source
L1 Data TLB L2 Data TLB L2 Cache Data System Data
load, store line-fill
6.4.2
Instruction Cache (IC)
The IC unit includes the level 1 instruction cache that holds instruction data and tags, as well as two levels of TLBs and instruction cache probing logic. 6.4.2.1 MC1_CTLIC Machine Check Control Register
This register enables the reporting, via the MCA interface, of a variety of errors detected by the Instruction Cache (IC) processor unit. For an error to be reported, both the global IC enable, MCG_CTL[ICE], and the corresponding local enable shown below must be set. If the local enable is off, the error is logged in MC1_STATUS, but not reported via the machine check exception. If the global IC enable is off, no error is logged or reported. MC1_CTL Register
63 reserved
MSR 0404h
32
31 reserved
10 9 RDDE
8 reserved
6 L2TP
5 L1TP
4 ISTP
3 IMTP
2 IDP
1 ECCM
0 ECCI
Bit 6310 9 87 6 5 4
Name reserved RDDE reserved L2TP L1TP ISTP
Function Read Data Errors L2 TLB Parity Errors L1 TLB Parity Errors Snoop tag array parity errors
R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0
Chapter 6
219
32559
Rev. 3.08 July 2007
Bit 3 2 1 0
Name IMTP IDP ECCM ECCI
Function Main tag array parity errors Data array parity errors Multi-bit ECC data errors Single-bit ECC data errors
R/W R/W R/W R/W R/W
Reset 0 0 0 0
Field Descriptions Single-bit ECC Data Errors (ECCI)Bit 0. Report single-bit ECC data errors during instruction cache line fills or TLB reloads from the internal L2 or the system. Multi-bit ECC Data Errors (ECCM)Bit 1. Report multi-bit ECC data errors during instruction cache line fills or TLB reloads from the internal L2 or the system. Data Array Parity Errors (IDP)Bit 2. Report instruction cache data array parity errors. Main Tag Array Parity Errors (IMTP)Bit 3. Report instruction cache main tag array parity errors. Snoop Tag Array Parity Errors (ISTP)Bit 4. Report instruction cache snoop tag array parity errors. L1 TLB Parity Errors (L1TP)Bit 5. Report instruction cache L1 TLB parity errors. L2 TLB Parity Errors (L2TP)Bit 6. Report instruction cache L2 TLB parity errors. Read Data Errors (RDDE)Bit 9. Report system read data errors for an instruction cache fetch if MC2_CTL[S_RDE_ALL] = 1. 6.4.2.2 MC1_CTL_MASKIC Machine Check Control Mask Register (MSR C001_0045h)
The MC1_CTL_MASK register can be used to mask the enabling of individual error reporting features in IC. Any bit set to 1 in this register inhibits writing 1s to the corresponding bits of the MC1_CTL register. See MCi_CTL_MASKMachine Check Control Mask Registers on page 210 for a description of this register. 6.4.2.3 MC1_STATUSIC Machine Check Status Register (MSR 0405h)
The MC1_STATUS MSR describes the error that was detected by IC and is very similar to MC0_STATUS. See MC0_STATUSDC Machine Check Status Register on page 215 for a description of the fields in this register.
220
Chapter 6
32559
Rev. 3.08
July 2007
Table 49.
Error Type System Line Fill L2 Cache Line Fill IC Data Load Tag Snoop/ Victim Tag Load/ Victim L1 TLB L1 TLB Multimatch L2 TLB L2 TLB Multimatch System Data Read Error
IC Error Codes
Extended Error Code 0000 0000 0000 0000 0000 0000 0001 0000 0001 0000 Error Code Type BUS Memory Memory Memory Memory TLB TLB TLB TLB BUS PP SRC SRC T 0 0 RRRR IRD IRD IRD Snoop/ Evict IRD/Evi ct IRD II MEM MEM LL LG L2 L1 L1 L1 L1 L1 L2 L2 LG TT Instruction Instruction Instruction Instruction Instruction Instruction Instruction Instruction -
Table 50.
Error Type System Line Fill L2 Cache Line Fill IC Data Load Tag Snoop/ Victim Tag Load/ Victim L1 TLB L1 TLB Multimatch L2 TLB
IC Error Status Bit Settings

UC If multi-bit If multi-bit 0 1 0 0 0 0 ADDRV 1 1 1 1/0 1/0 1 1 1 PCC If multi-bit If multi-bit 0 1 0 0 0 0 Syndrome N N N N N N N N CECC UECC SCRUB 0 0 0 0 0 0 0 0 If single- If multibit bit If single- If multibit bit 0 0 0 0 0 0 0 0 0 0 0 0
Chapter 6
221
32559
Rev. 3.08 July 2007
Table 50.
Error Type L2 TLB Multimatch System Data Read Error
IC Error Status Bit Settings (Continued)

UC 0 1 ADDRV 1 0 PCC 0 0 Syndrome N N CECC 0 0 0 UECC SCRUB 0 0
6.4.2.4
MC1_ADDRIC Machine Check Address Register (MSR 0406h)
The contents of this register are valid only if the ADDRV bit in the MC1_STATUS register is set to 1. The address can be up to 48 bits wide. In reality, not all address bits may be valid and the address field of the MC1_ADDR varies in width. The bits ranges depend on the values in the MC1_STATUS register (i.e., the type of error detected) and this is shown in Table 51. Table 51. Valid MC1_ADDR Bits
Access Type Snoop Victim Instruction Tag Array Instruction Data Array L1 TLB L2 TLB L2 Cache Data System Data System Address Out of Range None Code Read Victim Code Read Code Read Code Read Line-fill Valid Address Bits Physical[396] None Linear[476] None Linear[474] Linear[4712] for 4-Kbyte page Linear[4720] for 2-Mbyte page Linear[4712] for 4-Kbyte page Physical[396]
Error Source Snoop Tag Array
6.4.3
Bus Unit (BU)
The bus unit consists of the system bus interface logic and the L2 cache. 6.4.3.1 MC2_CTLBU Machine Check Control Register
This register enables the reporting, via the MCA interface, of a variety of errors detected in the Bus processor unit (BU). For an error to be reported, both the global BU enable, MCG_CTL[BUE], and the corresponding local enable shown below must be set. If the local enable is off, the error is logged in MC2_STATUS, but not reported via the machine check exception. If the global BU enable is off, no error is logged or reported.
222
Chapter 6
32559
Rev. 3.08
July 2007
MC2_CTL Register
63 reserved
MSR 0408h
32
31
20 19 18 17 16 15 14 13 12 11 10 9 L2D_ECCM_CPB L2D_ECCM_SNP L2T_ECCM_SCR L2D_ECCM_TLB L2D_ECC1_CPB L2D_ECC1_SNP L2T_ECC1_SCR L2D_ECC1_TLB L2T_PAR_SCR L2T_PAR_CPB L2T_PAR_SNP
8 L2T_PAR_TLB
7 L2T_PAR_ICDC
6 S_ECCM_HP
5 S_ECCM_TLB
4 S_ECC1_HP
3 S_ECC1_TLB
2 S_RDE_ALL
1 S_RDE_TLB
0 S_RDE_HP
reserved
Bit 6320 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
Name reserved L2T_ECCM_SCR L2T_ECC1_SCR L2D_ECCM_CPB L2D_ECCM_SNP L2D_ECCM_TLB L2D_ECC1_CPB L2D_ECC1_SNP L2D_ECC1_TLB L2T_PAR_SCR L2T_PAR_CPB L2T_PAR_SNP L2T_PAR_TLB L2T_PAR_ICDC S_ECCM_HP S_ECCM_TLB S_ECC1_HP S_ECC1_TLB S_RDE_ALL S_RDE_TLB S_RDE_HP
Function L2 tag array multi-bit ECC Scrub L2 tag array 1-bit ECC Scrub L2 data array multi-bit ECC copyback L2 data array multi-bit ECC snoop L2 data array multi-bit ECC TLB reload L2 data array 1-bit ECC copyback L2 data array 1-bit ECC snoop L2 data array 1-bit ECC TLB reload L2 tag array parity scrub L2 tag array parity copyback L2 tag array parity snoop L2 tag array parity TLB reload L2 tag array parity IC or DC fetch System data multi-bit ECC hardware prefetch System data multi-bit ECC TLB reload System data 1-bit ECC hardware prefetch System data 1-bit ECC TLB reload All system read data System read data TLB reload System read data hardware prefetch
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Field Descriptions System Read Data Hardware Prefetch (S_RDE_HP)Bit 0. Report system read data errors for a hardware prefetch. System Read Data TLB Reload (S_RDE_TLB)Bit 1. Report system read data errors for a TLB reload.
Chapter 6
223
32559
Rev. 3.08 July 2007
All System Read Data (S_RDE_ALL)Bit 2. Report system read data errors for any operation including a DC/IC fetch, TLB reload or hardware prefetch. System Data 1-bit ECC TLB Reload (S_ECC1_TLB)Bit 3. Report system data 1-bit ECC errors for a TLB reload. System Data 1-bit ECC Hardware Prefetch (S_ECC1_HP)Bit 4. Report system data 1-bit ECC errors for a hardware prefetch. System Data Multi-bit ECC TLB Reload (S_ECCM_TLB)Bit 5. Report system data multi-bit ECC errors for a TLB reload. System Data Multi-bit ECC Hardware Prefetch (S_ECCM_HP)Bit 6. Report system data multi-bit ECC errors for a hardware prefetch. L2 Tag Array Parity IC or DC Fetch (L2T_PAR_ICDC)Bit 7. Report L2 tag array parity errors for an IC or DC fetch. L2 Tag Array Parity TLB Reload (L2T_PAR_TLB)Bit 8. Report L2 tag array parity errors for a TLB reload. L2 Tag Array Parity Snoop (L2T_PAR_SNP)Bit 9. Report L2 tag array parity errors for a snoop. L2 Tag Array Parity Copyback (L2T_PAR_CPB)Bit 10. Report L2 tag array parity errors for a copyback. L2 Tag Array Parity Scrub (L2T_PAR_SCR)Bit 11. Report L2 tag array parity errors for a scrub. L2 Data Array 1-bit ECC TLB Reload (L2D_ECC1_TLB)Bit 12. Report L2 data array 1-bit ECC errors for a TLB reload. L2 Data Array 1-bit ECC Snoop (L2D_ECC1_SNP)Bit 13. Report L2 data array 1-bit ECC errors for a snoop. L2 Data Array 1-bit ECC Copyback (L2D_ECC1_CPB)Bit 14. Report L2 data array 1-bit ECC errors for a copyback. L2 Data Array Multi-bit ECC TLB Reload (L2D_ECCM_TLB)Bit 15. Report L2 data array multi-bit ECC errors for a TLB reload. L2 Data Array Multi-bit ECC Snoop (L2D_ECCM_SNP)Bit 16. Report L2 data array multi-bit ECC errors for a snoop. L2 Data Array Multi-bit ECC Copyback (L2D_ECCM_CPB)Bit 17. Report L2 data array multi-bit ECC errors for a copyback. L2 Tag Array 1-bit ECC Scrub (L2T_ECC1_SCR)Bit 18. Report L2 tag array 1-bit ECC errors for a scrub.
224
Chapter 6
32559
Rev. 3.08
July 2007
L2 Tag Array Multi-bit ECC Scrub (L2T_ECCM_SCR)Bit 19. Report L2 tag array multi-bit ECC errors for a scrub. 6.4.3.2 MC2_CTL_MASKBU Machine Check Control Mask Register (MSR C001_0046h)
The MC2_CTL_MASK register can be used to mask the enabling of individual error reporting features in BU. Any bit set to 1 in this register inhibits writing 1s to the corresponding bits of the MC2_CTL register. See MCi_CTL_MASKMachine Check Control Mask Registers on page 210 for a description of this register. 6.4.3.3 MC2_STATUSBU Machine Check Status Register (MSR 0409h)
The MC2_STATUS MSR describes the error that was detected by BU and is very similar to MC0_STATUS. See MC0_STATUSDC Machine Check Status Register on page 215 for a description of the fields in this register.
.
Table 52.
Error Type
BU Error Codes
Access Type Instruction Fetch Data Fetch Extended Error Code 0010 0010 0010 Error Code Type Memory Memory Memory PP T RRRR IRD DRD RD/ Snoop/ Evict GEN GEN RD Prefetch RD Prefetch RD Snoop/ Evict II LL L2 L2 L2 TT Instruction Data Generic
Tag Parity
TLB/ Snoop/ Evict Scrub Scrub TLB Hardware Prefetch TLB Hardware Prefetch TLB Snoop/ Evict
0010 0010 0000 0000 0000 0000 0000 0000
Memory Memory BUS BUS BUS BUS Memory Memory
SRC SRC SRC SRC -
0 0 0 0 -
MEM MEM/IO MEM/IO MEM -
L2 L2 LG LG LG LG L2 L2
Generic Instruction Generic Generic
Tag ECC System Data Read Error System Data ECC Data ECC Data Copyback
Chapter 6
225
32559
Rev. 3.08 July 2007
Table 53.
Error Type
BU Error Status Bit Settings

Access Type Instruction Fetch Data Fetch UC 1 1 1 ADDRV 1 1 1 PCC 1 1 1 CECC 0 0 0 UECC 0 0 0 SCRUB 0 0 0
Tag Parity
TLB/ Snoop/ Evict Scrub Scrub TLB Hardware Prefetch TLB
1 If multi-bit 1 1 If multi-bit If multi-bit If multi-bit If multi-bit
1 1 1 0 1 0 1 1
0 0 0 0
1 1 0 0 0 0 0 0
Tag ECC System Data Read Error System Data ECC Data ECC Data Copyback
If single- If multibit bit 0 0 0 0
If multi-bit If single- If multibit bit If multi-bit If single- If multibit bit If multi-bit If single- If multibit bit If multi-bit If single- If multibit bit
Hardware Prefetch TLB Snoop/ Evict
6.4.3.4
MC2_ADDRBU Machine Check Address Register (MSR 040Ah)
The contents of this register are valid only if the ADDRV bit in the MC2_STATUS register is set to 1. The address can be up to 48 bits wide. In reality, not all address bits may be valid and the address field of the MC2_ADDR varies in width. The bit ranges depend on the values in the MC2_STATUS register (i.e., the type of error detected) and this is shown in Table 54. Table 54. Valid MC2_ADDR Bits
Access Type Any Any Valid Address Bits Physical[396] MC2_ADDR[30] contains encoded cache way Physical[156] for 1-Mbyte L2 Physical[146] for 512-Kbyte L2 Physical[136] for 256-Kbyte L2 Physical[126] for 128-Kbyte L2 System Address Out of Range Any Physical[396]
Error Source L2 CacheData Array L2 CacheTag Array
226
Chapter 6
32559
Rev. 3.08
July 2007
6.4.4
6.4.4.1
Load Store Unit (LS)

MC3_CTLLS Machine Check Control Register
This register enables the reporting, via the MCA interface, of two types of errors detected by the Load/Store (LS) processor unit. For an error to be reported, both the global LS enable, MCG_CTL[LSE], and the corresponding local enable shown below must be set. If the local enable is off, the error is logged in MC3_STATUS, but not reported via the machine check exception. If the global LS enable is off, no error is logged or reported. MC3_CTL Register
63 reserved
MSR 040Ch
32
31 reserved
1 SRDES
0 SRDEL
Bit 633 2 1 0
Name reserved reserved S_RDE_S S_RDE_L
Function
R/W R/W
Reset 0 0 0 0
Read Data Errors on Store Read Data Errors on Load
R/W R/W
Field Descriptions Read Data Errors on Load (S_RDE_L)Bit 0. Report system read data errors on a load if MC2_CTL[S_RDE_ALL] = 1. Read Data Errors on Store (S_RDE_S)Bit 1. Report system read data errors on a store if MC2_CTL[S_RDE_ALL] = 1. 6.4.4.2 MC3_CTL_MASKLS Machine Check Control Mask Register (MSR C001_0047h)
The MC3_CTL_MASK register can be used to mask the enabling of individual error reporting features in LS. Any bit set to 1 in this register inhibits writing 1s to the corresponding bits of the MC3_CTL register. See MCi_CTL_MASKMachine Check Control Mask Registers on page 210 for a description of this register.
Chapter 6
227
32559
Rev. 3.08 July 2007
6.4.4.3
MC3_STATUSLS Machine Check Status Register (MSR 040Dh)
The MC3_STATUS MSR describes the error that was detected by LS and is very similar to MC0_STATUS. See MC0_STATUSDC Machine Check Status Register on page 215 for a description of the fields in this register.
.
Table 55.
Error Type Read Data on Store Read Data on Load
LS Error Codes
Extended Error Code 0000 0000 Error Code Type BUS BUS PP SRC SRC T 0 0 RRRR DWR DRD II MEM MEM/IO LL LG LG TT -
Table 56.
Error Type Read Data on Store Read Data on Load
LS Error Status Bit Settings

UC 1 1 ADDRV 1/0 1/0 PCC 1/0 1/0 Syndrome N N CECC 0 0 UECC 0 0 SCRUB 0 0
6.4.4.4
MC3_ADDRLS Machine Check Address Register (MSR 040Eh)
The contents of this register are valid only if the ADDRV bit in the MC3_STATUS register is set to 1. The only type of error recorded by the LS machine check mechanism is a system address out of range or read data error for which MC3_ADDR[39:0] store the physical address.
6.4.5
Northbridge (NB)
The Northbridge and DRAM memory controller are included in this block. 6.4.5.1 MC4_CTLNB Machine Check Control Register (MSR 0410h)
This register enables the reporting, via the MCA interface, of the errors detected by the North Bridge (NB) processor unit. MCA NB Control Register (Function 3, Offset 40h) maps to MC4_CTL[31:0]. See MCA NB Control Register on page 133 for detailed information on this register. For an error to be reported, both the global NB enable, MCG_CTL[NBE], and the corresponding local enable must be set. If the local enable is off, the error is logged in MC4_STATUS, but not reported via the machine check exception. If the global NB enable is off, no error is logged or reported.
228
Chapter 6
32559
Rev. 3.08
July 2007
6.4.5.2
MC4_CTL_MASKNB Machine Check Control Mask Register (MSR C001_0048h)
The MC4_CTL_MASK register can be used to mask the enabling of individual error reporting features in the Northbridge. Any bit set to 1 in this register inhibits writing1s to the corresponding bits of the MC4_CTL register. See MCi_CTL_MASKMachine Check Control Mask Registers on page 210 for a description of this register. 6.4.5.3 MC4_STATUSNB Machine Check Status Register (MSR 0411h)
The MC4_STATUS MSR describes the error that was detected by the Northbridge. MCA NB Status Low Register (Function 3, Offset 48h) maps to MC4_STATUS MSR[31:0] and MCA NB Status High Register (Function 3, Offset 4Ch) maps to MC4_STATUS MSR[63:32]. See MCA NB Status Low Register on page 139 and MCA NB Status High Register on page 142 for more detailed information. 6.4.5.4 MC4_ADDRNB Machine Check Address Register (MSR 0412h)
The contents of this register are valid only if the ErrAddrVal bit in the MC4_STATUS register is set to 1. MCA NB Address Low Register (Function 3, Offset 50h) maps to MC4_ADDR[31:0] and MCA NB Address High Register (Function 3, Offset 54h) maps to MC4_ADDR[63:32]. See MCA NB Address Low Register on page 145 and MCA NB Address High Register on page 145 for a description of this register. 6.4.5.5 MC4_MISCDRAM Errors Threshold Register
This register is used to count correctable and uncorrectable ECC errors that occur during DRAM read operations. MC4_MISCDRAM Errors Threshold Register
63 62 61 60 Locked CtrP Val reserved 56 55 LvtOff 52 51 50 49 48 47 IntType Ovrflw CntEn reserved 44 43 ErrCount
MSR 0413h
32
31 reserved
Bits 63 62 61 6056 5552
Mnemonic Val CtrP Locked reserved LvtOff
Function Valid Counter Present Locked LVT Offset
R/W R R R R R
Reset 1 1 0 0 0
Chapter 6
229
32559
Rev. 3.08 July 2007
Bits 51 5049 48 4744 4332 310
Mnemonic CntEn IntType Ovrflw reserved ErrCount reserved
Function Counter Enable Interrupt Type Overflow Error Count
R/W R/W R/W R/W R R/W R
Reset 0 X X 0 X 0
Field Descriptions Error Count (ErrCount)Bits 4332. This field is written by software to set the starting value of the error counter. This field is incremented by hardware when errors are logged. When this counter overflows, it stays at FFFh (no rollover). To set the threshold value, software should subtract the desired error count (the number of errors necessary in order for an interrupt to be taken) from FFFh and write the result into this field. This field is read only if the Locked bit (MSR 0413h) is set. Overflow (Ovrflw)Bit 48. This bit is set by hardware when ErrCount transitions from FFEh to FFFh. When this bit is set, the interrupt selected by the IntType field is generated. This bit can be cleared by writing a zero to the bit. This bit is read only if the Locked bit (MSR 0413h) is set. Interrupt Type (IntType)Bits 5049. This field determines the type of interrupt signaled when the counter overflows. This field is read only if the Locked bit (MSR 0413h) is set. 00b = No Interrupt. 01b = APIC based interrupt. 10b = SMI. 11b = Reserved. Counter Enable (CntEn)Bit 51. When this bit is set, DRAM ECC error counting is enabled. This field is read only if the Locked bit (MSR 0413h) is set. LVT Offset (LvtOff)Bits 5552. This field is shifted left by four bits and added to 500h to determine the address of the LVT entry in the APIC registers. Locked (Locked)Bit 61. This bit is set by BIOS to indicate that this register is not available for OS use. When this bit is set, write to this register are silently ignored. BIOS should set this bit if IntType is set to SMI. This bit is read/write if the MCi_STATUS_WREN bit (MSR C001_0015h bit 18) is set. Counter Present (CtrP)Bit 62. This bit indicates the presence of the ErrCount field in the MISC register. This bit is read/write if the MCi_STATUS_WREN bit (MSR C001_0015h bit 18) is set. Valid (Valid)Bit 63. This bit indicates the presence of a valid CtrP field in the MISC register. This bit is read/write if the MCi_STATUS_WREN bit (MSR C001_0015h bit 18) is set.
230
Chapter 6
32559
Rev. 3.08
July 2007
6.5
Initializing the Machine Check Mechanism
Following is the general process system software should follow to initialize the machine check mechanism: 1. Execute the CPUID and verify that the processor supports the machine check exception (MCE) and MCA. MCE is supported when EDX bit 7 is set to 1, and MCA is supported when EDX bit 14 set to 1. Software should not proceed with initializing the machine check mechanism if MCE is not supported. 2. If MCA is supported, system software should take the following steps: a. Check to see if the MCG_CTL_P bit in the MCG_CAP register is set to 1. If it is, then the MCG_CTL register is supported by the processor. When this register is supported, software should set its enable bits to 1 for the machine check features it uses. Software can load MCG_CTL with all 1s to enable all machine check features. b. Read the COUNT field from the MCG_CAP register to determine the number of errorreporting register banks supported by the processor. This is set to 5 in AMD NPT Family 0Fh Processors since there are five blocksDC, IC, BU, LS, and NB. For each error-reporting register bank, software should set the enable bits to 1 in the MCi_CTL register for the error types it wants the processor to report. Software can load each MCi_CTL register bit with a 1 to enable all error-reporting mechanisms. The error-reporting register banks are numbered from 0 to one less than the value found in the MCG_CAP.COUNT field. For example, when the COUNT field indicates that 5 register banks are supported, they are numbered 0 to 4. c. For each error-reporting register bank, software should clear all status fields in the MCi_STATUS register by writing all 0s to the register. It is possible that valid error status is reported in the MCi_STATUS registers at the time software clears them. The status can reflect fatal errors recorded before a processor reset or errors recorded during the system power-up and boot process. Prior to clearing the MCi_STATUS registers, software should examine their contents and log any errors found. 3. As a final step in the initialization process, system software should enable the machine check exception by setting CR4.MCE (bit 6) to 1.
6.6

Using Machine Check Features
System software can detect and handle machine check errors using two methods: Software can periodically examine the machine check status registers for reported errors, and log any errors found. Software can enable the machine check exception (#MC). When an uncorrectable error occurs, the processor immediately transfers control to the machine check exception handler. In this case, system software provides a machine check exception handler that, at a minimum, logs detected Machine Check Architecture 231
Chapter 6
32559
Rev. 3.08 July 2007
errors. The exception handler can be designed for a specific processor implementation or can be generalized to work on multiple implementations.
6.6.1
Handling Machine Check Exceptions
The processor uses the interrupt control-transfer mechanism to invoke an exception handler after a machine check exception occurs. This requires system software to initialize the interrupt-descriptor table (IDT) with either an interrupt gate or a trap gate that references the interrupt handler. At a minimum, the machine check exception handler must be capable of logging errors for later examination. Because most machine check errors are not recoverable, the ability to log errors can be sufficient for implementation of the handler. More thorough exception-handler implementations can analyze errors to determine if each error is recoverable. If a recoverable error is identified, the exception handler can attempt to correct the error and restart the interrupted program. Machine check exception handlers that attempt to correct recoverable errors must be thorough in their analysis and the corrective actions they take. The following guidelines should be used when writing such a handler: All status registers in the error-reporting register banks must be examined to identify the cause or causes of the machine check exception. Read the COUNT field from MCG_CAP to determine the number of status registers supported by the processor. The status registers are numbered from 0 to one less than the value found in the MCG_CAP.COUNT field. For example, if the COUNT field indicates five status registers are supported, they are numbered MC0_STATUS to MC4_STATUS. Check the valid bit in each status register (MCi_STATUS.VAL). The MCi_STATUS register does not need to be examined when its valid bit is clear. Check the valid MCi_STATUS registers to see if error recovery is possible. Error recovery is not possible when: The processor-context corrupt bit (MCi_STATUS.PCC) is set to 1. The error-overflow status bit (MCi_STATUS. OVER) is set to 1. This bit indicates that more than one machine check error has occurred, but only one error is reported by the status register.
If error recovery is not possible, the handler should log the error information and return to the operating system. Check the MCi_STATUS.UC bit to see if the processor corrected the error. If UC=1, the processor did not correct the error, and the exception handler must correct the error prior to restarting the interrupted program. If the handler cannot correct the error, it should log the error information and return to the operating system. When identifying the error condition, portable exception handlers should examine only the MCA error-code field of the MCi_STATUS register. See the error codes in tables 17 through 22 for information on interpreting this field.
232
Chapter 6
32559
Rev. 3.08
July 2007
If the MCG_STATUS.RIPV bit is set to 1, the interrupted program can be restarted reliably at the instruction-pointer address pushed onto the exception-handler stack. If RIPV = 0, the interrupted program cannot be restarted reliably, although it may be possible to restart it for debugging purposes. When logging errors, particularly those that are not recoverable, check the MCG_STATUS.EIPV bit to see if the instruction-pointer address pushed onto the exception-handler stack is related to the machine check error. If EIPV = 0, the address is not guaranteed to be related to the error. Prior to exiting the machine check handler, be sure to clear MCG_STATUS.MCIP to 0. MCIP indicates that a machine check exception occurred. If this bit is set when another machine check exception occurs, the processor enters the shutdown state. When an exception handler is able to, at a minimum, successfully log an error condition, clear the MCi_STATUS registers prior to exiting the machine check handler. Software is responsible for clearing at least the MCi_STATUS.VAL bit. Additional machine check exception-handler portability can be added by having the handler use the CPUID instruction to identify the processor and its capabilities. Implementation-specific software can be added to the machine check exception handler based on the processor information reported by CPUID. Reporting Correctable Machine Check Errors
6.6.1.1
Machine check exceptions do not occur if the error is correctable by the processor. If system software wishes to log and report correctable machine check errors, a system-service routine must be provided to check the contents of the machine check status registers for correctable errors. The service routine can be invoked by system software on a periodic basis, or it can be manually invoked by the user as needed. Assuming that the processor supports the machine check registers, a service routine that reports correctable errors should perform the following: 1. Examine each status register (MCi_STATUS) in the error-reporting register banks. For each MCi_STATUS register with a set valid bit (VAL = 1), the service routine should: a. Save the contents of the MCi_STATUS register. b. Save the contents of the corresponding MCi_ADDR register if MCi_STATUS.ADDRV = 1. c. Save the contents of the corresponding MCi_MISC register if MCi_STATUS.MISCV = 1. d. Check to see if MCG_STATUS.MCIP = 1, indicating that the machine check exception handler is in progress. If this is the case, then the machine check exception handler has called the service routine to log the errors. In this situation, the error-logging service routine should determine whether or not the interrupted program is restartable, and report the determination back to the exception handler. The program is not restartable if either of the following is true: MCi_STATUS.PCC = 1, indicating the processor context is corrupted. MCG_STATUS.RIPV = 0, indicating the interrupted program cannot be restarted reliably at the instruction-pointer address pushed onto the exception-handler stack. Chapter 6 Machine Check Architecture 233
32559
Rev. 3.08 July 2007
2. Once the information found in the error-reporting register banks is saved, the MCi_STATUS register should be cleared to 0. This allows the processor to properly report any subsequent errors in the MCi_STATUS registers. 3. In multiprocessor configurations, the service routine can save the processor node identifier. This can help locate a failing multiprocessor-system component, which can then be isolated from the rest of the system.
234
Chapter 6
32559
Rev. 3.08
July 2007
Advanced Programmable Interrupt Controller (APIC)
All AMD NPT Family 0Fh Processor-based systems must support 64-bit operating systems. As a result, all platforms must support APIC and declare it to the operating system in the appropriate tables. The local APIC contains logic to receive interrupts from a variety of sources and to send interrupts to other local APICs, as well as registers to control its behavior and report status. Interrupts can be received from: HyperTransport I/O devices, including the I/O hub (I/O APICs) Other local APICs (inter-processor interrupts) APIC timer Performance counters Legacy local interrupts from the I/O hub (INTR and NMI) APIC internal errors
The APIC timer, performance counters, local interrupts, and internal errors are all considered local interrupt sources, and their routing is controlled by local vector table entries. These entries assign a message type and vector to each interrupt, allow them to be masked, and track the status of the interrupt. I/O and inter-processor interrupts have their message type and vector assigned at the source and are unaltered by the local APIC. They carry a destination field and a mode bit that together determine which local APICs accepts them. The destination mode (DM) bit specifies if the interrupt request packet should be handled in Physical or Logical destination mode. In physical destination mode, if the destination field is FFh, the interrupt is a broadcast and is accepted by all local APICs. Otherwise, the interrupt is only accepted by the local APIC whose APIC ID matches the destination field of the interrupt. Physical mode allows up to 255 APICs to be addressed individually.
Chapter 7
235
32559
Rev. 3.08 July 2007
In logical destination mode, a local APIC accepts interrupts selected by the logical destination register (LOG_DEST) and the destination field of the interrupt using either cluster or flat format, as configured by the destination format register. If flat destinations are in use, bits 70 of the LOG_DEST field are checked against bits 70 of the arriving interrupts Destination Field. If any bit position is set in both fields, this APIC is a valid destination. Flat format allows up to 8 APICs to be addressed individually. If cluster destinations are in use, bits 74 of the LOG_DEST field are checked against bits 7 4 of the arriving interrupts destination field, identifying the cluster. If all of bits 74 match, then bits 30 of the LOG_DEST and the interrupt destination are checked for any bit positions that are set in both fields, choosing processors within the cluster. If both conditions are met, this APIC is a valid destination. Cluster format allows 15 clusters of 4 APICs each to be addressed. In both flat and cluster formats, if the destination field is FFh, the interrupt is a broadcast and is accepted by all local APICs.
7.1
Interrupt Delivery
SMI, NMI, INIT, Startup, and External interrupts are classified as non-vectored interrupts. When an APIC accepts a non-vectored interrupt, it is handled directly by the processor instead of being queued in the APIC. When an APIC accepts a fixed or lowest priority interrupt, it sets the bit in the interrupt request register (IRR) corresponding to the vector in the interrupt. (For local interrupt sources, this comes from the vector field in that interrupts local vector table entry.) If a subsequent interrupt with the same vector arrives when the corresponding IRR bit is already set, the two interrupts are collapsed into one. Vectors 015 are reserved.
7.2
Vectored Interrupt Handling
The task priority register (TPR_PRI) and processor priority register (PROC_PRI) each contain an 8bit priority, divided into a main priority (bits 74) and a priority sub-class (30). The task priority is assigned by software to set a threshold priority at which the processor is interrupted. To calculate processor priority, an 8-bit ISR vector is set based on the highest bit set in the in-service register (ISR). Like the TPR_PRI and PROC_PRI, this vector is divided into a main priority (74) and priority sub-class (30). The main priority of the TPR_PRI and ISR are compared and the highest is the PROC_PRI, as follows:
If (TPR_PRI[7:4] >= ISRVect[7:4])PROC_PRI = TPR_PRI Else PROC_PRI = {ISRVect[7:4],0h}
The resulting processor priority is used to determine if any accepted interrupts (indicated by IRR bits) are high enough priority to be serviced by the processor. When the processor is ready to service an interrupt, the highest bit in the IRR is cleared, and the corresponding bit is set in the ISR. The
236
Chapter 7
32559
Rev. 3.08
July 2007
corresponding trigger mode register (TMR) bit is set if the interrupt is level-triggered and cleared if edge-triggered. When the processor has completed service for an interrupt, it performs a write to the end of interrupt (EOI) register, clearing the highest ISR bit and causing the next-highest interrupt to be serviced. If the corresponding TMR bit is set, EOI messages is sent to all APICs to complete service of the interrupt at the source.
7.3
Spurious Interrupts
In the event that the task priority is set to or above the level of the interrupt to be serviced, the local APIC delivers a spurious interrupt vector to the processor, as specified by the spurious interrupt vector register. The ISR is not changed and no EOI occurs.
7.3.1
Spurious Interrupts Caused by Timer Tick Interrupt
A typical interrupt is asserted until it is serviced. Interrupt is deasserted when software clears the interrupt status bit within the interrupt service routine. Timer tick interrupt is an exception, since it is deasserted regardless of whether it is serviced or not. A BIOS programs the 8254 Programmable Timer to generate an 18.2 Hz square wave. This square wave represents the timer tick interrupt (PITIRQ) and in some interrupt configurations it is connected to the 8259 Programmable Interrupt Controller (PIC) IRQ0 input. The processor is not always able to service interrupts immediately (i.e. when interrupts are masked by clearing EFLAGS.IM or when the processor is in debug mode). The method of interrupt delivery to the processor (wire or messages) determines system behavior for the timer tick interrupt connected to the PIC after the processor has not been able to service it for some time. The PIC output INTR is identical to PITIRQ when interrupts are not serviced. If interrupts are delivered to the processor using a wire, and the processor is not able to service the timer tick interrupt for some time, timer tick interrupts asserted during that time are lost. The following cases are possible when the processor is ready to service interrupts: INTR is deasserted and it is not serviced by the processor. This happens almost 50 percent of the time. INTR is asserted and it is serviced by the processor. This happens almost 50 percent of the time. INTR is asserted, and it is detected by the processor. The processor sends the interrupt acknowledge cycle, but when the PIC receives it, INTR is deasserted, and the PIC sends a spurious interrupt vector. This happens a very small percentage of the time.
The probability of spurious interrupts, when interrupts are delivered to the processor using a wire, is very low. An example of a configuration that uses wire to deliver interrupts to the processor has PITIRQ connected to PIC IRQ0, and PIC INTR connected to LINT0 of an AMD Athlon processor.
Chapter 7
237
32559
Rev. 3.08 July 2007
If interrupts are delivered to the processor using messages, and the processor is not able to service the timer tick interrupt for an extended period of time, the INTR caused by the first timer tick interrupt asserted during that time is delivered to the local APIC in ExtInt mode and latched, and the subsequent timer tick interrupts are lost. The following cases are possible when the processor is ready to service interrupts: An ExtInt interrupt is pending, and INTR is asserted. This results in timer tick interrupt servicing. This occurs 50 percent of the time. An ExtInt interrupt is pending, and INTR is deasserted. The processor sends the interrupt acknowledge cycle, but when the PIC receives it, INTR is deasserted, and the PIC sends a spurious interrupt vector. This occurs 50 percent of the time.
There is a 50 percent probability of spurious interrupts when interrupts are delivered to the processor using messages. An example of an AMD NPT Family 0Fh Processor configuration that uses messages to deliver interrupts to the processor has PITIRQ connected to PIC IRQ0, PIC INTR connected to IOAPIC IRQ0, and interrupts are delivered to the processor by HyperTransport link messages. HyperTransport link messages are always used to deliver PIC interrupts to the processor in a system with AMD NPT Family 0Fh Processors. An example of an AMD Athlon processor configuration that uses messages to deliver interrupts to the processor has PITIRQ connected to PIC IRQ0, PIC INTR connected to IOAPIC IRQ0, and interrupts are delivered to the processor by APIC 3-wire messages.
7.4
Lowest-Priority Arbitration
Fixed, remote read, and non-vectored interrupts are accepted by their destination APICs without arbitration. Delivery of lowest-priority interrupts requires all APICs to arbitrate to determine which one accepts the interrupt. If focus processor checking is enabled (bit 9 of the spurious interrupt vector register cleared), then the focus processor for an interrupt always accept the interrupt. A processor is the focus of an interrupt if it is already servicing that interrupt (corresponding ISR bit is set) or if it already has a pending request for that interrupt (corresponding IRR bit is set). If there is no focus processor for an interrupt, or focus processor checking is disabled, then each APIC calculates an arbitration priority value, and the one with the lowest result accepts the interrupt. To calculate arbitration priority, an 8-bit IRR Vector is set based on the highest bit set in the IRR. Like the ISR, TPR_PRI, and PROC_PRI, this vector is divided into a main priority (74) and priority subclass (30). The main priority of the TPR_PRI, ISR, and IRR are compared and the highest is the new ARB_PRI, as follows:
If (TPR_PRI[7:4] >= IRRVect[7:4] and TPR_PRI[7:4] > ISRVect[7:4]) ARB_PRI = TPR_PRI Else if (IRRVect[7:4] > ISRVect[7:4]) ARB_PRI = {IRRVect[7:4],0h} Else ARB_PRI = {ISRVect[7:4],0h}
238
Chapter 7
32559
Rev. 3.08
July 2007
7.5
Inter-Processor Interrupts
In order to redirect an interrupt to another processor, originate an interrupt to another processor, or allow a processor to interrupt itself, the interrupt command register (ICR) provides a mechanism for generating interrupts. A write to the low doubleword of the ICR causes an interrupt to be generated, with the properties specified by the ICR low and ICR high fields.
7.6
APIC Timer Operation
The local APIC contains a 32-bit timer, controlled by the Timer LVT entry, initial count, and divide configuration registers. The processor bus clock is divided by the value in the timer divide configuration register to obtain a time base for the timer. When the timer initial count register is written, the value is copied into the timer current count register. The current count register is decremented at the rate of the divided clock. When the count reaches 0, a timer interrupt is generated with the vector specified in the Timer LVT entry. If the Timer LVT entry specifies periodic operation, the current count register is reloaded with the initial count value, and it continues to decrement at the rate of the divided clock. If the mask bit in the timer LVT entry is set, timer interrupts are not generated.
7.7
State at Reset
At power-up or reset, all registers take on the values listed in their descriptions. SMI, NMI, INIT, Startup, remote read, and LINT interrupts may be accepted. When the APIC is software-disabled, pending interrupts in the ISR and IRR are held, but further fixed, lowest-priority, and ExtInt interrupts are not accepted. All LVT entry mask bits are set and cannot be cleared. When a processor accepts an INIT interrupt, the APIC is reset as at power-up, with the exception that APIC_ID is unaffected.
Chapter 7
239
32559
Rev. 3.08 July 2007
7.8
Register Summary
All APIC registers are accessed with memory reads and writes of (APIC_BASE+Offset). APIC_BASE is MSR 001Bh and defaults to 00_FEE0_0000h. The APIC registers and their offsets are listed in Table 57. Table 57.
Offset (Hex) 20 30 80 90 A0 B0 C0 D0 E0 F0 100-170 180-1F0 200-270 280 300 310 320 330 340 350 360 370 380 390
APIC Register Summary

Mnemonic APIC_ID APIC_VER TPR_PRI ARB_PRI PROC_PRI EOI RMT_RD LOG_DEST DEST_FORMAT SPUR_INTR_VEC ISR TMR IRR ERR_STAT ICRLO ICRHI TIMER_LVT THERMAL_LVT PERF_CNT_LVT LINT0_LVT LINT1_LVT ERROR_LVT INIT_CNT CURR_CNT Name APIC ID Register APIC Version Register Task Priority Register Arbitration Priority Register Processor Priority Register End Of Interrupt Register Remote Read Register Logical Destination Register Destination Format Register Spurious Interrupt Vector Register In-Service Registers Trigger Mode Registers Interrupt Request Registers Error Status Register Interrupt Command Register Low (bits 310) Interrupt Command Register High (bits 6332) Timer Local Vector Table Entry Thermal Local Vector Table Entry Performance Counter Local Vector Table Entry Local Interrupt 0 Local Vector Table Entry Local Interrupt 1 Local Vector Table Entry Error Local Vector Table Entry Timer Initial Count Register Timer Current Count Register
240
Chapter 7
32559
Rev. 3.08
July 2007
Table 57.
3E0 400 500
APIC Register Summary (Continued)

TIMER_DVD_CFG EXT_APIC_FEAT THRCNT_INT0_LVT Timer Divide Configuration Register Extended APIC Feature Register Threshold Count Interrupt 0 Local Vector Table Entry
7.8.1
APIC ID Register
Offset 20h
24 23 APICID reserved 0
APIC_ID Register
31
Bit 3124 230
Name APICID reserved
Function APIC Identification
R/W R/W R/O
Reset Node ID 00_0000h
Field Descriptions APIC Identification (APICID)Bits 3124. Software must ensure that all APICs are assigned unique APIC IDs. When both ApicExtId and ApicExtBrdCst in the HyperTransport Transaction Control Register are set, all 8 bits of APIC ID are used. When either ApicExtID or ApicExtBrdCst is clear, only bits 30 of APIC ID are used, and bits 74 are reserved. Node ID is initially 00h if this is the boot strap processor or 07h for all other nodes.
7.8.2
APIC Version Register

Offset 30h
16 15 8 7 0
APIC_VER Register
31 30 ExtApicSpace 24 23
reserved
MaxLVTEntry
reserved
Version
Bit 31 3024 2316 158 70
Name ExtApicSpace reserved MaxLVTEntry reserved Version
Function Extended APIC Register Space Present Maximum LVT Entry Version
R/W R/O R/O R/O R/O R/O
Reset 1h 00h 04h 00h 10h
Chapter 7
241
32559
Rev. 3.08 July 2007
Field Descriptions Extended APIC Register Space PresentBit 31. This field indicates the presence extended APIC register space starting at 400h. Max LVT EntryBits 2316. This field indicates the number of entries in the Local Vector Table minus one. VersionBits 70. This field indicates the version number of this APIC implementation.
7.8.3
Task Priority Register

Offset 80h
8 reserved 7 Priority 0
TPR_PRI Register
31
Bit 318 70
Name reserved Priority
Function Priority
R/W R/O R/W
Reset 0000_00h 00h
Field Descriptions PriorityBits 70. This field is assigned by software to set a threshold priority at which the processor is interrupted.
7.8.4
Arbitration Priority Register

Offset 90h
ARB_PRI Register
31
Bit 318 70
Function Priority
R/W R/O R/O
Reset 0000_00h 00h
Field Descriptions PriorityBits 70. This field indicates the processors current priority, for a task being serviced, an interrupt being serviced, or an interrupt that is pending, and is used to arbitrate between processors to determine which accepts a lowest-priority interrupt request. 242 Advanced Programmable Interrupt Controller (APIC) Chapter 7
32559
Rev. 3.08
July 2007
7.8.5
Processor Priority Register

Offset A0h
PROC_PRI Register
31
Bit 318 70
Function Priority
R/W R/O R/O
Reset 0000_00h 00h
Field Descriptions PriorityBits 70. This field indicates the processors current priority servicing a task or interrupt, and is used to determine if any pending interrupts should be serviced.
7.8.6
End of Interrupt Register
This register is written by the software interrupt handler to indicate that servicing of the current interrupt is complete. EOI Register
31 reserved
Offset B0h
0
Bit 310
Name reserved
Function
R/W W/O
Reset XXXX_XXXXh
7.8.7
Remote Read Register

Offset C0h
0 RemoteReadData
RMT_RD Register
31
Bit 310
Name RemoteReadData
Function Remote Read Data
R/W R/O
Reset 0000_0000h
Chapter 7
243
32559
Rev. 3.08 July 2007
Field Descriptions Remote Read Data (RemoteReadData)Bits 310. This field contains the data resulting from a valid completion of a Remote Read Inter-Processor Interrupt.
7.8.8
Logical Destination Register

Offset D0h
0 reserved
LOG_DEST Register
31 Destination 24 23
Bit 3124 230
Name Destination reserved
Function Destination
R/W R/W R/O
Reset 00h 00_0000h
Field Descriptions Destination (Destination)Bits 3124. This field contains this APICs destination identification, and is used to determine which interrupts should be accepted.
7.8.9
Destination Format Register

Offset E0h
0 reserved
DEST_FORMAT Register
31 Format 28 27
Bit 3128 270
Name Format reserved
Function Format
R/W R/W R/O
Reset Fh FFF_FFFFh
Field Descriptions Format (Format)Bits 3128. 0h and Fh are the only allowed values. This field controls which format to use when accepting interrupts with a logical destination mode. 0h = Cluster destinations are used. Fh = Flat destinations are used.
244
Chapter 7
32559
Rev. 3.08
July 2007
7.8.10
Spurious Interrupt Vector Register

Offset F0h
10 9 FocusDisable 8 APICSWEn 7 0
SPUR_INTR_VEC Register
31
reserved
Vector
Bit 3110 9 8 70
Name reserved FocusDisable APICSWEn Vector
Function Focus Disable APIC Software Enable Vector
R/W R/O R/W R/W R/W
Reset 0000_00h 0 0 FFh
Field Descriptions Focus Disable (FocusDisable)Bit 9. This bit disables focus processor checking during lowestpriority arbitrated interrupts. APIC Software Enable (APICSWEn)Bit 8. When APICSWEnable is cleared, SMI, NMI, INIT, Startup, Remote Read, and LINT interrupts may be accepted, pending interrupts in the ISR and IRR are held, but further Fixed, Lowest-Priority, and ExtInt interrupts are not accepted. All LVT entry Mask bits are set and cannot be cleared. VectorBits 70. When ApicExtSpur in the HyperTransport Transaction Control Register is set, bits 30 of Vector are writable. When ApicExtSpur is clear, bits 30 are read-only 1111b. This field contains the vector that is sent to the processor in the event of a spurious interrupt.
TM
7.8.11
In-Service Registers
Offset 100h
16 15 InServiceBits reserved 0
ISR Register
31
Bit 3116 150
Name InServiceBits reserved
Function In-Service Bits
R/W R/O R/O
Reset 0000h 0000h
Chapter 7
245
32559
Rev. 3.08 July 2007
Field Descriptions In-Service Bits (InServiceBits)Bits 3116. These bits are set when the corresponding interrupt is being serviced by the CPU. ISR Registers
31 InServiceBits
Offsets 170h, 160h, 150h, 140h, 130h, 120h, 110h

0
Bit 310
Name InServiceBits
Function In-Service Bits
R/W R/O
Reset 0000_0000h
Field Descriptions In-Service Bits (InServiceBits)Bits 310. These bits are set when the corresponding interrupt is being serviced by the CPU. Interrupts are mapped as follows:
ISR Offset 110h Offset 120h Offset 130h Offset 140h Offset 150h Offset 160h Offset 170h Interrupt number 6332 9564 12796 159128 191160 223192 255224
7.8.12
Trigger Mode Registers

Offset 180h
16 15 TriggerModeBits reserved 0
TMR Register
31
Bit 3116 150
Name TriggerModeBits reserved
Function Trigger Mode Bits
R/W R/O R/O
Reset 0000h 0000h
246
Chapter 7
32559
Rev. 3.08
July 2007
Field Descriptions Trigger Mode Bits (TriggerModeBits)Bits 3116. The Trigger Mode bit for each corresponding interrupt is updated when an interrupt enters servicing. It is 0 for edge-triggered interrupts and 1 for level-triggered interrupts. TMR Registers
31 Trigger Mode Bits
Offsets 1F0h, 1E0h, 1D0h, 1C0h, 1B0h, 1A0h, 190h

0
Bit 310
Name TriggerModeBits
Function Trigger Mode Bits
R/W R/O
Reset 0000_0000h
Field Descriptions Trigger Mode Bits (TriggerModeBits)Bits 310. The Trigger Mode bit for each corresponding interrupt is updated when an interrupt enters servicing. It is 0 for edge-triggered interrupts and 1 for level-triggered interrupts. Interrupts are mapped as follows:
TMR Offset 190h Offset 1A0h Offset 1B0h Offset 1C0h Offset 1D0h Offset 1E0h Offset 1F0h Interrupt number 6332 9564 12796 159128 191160 223192 255224
7.8.13
Interrupt Request Registers

Offset 200h
16 15 RequestBits reserved 0
IRR Register
31
Bit 3116 150
Name RequestBits reserved
Function Interrupt Requests Bits
R/W R/O R/O
Reset 0000h 0000h
Chapter 7
247
32559
Rev. 3.08 July 2007
Field Descriptions Interrupt Request Bits (RequestBits)Bits 3116. Request bits are set when the corresponding interrupt is accepted by the APIC. IRR Registers
31 RequestBits
Offsets 270h, 260h, 250h, 240h, 230h, 220h, 210h

0
Bit 310
Name RequestBits
Function Interrupt Requests Bits
R/W R/O
Reset 0000_0000h
Field Descriptions Interrupt Request Bits (RequestBits)Bits 310. Request bits are set when the corresponding interrupt is accepted by the APIC. Interrupts are mapped as follows:
IRR Offset 210h Offset 220h Offset 230h Offset 240h Offset 250h Offset 260h Offset 270h Interrupt number 6332 9564 12796 159128 191160 223192 255224
7.8.14
Error Status Register
This register must be written to trigger an update before it can be read. Each write causes the internal error state to be loaded into this register, clearing the internal error state. A second write before another error occurs causes this register to be cleared. ERR_STAT Register
31 8 7 IllegalRegAddr 6 RcvdIllegalVector 5 SentIllegalVector 4 reserved
Offset 280h
3 RcvAcceptError 2 SendAcceptError 1 reserved 0
reserved
248
Chapter 7
32559
Rev. 3.08
July 2007
Bit 318 7 6 5 4 3 2 10
Name reserved IllegalRegAddr RcvdIllegalVector SentIllegalVector reserved RcvAcceptError SendAcceptError reserved
Function Illegal Register Address Received Illegal Vector Sent Illegal Vector Receive Accept Error Send Accept Error
R/W R/O W/R W/R W/R R/O W/R W/R R/O
Reset 0000_00h 0 0 0 0 0 0 00b
Field Descriptions Illegal Register Address (IllegalRegAddr)Bit 7. This bit indicates that an access to a nonexistent register location within this APIC was attempted. Received Illegal Vector (RcvdIllegalVector)Bit 6. This bit indicates that this APIC received a message with an illegal Vector (00h to 0Fh for fixed and lowest priority interrupts). Sent Illegal Vector (SentIllegalVector)Bit 5. This bit indicates that this APIC attempted to send a message with an illegal Vector (00h to 0Fh for fixed and lowest priority interrupts). Receive Accept Error (RcvAcceptError)Bit 3. This bit indicates that a message received by this APIC was not accepted by this or any other APIC. Send Accept Error (SendAcceptError)Bit 2. This bit indicates that a message sent by this APIC was not accepted by any APIC.
7.8.15
Interrupt Command Register Low

Offset 300h
20 19 18 17 16 15 14 13 12 11 10 RemoteRdStat DestShrthnd DlvryStat reserved 8 7 0
ICRLO Register
31
Level
reserved
DM
TM
Msg Type
Vector
Bit 3120 1918 1716 15 14 13
Name reserved DestShrthnd RemoteRdStat TM Level reserved
Function Destination Shorthand Remote Read Status Trigger Mode Level
R/W R/O R/W R/O R/W R/W R/O
Reset 000h 00b 00b 0 0 0
Chapter 7
249
32559
Rev. 3.08 July 2007
Bit 12 11 108 70
Name DlvryStat DM MsgType Vector
Function Delivery Status Destination Mode Message Type Vector
R/W R/O R/W R/W R/W
Reset 0 0 000b 00h
Field Descriptions Destination Shorthand (DestShrthnd)Bits 1918. This field provides a quick way to specify a destination for a message. If All Including Self or All Excluding Self are used, then DM is ignored and physical is automatically used. 00b =Destination Field 01b =Self 10b =All Including Self 11b =All Excluding Self (Note that this sends a message with a destination encoding of all 1s, so if lowest priority is used, the message could end up being reflected back to this APIC.) Remote Read Status (RemoteRdStat)Bits 1716. The encoding for this field is as follows: 00b =Read was Invalid 01b =Delivery Pending 10b =Delivery Done and access was valid 11b =Reserved
Trigger Mode (TM)Bit 15. This bit can be 0 for Edge or 1 for Level. Level (Level)Bit 14. This bit can be 0 for deasserted or 1 or asserted. Delivery Status (DlvryStat)Bit 12. This bit is set to indicate that the interrupt has not yet been accepted by the destination CPU(s). Destination Mode (DM)Bit 11. This bit can be 0 for Physical or 1 for Logical.Message Type (MsgType)Bits 108. The encoding for this field is as follows: 000b = Fixed 001b = Lowest Priority 010b = SMI 011b = Remote Read 100b = NMI 101b = INIT 110b = Startup 111b = External Interrupt
250
Chapter 7
32559
Rev. 3.08
July 2007
Note: Not all combinations of ICR fields are valid. Only those combinations listed in Table 58 are valid. Table 58. Valid Combinations of ICR Fields
Trigger Mode (TM) Edge Level Lowest Priority, SMI, NMI, INIT Startup
Note: X indicates a dont care.
Message Type (MsgType) Fixed
Level (Lvl) X Assert X Assert X
Destination Shorthand (DestShrthnd) X X Dest or All Excluding Self Dest or All Excluding Self Dest or All Excluding Self
Edge Level X
Vector (Vector)Bits 70. This field contains the vector that is sent for this interrupt source.
7.8.16
Interrupt Command Register High

Offset 310h
56 55 32 reserved
ICRHI Register
63 DestinationField
Bit 6356 5532
Name DestinationField reserved
Function Destination Field
R/W R/W R/O
Reset 00h 00_0000h
Field Descriptions Destination Field (DestinationField)Bits 6356. This field contains the destination encoding used when the destination shorthand is 00b.
7.8.17
Timer Local Vector Table Entry

Offset 320h
18 17 16 15 Mode reserved Mask 13 12 11 DlvryStat reserved 8 7 Vector 0
TIMER_LVT Register
31
reserved
Chapter 7
251
32559
Rev. 3.08 July 2007
Bit 3118 17 16 1513 12 118 70
Name reserved Mode Mask reserved DlvryStat reserved Vector
Function Mode Mask Delivery Status Vector
R/W R/O R/W R/W R/O R/O R/O R/W
Reset 0000h 0 1 000b 0 0h 00h
Field Descriptions Mode (Mode)Bit 17. This bit is 0 for One-shot and 1 for Periodic. Mask (Mask)Bit 16. If this bit is set, this LVT Entry does not generate interrupts. Delivery Status (DlvryStat)Bit 12. This bit is set to indicate that the interrupt has not yet been accepted by the CPU. Vector (Vector)Bits 70. This field contains the vector that is sent for this interrupt source.
7.8.18
Thermal Local Vector Table Entry

Offset 330h
17 16 15 reserved Mask 13 12 11 10 DlvryStat reserved 8 7 Vector 0
THERMAL_LVT Register
31
reserved
MsgType
Bit 3117 16 1513 12 11 108 70
Name reserved Mask reserved DlvryStat reserved MsgType Vector
Function Mask Delivery Status Message Type Vector
R/W R/O R/W R/O R/O R/O R/W R/W
Reset 0000h 1 000b 0 0 000b 00h
Field Descriptions Mask (Mask)Bit 16. If this bit is set, this LVT Entry does not generate interrupts. Delivery Status (DlvryStat)Bit 12. This bit is set to indicate that the interrupt has not yet been accepted by the CPU.
252
Chapter 7
32559
Rev. 3.08
July 2007
Message Type (Msg Type)Bits 108. Only Message Types 000b (Fixed), 010b (SMI) and 100b (NMI) are legal. Vector (Vector)Bits 70. This field contains the vector that is sent for this interrupt source.
7.8.19
Performance Counter Local Vector Table Entry

Offset 340h
17 16 15 reserved Mask 13 12 11 10 DlvryStat reserved 8 7 Vector 0
PERF_CNT_LVT Register
31
reserved
Msg Type
Bit 3117 16 1513 12 11 108 70
Name reserved Mask reserved DlvryStat reserved MsgType Vector
Reset 0000h 1 000b 0 0 000b 00h
Field Descriptions Mask (Mask)Bit 16. If this bit is set, this LVT Entry does not generate interrupts. Delivery Status (DlvryStat)Bit 12. This bit is set to indicate that the interrupt has not yet been accepted by the CPU. Message Type (MsgType)Bits 108. Message Types 000b (Fixed), 010b (SMI), and 100b (NMI) are legal. Vector (Vector)Bits 70. This field contains the vector that is sent for this interrupt source.
7.8.20
Local Interrupt 0 (Legacy INTR) Local Vector Table Entry Register

Offset 350h
17 16 15 14 13 12 11 10 DlvryStat reserved reserved RmtIRR PinPol Mask TM 8 7 Vector 0
LINT0_LVT Register
31
MsgType
Chapter 7
253
32559
Rev. 3.08 July 2007
Bit 3117 16 15 14 13 12 11 108 70
Name reserved Mask TM RmtIRR PinPol DlvryStat reserved MsgType Vector
Function Mask Trigger Mode Remote IRR Pin Polarity Delivery Status Message Type Vector
R/W R/O R/W R/W R/O R/W R/O R/O R/W R/W
Reset 0000h 1 0 0 0 0 0 000b 00h
Field Descriptions Mask (Mask)Bit 16. If this bit is set, this LVT Entry does not generate interrupts. Trigger Mode (TM)Bit 15. This bit is set for level-triggered interrupts and clear for edge-triggered interrupts. Remote IRR (RmtIRR)Bit 14. If trigger mode is level, remote IRR is set when the interrupt has begun service. Remote IRR is cleared when the EOI has occurred. Pin Polarity (PinPol)Bit 13. This bit is not used because LINT interrupts are delivered by HyperTransport messages instead of individual pins. Delivery Status (DlvryStat)Bit 12. This bit is set to indicate that the interrupt has not yet been accepted by the CPU. Message Type (MsgType)Bits 108. Only Message Types 000b (Fixed), 100b (NMI), and 111b (External Interrupt) are legal. Vector (Vector)Bits 70. This field contains the vector that is sent for this interrupt source.
7.8.21
Local Interrupt 1 (Legacy NMI) Local Vector Table Entry

Offset 360h
17 16 15 14 13 12 11 10 DlvryStat reserved reserved RmtIRR PinPol Mask TM 8 7 Vector 0
LINT1_LVT Register
31
MsgType
Bit 3117 16 15 14
Name reserved Mask TM RmtIRR
Function Mask Trigger Mode Remote IRR
R/W R/O R/W R/W R/O
Reset 0000h 1 0 0
254
Chapter 7
32559
Rev. 3.08
July 2007
Bit 13 12 11 108 70
Name PinPol DlvryStat reserved MsgType Vector
Function Pin Polarity Delivery Status Message Type Vector
R/W R/W R/O R/O R/W R/W
Reset 0 0 0 000b 00h
Field Descriptions These fields are the same as those for Local Interrupt 0.
7.8.22
Error Local Vector Table Entry

Offset 370h
17 16 15 reserved Mask 13 12 11 DlvryStat reserved 8 7 Vector 0
ERROR_LVT Register
31
reserved
Bit 3117 16 1513 12 118 70
Name reserved Mask reserved DlvryStat reserved Vector
Function Mask Delivery Status Vector
R/W R/O R/W R/O R/O R/O R/W
Reset 0000h 1 000b 0 0h 00h
Field Descriptions Mask (Mask)Bit 16. If this bit is set, this LVT Entry does not generate interrupts. Delivery Status (DlvryStat)Bit 12. This bit is set to indicate that the interrupt has not yet been accepted by the CPU. Vector (Vector)Bits 70. This field contains the vector that is sent for this interrupt source.
7.8.23
Timer Initial Count Register

Offset 380h
0 Count
INIT_CNT Register
31
Chapter 7
255
32559
Rev. 3.08 July 2007
Bit 310
Name Count
Function Initial Count Value
R/W R/W
Reset 0000_0000h
Field Descriptions Count (Count)Bits 310. This field contains the value copied into the current count register when the timer is loaded or reloaded.
7.8.24
Timer Current Count Register

Offset 390h
0 Count
CURR_CNT Register
31
Bit 310
Name Count
Function Current Count Value
R/W R/O
Reset 0000_0000h
Field Descriptions Count (Count)Bits 310. This field contains the current value of the counter.
7.8.25
Timer Divide Configuration Register

Offset 3E0h
4 reserved 3 Div[3] 2 reserved 1 Div[1-0] 0
TIMER_DVD_CFG Register
31
Bit 314 3 2 10
Name reserved Div[3] reserved Div[10]
Function Div[3] Div[10]
R/W R/O R/W R/O R/W
Reset 0000_00h 0 0 00b
Field Descriptions Div[3] and Div[10]Bits 3 and 10. The Div bits are encoded as follows:
Div[3] Div[10] Resulting Timer Divide
256
Chapter 7
32559
Rev. 3.08
July 2007
0 0 0 0 1 1 1 1
00 01 10 11 00 01 10 11
2 4 8 16 32 64 128 1
7.8.26
Extended APIC Feature Register
See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. EXT_APIC_FEAT Register
31 reserved 24 23 NumExtLvt 16 15 reserved
Offset 400h
2 1 0
Bit 3124 2316 150
Name reserved NumExtLvt reserved
Function Number of Extended LVTs Supported
R/W R/O R/O R/O
Reset 0 1h 0
Field Descriptions Number of Extended LVTs Supported (NumExtLvt)Bits 2316. This field indicates the number of extended LVT entries supported.
7.8.27
Threshold Count Interrupt 0 Local Vector Table Entry
This register enables and controls interrupts caused by the DRAM ECC threshold counter overflow in MC4_MISC (see MC4_MISCDRAM Errors Threshold Register on page 229). See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. THRCNT_INT0_LVT Register
31 reserved 17 16 15 Mask 13 12 11 10 DlvryStat reserved 8 7 Vector
Offset 500h
0
reserved
MsgType
Chapter 7
257
32559
Rev. 3.08 July 2007
Bit 3117 16 1513 12 11 108 70
Name reserved Mask reserved DlvryStat reserved Msgtype Vector
Reset 0000h 1 000b 0 0h 0h 00h
Field Descriptions Mask (Mask)Bit 16. If this bit is set, this LVT Entry does not generate interrupts. Delivery Status (DlvryStat)Bit 12. This bit is set to indicate that the interrupt has not yet been accepted by the CPU. Message Type (Msg Type)Bits 108. Only Message Types 000b (Fixed), 010b (SMI), 100b (NMI) and 111b (ExtInt) are legal. Vector (Vector)Bits 70. This field contains the vector that is sent for this interrupt source.
258
Chapter 7
32559
Rev. 3.08
July 2007
System Management Mode (SMM)
System management mode (SMM) is an operating mode entered when system management interrupt SMI is asserted. The SMI is handled by a dedicated interrupt handler pointed to by processor model specific registers. SMM is used for system control activities such as power management. These activities are transparent to conventional operating systems (such as MS-DOS and Windows operating system). SMM is used by the BIOS and specialized low level device drivers. The code and data for SMM are stored in the SMM memory area, which may be isolated from the main memory accesses using special processor functions.
This chapter describes the SMM state save area, entry into and exit from SMM, exceptions and interrupts in SMM, and memory allocation and addressing in SMM.
8.1
SMM Overview
The processor enters SMM after the system logic asserts the SMI interrupt and the processor recognizes SMI active. The processor may be programmed to send a special bus cycle to the system, indicating that it is entering SMM mode. The processor saves its state into the SMM memory state save area and jumps to the SMM service routine. The processor returns from SMM by executing the RSM instruction in the SMM service routine. The processor restores its state from the SMM state save area and resumes execution of the instruction following the point where it entered SMM. The processor may be programmed to send a special bus cycle to the system, indicating that it is exiting SMM mode. The system logic asserting the SMI interrupt may be any I/O device connected to a noncoherent HyperTransport link. If more than one device exists in the system then the BIOS must take special care to ensure that all processors have entered SMM prior to accessing shared I/O resources and that all processors SMI interrupt status bits are synchronized. See Section 8.13 on page 274 for more information on multiple system sources of SMIs. The processor core also enters SMM after the Northbridge block of the processor asserts the core specific SMI interrupt and the processor core recognizes SMI active. This is a processor core-local SMI and is not recognized by other processor cores in the system. The BIOS must take special care to ensure that other processor cores have entered SMM prior to accessing shared I/O resources. See Section 8.13 on page 274 for more information on processor core local SMIs.
8.2

Operating Mode and Default Register Values
The software environment after entering SMM mode has the following characteristics: Addressing and operation in Real mode. 4-Gbyte segment limits. System Management Mode (SMM) 259
Chapter 8
32559
Rev. 3.08 July 2007
Default 16-bit operand, address, and stack sizes (instruction prefixes can override these defaults). Control transfers that do not override the default operand size truncate the EIP to 16 bits. Far jumps or calls cannot transfer control to a segment with a base address requiring more than 20 bits, like in Real mode segment-base addressing, unless a change is made into protected mode A20M# is disabled. A20M# assertion or deassertion have no effect on SMM mode. Interrupt vectors use the Real mode interrupt vector table. The IF flag in EFLAGS is cleared (INTR is not recognized). The TF flag in EFLAGS is cleared. The NMI and INIT interrupts are masked. Debug register DR7 is cleared (debug traps are disabled).
Figure 6 on page 261 shows the default map of the SMM memory area. It consists of a 64-Kbyte area, between 0003_0000h and 0003_FFFFh. The top 32 Kbytes (0003_80000003_FFFFh) must be populated with RAM. The default code-segment (CS) base address for the area (called the SMM_BASE address) is 0003_0000h. The top 512 bytes (0003_FE000003_FFFFh) are the SMM state save area. The default entry point for the SMM service routine is 0003_8000h.
260
Chapter 8
32559
Rev. 3.08
July 2007
Fill Down
SMM state save Area
0003_FFFFh
0003_FE00h
32-Kbyte Minimum
SMM Service Routine Service Routine Entry Point 0003_8000h
SMM_BASE Address (CS)
0003_0000h
Figure 6.
Default SMM Memory Map
8.3
Table 59.
SMM State Save Definition

SMM Save State (Offset FE00FFFFh)
Note:
Offset (Hex) from SMM_BASE FE00h FE02h FE04h FE08h FE10h FE12h FE14h FE18h
Contents ES Selector Attributes Limit Base CS Selector Attributes Limit Base
Size Word Word Doubleword Quadword Word Word Doubleword Quadword
Allowable Access Read-Only
Read-Only
Note: The offset for the SMM-revision identifier is compatible with previous implementations.
Chapter 8
261
32559
Rev. 3.08 July 2007
Table 59.
SMM Save State (Offset FE00FFFFh) (Continued)

Note:
Offset (Hex) from SMM_BASE FE20h FE22h FE24h FE28h FE30h FE32h FE34h FE38h FE40h FE42h FE44h FE48h FE50h FE52h FE54h FE58h FE60hFE61h FE62h FE64h FE66hFE67h FE68h FE70h FE72h FE74h FE78h FE80hFEB1h FE82h FE84h FEB6hFEB7h FE88h FE90h FE92h FE94h FE98h
Contents SS Selector Attributes Limit Base DS Selector Attributes Limit Base FS Selector Attributes Limit Base GS Selector Attributes Limit Base GDTR reserved reserved Limit reserved Base LDTR Selector Attributes Limit Base IDTR reserved reserved Limit reserved Base TR Selector Attributes Limit Base
Size Word Word Doubleword Quadword Word Word Doubleword Quadword Word Word Doubleword Quadword Word Word Doubleword Quadword 2 Bytes Word Word 2 Bytes Quadword Word Word Doubleword Quadword 2 Bytes Word Word 2 Bytes Quadword Word Word Doubleword Quadword
Allowable Access Read-Only
Read-Only
Read-Only
Read-Only
Read-Only
Read-Only
Read-Only
Read-Only
262
Chapter 8
32559
Rev. 3.08
July 2007
Table 59.

Note:
Offset (Hex) from SMM_BASE FEA0hFEBFh FEC0hFEC3h FEC4hFEC7h FEC8h FEC9h FECAh FECABhFECFh FED0h FED8hFEFBh FEFCh FF00h FF04hFF47h FF48h FF50h FF58h FF60h FF68h FF70h FF78h FF80h FF88h FF90h FF98h FFA0h FFA8h FFB0h FFB8h
Contents reserved SMM I/O Trap reserved I/O Instruction Restart Auto-Halt Restart NMI Mask reserved EFER reserved SMM-Revision Identifier SMM_BASE reserved CR4 CR3 CR0 DR7 DR6 RFLAGS RIP R15 R14 R13 R12 R11 R10 R9 R8
1
Size 32 Bytes Doubleword 4 Bytes Byte Byte Byte 6 Bytes Quadword 36 Bytes Doubleword Doubleword 68 Bytes Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword
Allowable Access Read-Only Read/Write
Read-Only Read-Only Read/Write Read-Only
Read-Only Read/Write Read/Write
Chapter 8
263
32559
Rev. 3.08 July 2007
Table 59.

Note:
Offset (Hex) from SMM_BASE FFC0h FFC8h FFD0h FFD8h FFE0h FFE8h FFF0h FFF8h
Contents RDI RSI RBP RSP RBX RDX RCX RAX
Size Quadword Quadword Quadword Quadword Quadword Quadword Quadword Quadword
Allowable Access Read/Write
After recognizing the SMI assertion, the processor saves almost the entire integer processor state to memory. Exceptions are: the floating point state, the model specific registers, and CR2. Any processor state not saved in the save state must be saved and restored by the SMM handler. With the AMD NPT Family 0Fh Processor extension of register size to 64-bits, the SMM save state has changed considerably compared to the legacy 32-bit version. One exception is the SMM Revision Identifier, which is located in the same position as before, for proper SMM identification. Note: The save state is always be 64-bit, regardless of the operating mode (32-bit or 64-bit).
8.4
SMM Initial State
After storing the save state, the processor starts executing the handler at address SMM_BASE + 08000h. It starts with the entry state listed in Table 60 on page 265.
264
Chapter 8
32559
Rev. 3.08
July 2007
Table 60.
Register CS DS ES FS GS SS
SMM Entry State

Initial Contents Selector SMM_BASE[19:4] 0000h 0000h 0000h 0000h 0000h Unmodified 0000_0002h 0000_8000h Bits 0, 2, 3, and 31 cleared (PE, EM, TS, and PG); remainder is unmodified 0000_0000h Unmodified Unmodified Unmodified Unmodified 0000_0400h Undefined Base SMM_BASE[31:0] 0000_0000h 0000_0000h 0000_0000h 0000_0000h 0000_0000h Limit
4 Gbytes 4 Gbytes 4 Gbytes 4 Gbytes 4 Gbytes 4 Gbytes
General-Purpose Registers EFLAGS EIP CR0 CR4 GDTR LDTR IDTR TR DR7 DR6
8.5
SMM-Revision Identifier
Offset FEFCh
18 17 16 15 reserved IOTRAP BRL REV 0
The SMM-Revision Identifier specifies the version of SMM and the processor extensions. SMM-Revision Identifier
31
Bit 3118 17 16 150
Name reserved BRL IOTRAP REV
Function SBZ SMM Base Relocation Supported SMM I/O Trap Supported SMM Revision Level
R/W R R R R
Reset 1 1 0064h
Chapter 8
265
32559
Rev. 3.08 July 2007
8.6
SMM Base Address
SMM_BASE register holds the base of the system management memory region, and its default value is 30000h. The value of this register is stored in the save state on entry into SMM and it is restored on returning from SMM. The first SMI handler instruction is fetched at SMM_BASE + 8000h. The 16 bit code segment selector is formed on entering SMM from SMM_BASE[194]. SMM_BASE[31 20] and SMM_BASE[30] have to be 0; if not, the CS base will not be equal to the (CS selector << 4), and attempts to load the CS_BASE into other descriptors will not work properly, since the selector cannot properly represent the base. If there is a far branch in the SMM handler, it will only be able to address the lower 1M of memory and will not be able to restore the SMM base to a value above 1M, unless it first switches to protected mode. The SMM base address can be changed in two ways: The SMM base address, at offset FF00h in the SMM state save area, can be changed by the SMM service routine. The RSM instruction updates the SMM_BASE with the new value. The SMM_BASE is mapped to MSR C001_0111h and WRMSR instruction can be used to update it. Offset FF00h MSR C001_0111h
32 reserved 31 SMM_BASE 0
SMM_BASE Address
63
Bit 6332 310
Name reserved SMM_BASE
Function RAZ System management mode base
R/W R/W
8.7
Auto Halt Restart
During entry into SMM, the auto halt restart byte at offset FEC9h in the SMM state save area indicates whether SMM was entered from the Halt state. Before returning from SMM, the halt restart byte bit can be written by the SMM service routine to specify whether the return from SMM should take the processor back to the Halt state or to the instruction-execution state specified by the SMM state save area (the instruction after the HLT).
266
Chapter 8
32559
Rev. 3.08
July 2007
If the return from SMM takes the processor back to the Halt state, the HLT instruction is not refetched and reexecuted, but the Halt special bus cycle is sent to the system on the bus and the processor enters the HLT state. Auto Halt Restart
7 reserved 0 HLT
Offset FEC9h
Bit 71 0
Name reserved HLT
Function SBZ HLT Restart Byte
R/W R/W R/W
Field Description HLT Restart Byte (HLT)Bit 0. On entry: 0 = Entered from normal x86 instruction boundary 1 = Entered from HLT State After entry and before returning: 0 = Return to SMM State 1 = Return to HLT State
8.8
SMM I/O Trap and I/O Restart
If the assertion of SMI is recognized on the boundary of an I/O instruction, the I/O trap doubleword at offset FEC0h in the SMM state save area contains information about the executed I/O instruction. If the Valid bit is equal to 0, not all information is valid. This is used in the I/O restart implementation. When an I/O access is done to a device that may be unavailable, the system can assert SMI and trap the I/O instruction. The system can then determine which device is not responding by using the I/O port information in the I/O Trap doubleword. It can then figure out why the device is not responding, fix the device, and re-execute the I/O instruction. It can reconstruct and then re-execute the I/O instruction in the SMM handler by using the SMM I/O trap information. More likely, it can use the SMM I/O restart mechanism to cause the processor to restart the I/O instruction immediately after the RSM.
Chapter 8
267
32559
Rev. 3.08 July 2007
8.8.1
SMM I/O Trap

Offset FEC0h
16 15 PORT BRP 12 11 10 TF reserved 7 6 SZ32 5 SZ16 4 SZ8 3 REP 2 STR 1 V 0 R/W
SMM I/O Trap

31
Bit 3116 1512 11 107 6 5 4 3 2 1 0
Name PORT BRP TF reserved SZ32 SZ16 SZ8 REP STR V R/W
Function Trapped I/O Port I/O Breakpoint Matches EFLAGS TF Value RAZ Port Access was 32-bit Port Access was 16-bit Port Access was 8-bit Repeated Port Access String Based Port Access I/O Trap Word Valid Bit Port Access Type 0 = Write (OUT instruction) 1 = Read (IN instruction)
R/W R R R R R R R R R R R
8.8.2
SMM I/O Restart Byte
The processor initializes the I/O trap restart slot to 00h on entry into SMM. If SMM is entered as a result of a trapped I/O instruction, the processor indicates the validity of the I/O instruction by setting or clearing bit 1 of the I/O trap doubleword at offset FEC0h in the SMM state save area. The SMM service routine should test bit 1 of the I/O trap doubleword to determine if a valid I/O instruction was being executed when entering SMM and before writing the I/O trap restart slot. If the I/O instruction is valid, the SMM service routine can safely rewrite the I/O trap restart slot with the value FFh, causing the processor to re-execute the trapped I/O instruction when the RSM instruction is executed. If the I/O instruction is invalid, writing the I/O trap restart slot has undefined results. If a second SMI is asserted and a valid I/O instruction is trapped by the first SMM handler, the CPU services the second SMI prior to re-executing the trapped I/O instruction. The second entry into SMM does not have bit 1 of the I/O trap doubleword set, and the second SMM service routine must not rewrite the I/O trap restart slot. During a simultaneous SMI I/O instruction trap and debug breakpoint trap, the processor first responds to the SMI and postpones recognizing the debug exception until after the RSM instruction returns from SMM. If the debug registers DR3DR0 are used while in SMM, they must be saved and restored by the SMM handler. The processor automatically saves and restores DR7 and DR6. If the I/O trap restart slot in the SMM state save area contains the value FFh when the RSM instruction is executed, the debug trap does not occur until after the I/O instruction is re-executed. 268 System Management Mode (SMM) Chapter 8
32559
Rev. 3.08
July 2007
SMM I/O Restart Byte

7 RST 0
Offset FEC8h
Bit 70
Name RST
Function I/O Restart Byte 00h = Do not restart FFh = Restart I/O instruction
R/W R/W
8.9
Exceptions and Interrupts in SMM
When SMM is entered, the processor masks INTR, NMI, SMI, INIT, and A20M interrupts. The processor clears the IF flag to disable INTR interrupts. To enable INTR interrupts within SMM, the SMM handler must set the IF flag to 1. A20M is disabled so that address bit 20 is always generated externally when in SMM. Generating an INTR interrupt is a method for unmasking NMI interrupts in SMM. The processor recognizes the assertion of NMI within SMM immediately after the completion of an IRET instruction. The NMI can be enabled by using an INTR interrupt. Once NMI is recognized within SMM, NMI recognition remains enabled until SMM is exited, at which point NMI masking is restored to the state it was in before entering SMM. Because the IF flag is cleared when entering SMM, the HLT instruction should not be executed in SMM without first setting the IF bit to 1. Setting this bit to 1 enables the processor to exit the Halt state by means of an INTR interrupt. The processor still responds to the DBREQ and STPCLK interrupts as well as to all exceptions that might be caused by the SMM handler.
8.10
NMI Mask
During entry into SMM, the NMI Mask byte at offset FECAh in the SMM state save area indicates whether NMI was masked when SMM was entered. NMI Mask
7 reserved 0 NMIM
Offset FECAh
Chapter 8
269
32559
Rev. 3.08 July 2007
Bit 71 0
Name reserved NMI
Function SBZ NMI Mask
R/W R/W R/W
Field Description NMI Mask (NMIM)Bit 0. On entry: 0 = NMI not masked 1 = NMI Masked
8.11
Protected SMM and ASeg/TSeg
System management memory, SMRAM, is defined by two ranges. The ASeg range is located at a fixed range from A0000hBFFFFh. The TSeg range is located at a variable base. The size of the TSeg range is controlled by a variable mask. SMRAM provides a safe location for system management code and data that is not readily accessible by applications. The SMM interrupt handler can be located in one of these two ranges if protection is needed, or it can be located outside these ranges in the rest of memory.
8.11.1
SMM_MASK Register
This register holds the mask of the TSeg range as well as configuration information for both the fixed ASeg range and the variable TSeg range. SMM_MASK Register
63 reserved 40 39 MASK (2215)
MSR C001_0113h
32
31
17 16 15 14 TMTypeDram reserved
12 11 10 AMTypeDram reserved
7 reserved
5 TMTypeIoWc
4 AMTypteIcWc
3 TClose
2 AClose
1 TValid
0 AValid
MASK (140)
Bit 6340 3917 1615 1412 11
Name reserved MASK reserved TMTypeDram reserved
Function RAZ SMRAM TSeg Range Mask RAZ SMRAM TSeg Range Memory Type RAZ
R/W R/W R/W
270
Chapter 8
32559
Rev. 3.08
July 2007
Bit 108 76 5 4 3 2 1 0
Name AMTypeDram reserved TMTypeIoWc AMTypteIcWc TClose AClose TValid AValid
Function SMRAM ASeg Range Memory Type RAZ Non-SMRAM TSeg Range Memory Type Non-SMRAM ASeg Range Memory Type Send TSeg Range Data Accesses to Non-SMRAM Send ASeg Range Data Accesses to Non-SMRAM Enable TSeg SMRAM Range Enable ASeg SMRAM Range
R/W R/W R/W R/W R/W R/W R/W R/W
SMRAM TSeg Range Mask (MASK)Bits 3917. Mask of the SMRAM TSeg Range. SMRAM TSeg Range Memory Type (TMTypeDram)Bits 1412. Memory Type for the SMRAM TSeg Range. SMRAM ASeg Range Memory Type (AMTypeDram)Bits 108. Memory Type for the SMRAM ASeg Range. Non-SMRAM TSeg Range Memory Type (TMTypeIoWc)Bit 5. Memory Type for the nonSMRAM TSeg Range. Non-SMRAM ASeg Range Memory Type (AMTypteIcWc)Bit 4. Memory Type for the nonSMRAM ASeg Range. Send TSeg Range Data Accesses to Non-SMRAM (TClose)Bit 3. Send Data Accesses to the TSeg Range to Non-SMRAM. Send ASeg Range Data Accesses to Non-SMRAM (AClose)Bit 2. Send Data Accesses to the ASeg Range to Non-SMRAM. Enable TSeg SMRAM Range (TValid)Bit 1. Enables the TSeg SMRAM Range. Enable ASeg SMRAM Range (AValid)Bit 0. Enables the ASeg SMRAM Range.
8.11.2
SMM_ADDR Register
This register holds the base of the variable TSeg range. SMM_ADDR Register
63 reserved 31 ADDR (140) 17 16 reserved 40 39 ADDR (2215) 0
MSR C001_0112h
32
Chapter 8
271
32559
Rev. 3.08 July 2007
Bit 6340 3917 160
Name reserved ADDR reserved
Function RAZ Base of the SMRAM TSeg Range RAZ
R/W R/W
8.11.3
SMM ASeg
The SMM ASeg address range is A0000BFFFFh. The ASeg is enabled by the AValid bit in the SMM_MASK MSR. When the ASeg is enabled, the MTRR memory maps are not used for the addresses in A0000 BFFFFh memory range. Instead, settings in the SMM_MASK register and whether the processor is in SMM or not determine how those addresses are handled. If not in SMM, the addresses are mapped to I/O space and do not access DRAM. The memory type used is either UC or WC, based on the AMTypeIoWc bit in the SMM_MASK register. When inside SMM, the accesses to the ASeg memory range are directed to DRAM. The memory type is determined using the normal memory type encoding and is specified in AMTypeDram. Table 61 lists these types. Table 61. SMM ASeg-Enabled Memory Types
Out of SMM Address Range 09FFFh A0000BFFFFh C0000hMax In SMM Normal MTRR DRAM, AMTypeDram Normal MTRR AMTypeIoWc=0 Normal MTRR I/O, UC Normal MTRR AMTypeIoWc=1 Normal MTRR I/O, WC Normal MTRR
The SMM save state and code can be cached. After leaving SMM, the same addresses bypasses the caches and DRAM and accesses the I/O memory subsystem. This protects the SMM memory from corruption by programs outside of SMM.
8.11.4
SMM TSeg
The TSeg is similar to the ASeg except it provides more programmability and therefore allows more space to be allocated in a different area of memory for the SMM handler. It is enabled by the TValid bit in the SMM_MASK register. It uses the SMM_ADDR register to specify its base. It can be used with ASeg or by itself. When it is used with ASeg, the two spaces should not overlap. If they do overlap, the memory types should be consistent in both the ASeg map and the TSeg map for the overlapping addresses. Otherwise, undefined behavior occurs. The SMM TSeg begins at the address specified in SMM_ADDR concatenated with 0s in the lower 17 bits, forcing the TSeg to begin on a 128-Kbyte boundary. The ending of TSeg is specified by the SMM_Mask[39:17]. The SMM_Mask and SMM_ADDR function as the variable-range MTRRs. 272 System Management Mode (SMM) Chapter 8
32559
Rev. 3.08
July 2007
Each address generated by the processor is ANDed with the SMM_MASK and compared to the SMM_ADDR ANDed with the mask. If the values are equal, the address is within the TSeg range. For example, suppose you wish to create a TSeg starting at the 1-Mbyte address boundary and extending 256 Kbytes. The SMM_ADDR register would be set to 0010_0000h and the SMM_MASK to FFFC_0002h. This makes the TSeg range from 0010_00000013_FFFFh. The TSeg memory type table is similar to the ASeg memory type table. When not in SMM and TSeg is enabled, the addresses within the TSeg range are directed to I/O space with either a UC memory type or a WC memory type, based on the TMTypeIoWc bit in the SMM_MASK register. When within SMM, the accesses are directed to DRAM with a memory type specified by TMTypeDram. Table 62. SMM TSeg-Enabled Memory Types
Out of SMM Address Range <TSeg Range TSeg Range > TSeg Range In SMM Normal MTRR DRAM, TMTypeDram Normal MTRR TMTypeIoWc=0 Normal MTRR I/O, UC Normal MTRR TMTypeIoWc=1 Normal MTRR I/O, WC Normal MTRR
8.11.5
Closing SMM
Sometimes within SMM code with ASeg or TSeg enabled, there is a requirement to access the I/O space at the same address as the current SMM segment. That is typically only accessible outside of SMM. To accomplish this function, the Aclose and Tclose bits from SMM_MASK register are used. When the Aclose bit is set, data cache accesses to the ASeg that would normally go to DRAM are redirected to I/O, with the memory type specified by AMTypeIoWc. The same function applies to the TSeg. Instruction cache accesses and Page Directory/Table accesses still access the SMM code in DRAM. When the SMM handler is done accessing the I/O space, it must clear the appropriate close bit. Failure to do so and then issuing an RSM will probably cause the processor to enter shutdown, as the save state is read from I/O space.
8.11.6
Locking SMM
The SMM registers can be locked by setting the SMMLOCK (HWCR, bit 0). Once set, the SMM_BASE, the SMM_ADDR, all but the two close bits of SMM_MASK and the RSMSPCYCDIS, SMISPCYCDIS, and SMMLOCK bits of HWCR are locked and cannot be changed. The only way to unlock the SMM registers is to assert reset. This provides security to the SMM mechanism. The BIOS can lock the SMM environment after setting it up so that it can not be tampered with.
Chapter 8
273
32559
Rev. 3.08 July 2007
8.12
SMM Special Cycles
Special cycles can be initiated on entry and exit from SMM to acknowledge to the system that these transitions are occurring. This allows secure spaces external to the processor to be secured within the SMM environment as well. It also helps with synchronizing the SMI interrupt across multiple processor systems. These special cycles can be disabled by setting the SMISPCYCDIS to 1 to disable the entry special cycle and setting the RSMSPCYCDIS bit to 1 to disable the exit special cycle.
8.13
MP-SMM Spring Boarding
When processor core-local SMI sources are enabled, or when more than one device in the system is enabled to signal SMI, BIOS must take special care to ensure that all cores have entered SMM prior to accessing shared I/O resources and all processors SMI interrupt status bits are synchronized. The act of synchronizing cores into SMM is called spring boarding. MP-SMM spring boarding applies to dual-core uniprocessor systems (CMP-SMM) as well as symmetric multiprocessing systems (SMPSMM). An ACPI-compliant I/O hub is required for MP-SMM spring boarding. Depending on the I/O hub design, BIOS may have to set additional end-of-SMI bits to trigger an SMI from within SMM. The software requirements for MP-SMM spring boarding are as follows. A binary semaphore located in SMRAM, accessible by all processors. For the purpose of this discussion, the semaphore is called CheckSpringBoard. CheckSpringBoard is initialized to zero. Two semaphores located in SMRAM, accessible by all processors. For the purpose of this discussion, the semaphores are called NotInSMM and WaitInSMM. NotInSMM and WaitInSMM are initialized to a value equal to the number of processor cores in the system (NumCPUs).
The following BIOS algorithm describes spring boarding and is optimized to reduce unnecessary SMI activity. This algorithm must be made part of the SMM instruction sequence for each processor core in the system. 1. Attempt to obtain ownership of the CheckSpringBoard semaphore with a read-modify-write instruction. If ownership was obtained then do the following, else proceed to step 2: Check all enabled SMI status bits in the I/O hub. Let Status=enable1&status1 | enable2&status2 | enable3&status3enable n&status n. If Status=0 then perform the following sub-actions. Trigger an SMI broadcast assertion from the I/O hub by writing to the software SMI command port. Resume from SMM with the RSM instruction.
//Example: InLineASM{ BTS CheckSpringBoard,0
;Try to obtain ownership of semaphore
274
Chapter 8
32559
Rev. 3.08
July 2007
JC Step_2: CALL CheckIOHUB_SMIEVT JNZ Step_2: CALL Do_SpringBoard Step_2: }
;proc returns ZF=1 for no events ;Trigger SMI and then RSM
2. Decrement the NotInSMM variable. Wait for NotInSMM=0. See Note 1. 3. Execute the core-local event SMI handler. Using a third semaphore (not described here), synchronize processor core execution at the end of the task. After all processor cores have executed, proceed to step 4. The following is a brief description of the task for each processor core: Check all enabled processor-core-local SMI status bits in the cores private or MSR address space. Handle the event if possible, or pass information necessary to handle the event to a mailbox for the boot strap processor to handle. An exclusive mailbox must exist for each processor core for each core local event. On-line spare events should be handled in this task by the individual core for optimal performance. Assign one core of a dual core processor to handle On-line spare. These events may be optionally handled by the BSP just as other global events. Wait for all processor cores to complete this task at least once. 4. If the current processor core executing instructions is not the boot strap processor then jump to step 5. If the core executing instructions is the boot strap processor then jump to the modified main SMI handler task, described below. Check all enabled SMI status bits in the I/O hub. Check mailboxes for event status. For each event, handle the event and clear the corresponding status bit. Repeat until all enabled SMI status bits are clear and no mailbox events remain. Set NotInSMM=NumCPUs. (Jump to step 5.) 5. Decrement the WaitInSMM variable. Wait for WaitInSMM=0. See Note 2. 6. Increment the WaitInSMM variable. Wait for WaitInSMM=NumCPUs. 7. If the current processor core executing instructions is the boot strap processor then reset CheckSpringBoard to zero. 8. Resume from SMM with the RSM instruction. Note: 1. To support a secure startup by the secure loader the BIOS must provide a timeout escape from the otherwise endless loop. The timeout value should be large enough to account for the latency of all processor cores entering SMM. The maximum SMM entrance latency is defined by the platforms I/O sub-system, not the processor. AMD recommends a value of twice the watchdog timer count. See MCA NB Configuration Register for more information on the watchdog time-out value. If a time-out occurs in the wait loop, the BIOS (the last core to decrement NotInSMM) Chapter 8 System Management Mode (SMM) 275
32559
Rev. 3.08 July 2007
should record the number of cores that have not entered SMM and all cores must fall out of the loop. 2. If a time-out occurs in the wait loop in step 2, the BIOS must not wait for WaitInSMM=0. Instead it must wait for WaitInSMM=the number of cores recorded in step 2.
276
Chapter 8
32559
Rev. 3.08
July 2007
HyperTransport Technology Configuration and Enumeration
An AMD NPT Family 0Fh Processor is a node connected to other nodes with coherent HyperTransport links. Function 0 HyperTransport technology configuration register values for each node must be initialized. A node is identified in the Node ID register (Function 0, Offset 60h) with a number from 0 to 7. Routing tables are used by the node to decide whether to accept a packet and/or forward it through any or all of its HyperTransport links. Three examples of coherent HyperTransport technology initialization sequence are shown in this chapter: 1-node initialization, 2-node initialization and generic initialization that can enumerate any number of nodes. This chapter also provides information on how to determine if a given routing table is valid.
9.1
Initial Configuration Steps
For coherent HyperTransport technology initialization, the following steps are required before enumeration: 1. Coherent HyperTransport Technology initialization is only performed by Node 0 (BSP). All the other nodes (APs) should bypass the HyperTransport initialization process. 2. Clearing RoutTblDis (Function 0, Offset 6Ch) enables the routing table for Node 0, and allows access to the memory controller on Node 0. 3. Node 0 is by definition connected to the HyperTransport I/O Hub, which means that Node 0 owns the compatibility chain. The link number that connects to HyperTransport I/O Hub should be written to SbLink (Function 0, Offset 64h). 4. Count coherent HyperTransport links on Node 0. To perform the step 4 above, it is necessary to detect whether each HyperTransport link on a node is connected, and if the connected link type is coherent or noncoherent. The AMD Opteron processor supports up to three links, and the AMD Athlon 64 processor supports one link. For each HyperTransport link, the following steps are performed to detect the link connection status and type: 1. Check whether LinkConPend (Function 0, Offset 98h, D8h, B8h) is set. 2. If the LinkConPend is clear, check whether LinkCon (Function 0, Offset 98h, D8h, B8h) is set. If the Link Connected bit is set, the testing port is connected to other node. 3. Check whether InitComplete (Function 0, Offset 98h, D8h, B8h) is set. 4. If InitComplete is set, check NC (Function 0, Offset 98h, D8h, B8h) bit. The bit value 0 indicates a coherent HyperTransport link, while value 1 indicates a noncoherent link. Chapter 9 HyperTransport Technology Configuration and Enumeration 277
32559
Rev. 3.08 July 2007
5. Record the number of coherent and noncoherent ports and their respective port numbers.
9.2
One-Node Coherent HyperTransport Technology Initialization
The number of nodes that exist in the system is determined on page 277. The following configuration steps should be executed in single (one-node) processor systems: 1. Set LimitCldtCfg (Function 0, Offset 68h) to enable limiting the extent of coherent HyperTransport configuration space based on the number of nodes in the system. 2. Disable read/write/fill probes in Function 0, Offset 68h for single-core one-node systems, since single processor systems do not need probes. For dual-core one-node systems, enable read/write/fill probes in Function 0, Offset 68h.
9.3
Two-Node Coherent HyperTransport Technology Initialization
For two-node systems, Node ID and Routing Tables need to be written for both nodes, as follows. 1. Initialize Node 0 Routing Table rows 0 and 1. Based on the coherent HyperTransport link number on Node 0. 2. Verify if Node 0-to-Node 1 link is established by reading Vendor ID/Device ID of Node 1. 3. Initialize Node 1 Routing Table row 0 and 1 using the coherent HyperTransport link number on Node 1. 4. Initialize the node ID of Node 1 by programming NodeId (Function 0, Offset 60h). 5. Write CPU Count and Node Count to CpuCnt and NodeCnt (Function 0, Offset 60h). 6. Set LimitCldtCfg (Function 0, Offset 68h) for Node 0 and Node 1. 7. Clear RouteTblDis (Function 0, Offset 6Ch) for Node 1.
9.4
Generic HyperTransport Technology Configuration
For multi-node systems, Node ID and Routing Tables need to be written for all nodes, as follows. This algorithm assumes the BIOS knows the topology of the nodes and how each link between nodes is routed on the system board. 1. Initialize Node 0 Routing Table rows 0 through N based on the coherent HyperTransport link numbers on Node 0.
278
Chapter 9
32559
Rev. 3.08
July 2007
2. For each node n in the system (starting with the nodes connected to Node 0 and working outward to the nodes furthest from Node 0) repeat steps a though e. a. Verify if Node 0-to-Node n link is established by reading Vendor ID/Device ID of Node n. If the link node n is not established, disable the Routing Table on each node and go back to step 1 and build the Routing Tables for each node based on the new configuration. b. Poll the ReqDis bit (Function 0, Offset 6Ch) on Node n until set. c. Initialize Node n Routing Table row 0 through N. Using the coherent HyperTransport link numbers on Node n. d. Initialize the node ID of Node n by programming NodeId (Function 0, Offset 60h). e. Clear RouteTblDis (Function 0, Offset 6Ch) for Node 1. 3. Write CPU Count and Node Count to CpuCnt and NodeCnt (Function 0, Offset 60h) of each Node. 4. Set LimitCldtCfg (Function 0, Offset 68h) bit for each Node.
9.5
Rules for Valid Routing Tables
This section provides rules for determining if a given routing table is valid. A routing table is valid if and only if the routing table is deadlock free and probes are delivered only once. A routing table is deadlock-free if it contains no Open Paths and no two-hop cycles. An Open Path is a routing path between nodes that passes through one or more nodes that contains a subpath that is not a routing path in the routing table. For example if the routing path between Nodes 0 and 2 in Figure 7 was Node 0->Node 1->Node 3->Node 2 and the routing path between Nodes 3 and 2 was not Node 3->Node 2 then the routing path between Nodes 0 and 2 would be open because the subpath Node 3->Node 2 is not a path in the routing table. A 2-hop cycle is a group of two hop routing paths (routing paths between two nodes that pass through a third node) such that the first and second nodes in the each two hop routing path are also the second and third nodes in a two hop routing path in the group.
9.5.1
2-Hop Cycle Example
Consider the four node configuration shown in Figure 7. A 2-hop cycle would occur in the configuration if the routing table was configured with the following routing paths: The routing path from Node 0 to Node 3 is: Node 0->Node 1->Node 3. The routing path from Node 1 to Node 2 is: Node 1->Node 3->Node 2. The routing path from Node 2 to Node 1 is: Node 2->Node 0->Node 1. The routing path from Node 3 to Node 0 is: Node 3->Node 2->Node 0.
Chapter 9
279
32559
Rev. 3.08 July 2007
To break this cycle at least one but no more than three of these routing paths must be modified to use a different intermediate node. Reconfiguring the routing paths as follows eliminates the 2-hop cycle. The routing path from Node 0 to Node 3 is: Node 0->Node 1->Node 3. The routing path from Node 1 to Node 2 is: Node 1->Node 3 to Node 2. The routing path from Node 2 to Node 1 is: Node 2->Node 0->Node 1. The routing path from Node 3 to Node 0 is: Node 3->Node 1->Node 0.
Node 3 Node 1
Node 2 Node 0
Figure 7.
A Four-Node Configuration
280
Chapter 9
32559
Rev. 3.08
July 2007
10
Power and Thermal Management
AMD NPT Family 0Fh Processors support ACPI-compliant power management for all classes of systems from notebook PCs to multiprocessor servers. Table 63 lists various levels of power management. Table 64 on page 282 lists ACPI-state support by system class. This chapter covers processor-level power management and processor-specific aspects of system-level power management. Table 63. Power Management Categories
Comments Coarse level power management applied to the entire system, typically after a predefined period of inactivity, or in response to a user action such as pressing the system power button. System Sleep States (ACPI-defined S-states) Processor Level Processor Power States (ACPI-defined C-states) Processor Performance States (ACPI-defined P-states) Software Transparent Power Management (Hardware enforced throttling). Device Level Device Power States (ACPI-defined D-states) Device Performance States Software Transparent Power Management. Bus Level Sub-system Level Bus Power States Software Transparent Power Management Typically not used in PC systems, and beyond the scope of this document.
Power Management Category System Level
Chapter 10
281
32559
Rev. 3.08 July 2007
Table 64.
ACPI-State Support by System Class

Mobile Systems Yes Yes Yes Yes Yes Yes
1,2
ACPI State Support G0/S0/C0: Working G0/S0/C0: Processor performance state (Pstate) transitions under OS control. G0/S0/C0: Thermal clock throttling (I/O hub hardware enforced) G0/S0/C0: Thermal clock throttling (operating system enforced) G0/S0/C1: Halt G0/S0/C2: Stop Grant Caches snoopable
Uniprocessor Desktop PCs Yes Yes Yes Yes Yes Revision F:No Revision G: Yes
1,2
Multiprocessor Systems Yes Yes Yes No Yes No
G0/S0/C3: Stop Grant Caches not snoopable
Yes
1,2
Revision F:No Revision G: Yes

1,2
No
G1/S1: Stand By (Powered On Suspend) G1/S3: Stand By (Suspend to RAM) G1/S4: Hibernate (Suspend to Disk) G2/S5: Shut Down, Turn Off (Soft Off) G3 Mechanical Off
Yes Yes Yes Yes Yes
Yes Yes Yes Yes Yes
Yes Yes Yes Yes Yes
Notes: 1. See Advanced Programmable Interrupt Controller (APIC) on page 235. 2. Hardware can reach these states when both cores of a dual core processor are halted (C1E), but the operating system does not directly initiate C2 or C3 for dual core processors.
10.1
Stop Grant
The processor and chipset use HyperTransport technology STPCLK and Stop Grant system management messages to sequence into all power management states except for Halt. These messages carry a 3-bit System Management Action Field (SMAF) to differentiate the various reasons for placing the processor into the Stop Grant state. Table 65 maps Stop Grant SMAF codes to ACPI states and actions that the BIOS is required to configure the processor to take for each ACPI state.
282
Chapter 10
32559
Rev. 3.08
July 2007
Table 65.
Reason for STPCLK Message
Required SMAF Code to Stop Grant Mapping

SMAF Field of STPCLK and Stop Grant Messages 000b Processors Programmed Response after Entering the Stop Grant State (See Table 75 Power Management Control Registers.) Ramp the CPU clock grid down when no probes need to be serviced. (CPU Low Power Enable) Ramp the CPU clock grid frequency down. (CPU Low Power Enable) After LDTSTOP_L is asserted, place memory into self-refresh and ramp down the processor host bridge and memory controller clock grid. (Northbridge Low Power Enable) After LDTSTOP_L assertion, place memory into self-refresh, ramp the processor clock grids down, then drive new FID values to PLL. (CPU Low Power Enable, Northbridge Low Power Enable, FID/VID Change Enable) Same response as C3.
Mechanism that Forces the I/O Hub to Send the STPCLK Message
C2 Stop Grant Caches Snoopable C3 Stop Grant Caches not Snoopable. Used only by mobile systems.
2
Sent in response to a read of the ACPIdefined P_LVL2 register.
001b
Sent either in response to a read of the ACPI-defined P_LVL3 register or in response to a write to the Link Frequency Change and Resize LDTSTOP_L Command register in the I/O hub.
1
VID/FID Change Or Link Width/ Frequency changes
010b
Sent either in response to a VID/FID Change special cycle from the processor or in response to a write to the Link Frequency Change and Resize LDTSTOP_L Command register.
S1 Sleep state
011b
Sent in response to writing the S1 value to the SLP_TYP[2:0] field and setting the SLP_EN bit of the ACPI-defined PM1 control register in the I/O hub. Sent in response to writing the S3 value to the SLP_TYP[2:0] field of the PM1 control register.
S3 Sleep state
100b
Same response as C3. Additionally, after LDTSTOP_L has been asserted, the I/O hub powers off the main power planes. Ramp down the CPU grid by the programmed amount. (CPU Low Power Enable)
Throttling
3,4
101b
Occurs based upon hardware initiated thermal throttling or ACPI-controlled throttling.
1. GfxMode (Function3, Offset D8h) must be 0 when using this SMAF code to change the link Width/Frequency. 2. This SMAF code may be used in desktop and server systems to change link Width/Frequency. NBLowPwrEn and CpuLowPwrE must be 1 when using this SMAF code to change the link Width/Frequency. 3. AMD recommends using PROCHOT_L for thermal throttling and not implementing stop clock based throttling. 4. STPCLK throttling without ramping the CPU clock grid is supported. This will result in less power savings.
Chapter 10
283
32559
Rev. 3.08 July 2007
Table 65.
Reason for STPCLK Message S4/S5
Required SMAF Code to Stop Grant Mapping (Continued)

SMAF Field of STPCLK and Stop Grant Messages 110b Processors Programmed Response after Entering the Stop Grant State (See Table 75 Power Management Control Registers.) Same response as S3 for processor. Additionally, all power is removed from processor during S4 and S5. Same response as C2, except a Halt special cycle is broadcast.
Mechanism that Forces the I/O Hub to Send the STPCLK Message
Sent in response to writing the S4 or S5 value to the SLP_TYP[2:0] field of the PM1 control register. No I/O hub STPCLK message uses this SMAF code. The power management register field corresponding to this SMAF code is used by the processor in response to executing the Halt instruction.
Reserved for use by processor
111b
1. GfxMode (Function3, Offset D8h) must be 0 when using this SMAF code to change the link Width/Frequency. 2. This SMAF code may be used in desktop and server systems to change link Width/Frequency. NBLowPwrEn and CpuLowPwrE must be 1 when using this SMAF code to change the link Width/Frequency. 3. AMD recommends using PROCHOT_L for thermal throttling and not implementing stop clock based throttling. 4. STPCLK throttling without ramping the CPU clock grid is supported. This will result in less power savings.
10.2
C-States
As Table 63 on page 281 indicates, the processor supports ACPI-defined processor power states, which are referred to as C-states. Table 64 on page 282 indicates which C-states are supported by system class. The operating system places the processor into C-states when the processor is idle an operating-system-determined percentage of the time.
10.2.1
C1 Halt State
C1 is the Halt state. C1 is entered after the HLT instruction has been executed. The BIOS configures the processor to enter a low-power state during C1, in which the CPU core clock grid is ramped down from its operating frequency. See Table 75 on page 316.
10.2.2
C2 and C3
C2 is the Stop Grant state in which the processor caches can be snooped. When the processor is in the Stop Grant state and there is no probe activity to the processors caches, the CPU core clock grid is ramped down from its operating frequency. For more information, see Table 75 on page 316. C3 is the Stop Grant state in which the processors caches cannot be snooped. The chipset asserts LDTSTOP_L during the C3 state. The chipset may require BIOS to enable LDTSTOP_L assertion for the C3 state. When the processor is in the Stop Grant state, the CPU core clock grid is ramped 284 Power and Thermal Management Chapter 10
32559
Rev. 3.08
July 2007
down from its operating frequency. After LDTSTOP_L assertion, the processors system memory is placed into self-refresh mode and the host bridge/memory controller clock grid is ramped down from their operational frequency. For more information, see Table 75 on page 316.
10.2.3
C3 and AltVID
For mobile processors that use discrete graphics, the hardware can be programmed to drive an alternate VID code during the C3 processor power state. The purpose of the AltVID is to reduce the processor voltage to the minimum operational level while the processor is in the C3 state. AltVID settings can be determined by subtracting the AltVidOffset (MSRC001_0042) from the startup voltage.
10.2.4
C1 Enhanced (C1E) Halt State
C1E is a Stop Grant state like C3 supported by dual core mobile processors. The difference between C1E and C3 is that transition into C1E is not initiated by the operating system. The C1E state is initiated by an I/O transaction to the chipset when both cores transition into the C1 state. The chipset will either generate an SMI when receiving the I/O transaction. 10.2.4.1 BIOS Requirements for C1E
Chipsets that support an SMI command port can place the processor in the C1E state by generating an SMI when a write to the port is received. BIOS configures the CPU to generate the I/O write as follows: Program the Interrupt Pending Message register (MSR C001_0055h) in both cores with the address of the SMI command port in the chipset and an the eight bit message data that indicates to the SMM handler that the SMI was generated by both cores entering halt. Set the SmiOnCmpHalt bit and clear the IORd bit in MSR C001_0055h in both cores
The BIOS must not report C2 or C3 support to the OS. The BIOS must not enable C1E when the platform is not in ACPI power management mode. 10.2.4.1.1 C1E SMM Handler Requirements The SMM handler in the BSP is responsible for handling the C1E code for both cores. The SMM handler on the AP should handshake with the BSP to determine when to exit and what state to exit into. The SMM handler should read the SMI command port to determine if the SMI was initiated because both cores went into halt. The SMM handler should read the Auto-Halt Restart byte for both cores to determine if both cores were in halt when entering SMM. If the both cores were not in halt, the SMM handler should exit SMM.
Chapter 10
285
32559
Rev. 3.08 July 2007
Set bit 1 (BM_RLD) of the ACPI PM1 Control register to enable the chipset to bring the cores out of C1E when bus master activity occurs. The SMM handler should then read the BM_STS bit to determine if bus master activity is occurring. If the BM_STS bit is set, the SMM handler should clear the bit, look at the ACPI timer, log the time, and put both cores into C2 by reading the P_LVL2 register. If the BM_STS bit is clear, the SMM handler should read the ACPI timer to see how long it has been since bus master activity last occurred. If the BM_STS bit has been clear for less than 20ms, then the SMM handler should exit SMM mode, and put both cores into the C2 state. If the BM_STS bit has been clear for more than 30ms, then the SMM handler should: Set bit 0 (ARB_DIS) of the ACPI PM2 Control register to disable the system arbiter. Read the P_LVL3 register in the chipset to force both cores into the C3 state. When C1E has been entered, the next timer tick interrupt or bus master activity should cause the chipset to de-assert LDTSTOP_L and issue a STPCLK deassertion message. The SMM handler should then exit both cores back to normal code execution.
10.3
Throttling
The BIOS must declare a DUTY_WIDTH value of zero in the Fixed ACPI Description table if the system does not support the _PSV object as part of a processor ACPI thermal zone. Chipsets must not be configured to assert LDTSTOP_L during STPCLK throttling.
10.3.1
BIOS Requirements for PROCHOT_L Throttling
If PROCHOT_L throttling is being used by the system and HTC is not being used, the BIOS must program the HTC Register (Function 3, offset 64h) as follows to configure the processor for PROCHOT_L throttling. Program HtcEn=1b. Program HtcClkAct and HtcClkInact to 3 us (111b). Program HtcTmpLmt to 1_1111b.
10.4
Processor ACPI Thermal Zone
This section will be part of a future revision of this document.
286
Chapter 10
32559
Rev. 3.08
July 2007
10.5
Processor Performance States
Processor performance states (P-states) are valid operating combinations of processor core voltage and frequency. The hardware supporting P-state transitions in AMD processors is referred to as AMD PowerNow! technology for mobile systems and AMD CoolnQuiet technology for desktop systems. In this document all descriptions of AMD PowerNow! technology, software and driver apply to AMD CoolnQuiet technology, software, and driver. For operating systems without native support for P-state transitions (legacy operating systems), AMD PowerNow! software is used to perform P-state transitions. In this case, a BIOS generated Performance State Block (PSB) is used by AMD PowerNow! software to determine what P-states are supported by the processor in the system. For operating systems that have native support for P-state transitions, a processor-specific driver is used by the operating system to make P-state transitions. In this case, BIOS-provided ACPIdefined P-state objects inform the operating system that P-state transitions are supported by the system, which P-states are supported, and when given P-states are available for use by the operating system.
The valid P-states for each processor are defined using the P-state recognition algorithm. The BIOS must only use P-states defined for a given processor when building the ACPI P-state objects and PSB for use by higher-level software. BIOS is not required to implement all of the valid P-states for a processor. MSR C001_0041h (FIDVID_CTL) and MSR C001_0042h (FIDVID_STATUS) are used for initiating P-state transitions and verifying that they are complete. Chipsets provide support for P-state transitions. Refer to the chipset documentation for information on chipset configuration and P-state-related considerations.
10.5.1
BIOS Requirements for P-State Transitions
P-state transitions can be used only if they are supported by the processor and by the system. BIOS is required to: Have a valid set of P-states for the processor, based on the P-state recognition algorithm. Take chipset or system limitations into account before providing the ACPI-defined P-state objects described in 10.6 on page 303, or the PSB described in 10.8 on page 313 of this document. BIOS is required to provide both: ACPI-defined P-state objects for operating systems that support native P-state control. a PSB in support of AMD PowerNow! software with legacy operating systems.
The BIOS must not provide the ACPI P-state objects or the PSB if: The chipset or system does not support P-state transitions. Power and Thermal Management 287
Chapter 10
32559
Rev. 3.08 July 2007
The BIOS does not have a set of valid P-states P-state Recognition Algorithm
10.5.1.1
If CPUID extended function 8000_0007h returned EDX[2:1] = 11b, the BIOS must correlate the processor in the system to a valid set of P-states using this algorithm. 1. Determine Maximum (P[Max] or P[0]) and Minimum (P[Min]) P-states. If MaxFID = 10_1010b and MaxVID != 00_0000b P[Max] FID = start FID + 00_1010b (Start Frequency + 1GHz) P[Max] VID = MaxVID + 00_0010b (Maximum Voltage - 50mV) P[Min] FID = 00_0010b P[Min] VID = start VID else P[Max] FID = MaxFID P[Max] VID = MaxVID + 00_0010b P[Min] FID = start FID P[Min] VID = start VID 2. Generate intermediate P-states based on the following rules: No intermediate P-states are supported if IntPstateSup = 0b. No intermediate P-state can have a VID >= P[Min] VID. No intermediate P-state can have a VID <= P[Max] VID. No intermediate P-state can have a FID < P[Min] FID + 00_1000b. No intermediate P-state can have a FID >= P[Max] FID. The highest intermediate P-state (P[1]) must have an even FID. If P[Max] FID is odd: P[1] FID = P[Max] FID - 00_0001b else: P[1] FID = P[Max] FID - 00_0010b The P[1] VID is based on P[Max] VID. If P[Max] FID is odd: P[1] VID = P[Max] VID + 00_0001b else: P[1] VID = P[Max] VID + 2^PstateStep The FID/VID for the current P-state (P[N]) is based on the FID/VID of the next highest P-state (P[N-1]). If P[N-1] VID + 2^PstateStep >= P[Min]: P[N] = P[N-1] else: P[N] = P[N-1] VID + 2^PstateStep P[N] FID = P[N-1] FID - 00_0010b 3. Calculate the power for each P-state according to the following formula: Power = (PwrLmt * P[N] Frequency * (P[N] Voltage^2))/(P[0] Frequency* (P[0] Voltage^2)). 288 Power and Thermal Management Chapter 10
32559
Rev. 3.08
July 2007
The PwrLmt (CPUID Fn[8000_0001]_EBX[8:6,14]) field encodes the TDP for the processor. The TDP associated with the PwrLmt encoding can be determined from the appropriate Power and Thermal Datasheet. If the BIOS does not contain a TDP for the PwrLmt encoding, the BIOS should increment the PwrLmt value by one until a valid TDP is found. This TDP should be used to calculate the power. If no valid TDP is found the BIOS must disable P-state support.
10.5.2
BIOS Requirements for P-State Transitions in a Multiprocessor System
In a multiprocessor system, P-state transitions can only be used if they are supported by all processors. In a multiprocessor system, the BIOS is required to: Have a valid set of P-states for all processor. BIOS should follow the steps outlined in 10.5.1.1 on page 288. Take chipset or system limitations into account before providing the ACPI-defined P-state objects. Provide the same number of P-states for all processors, as per Section 8.3.3 of the ACPI 2.0 specification. Ensure that all processor performance states in the _PSS object have identical frequency and power-consumption parameters as per Section 8.3.3 of the ACPI 2.0 specification. Transition all processors to a voltage that is equal to or greater than the recommended voltage for the current frequency (valid voltage and frequency combinations are determined using the P-state recognition algorithm). Transition all processors to the maximum frequency of the slowest processor. Ensure that the frequency of each processor is one of the valid P-state frequencies. Copy the CurrVID (MSR C001_0042h, bits 36-32) value to NewVID (MSR C001_0041h, bits 12-8) for each processor. Copy the CurrFID (MSR C001_0042h, bits 5-0) value to NewFID (MSR C001_0041h, bits 5-0) for each processor. Set StpGntTOCnt (MSR C001_0041h, bits 51-32) to a value of 1h for each processor. Set the Clock Power/Timing High register (Function 3, Offset D8h) to a value of 2000_2710h for each processor.
In a multiprocessor system, the BIOS must not provide the ACPI P-state objects if: Any of the processors cannot be identified. Power and Thermal Management 289
Chapter 10
32559
Rev. 3.08 July 2007
The chipset or system does not support P-state transitions. The BIOS does not have a set of valid P-states.
In a multiprocessor system, the BIOS should follow the P-state transitioning algorithm for each processor as defined in 10.5.6 on page 291.
10.5.3
BIOS Requirements for P-State Transitions in Systems Using Dual Core Processors
For systems with a single processor and unbuffered DIMMs or SODIMMs, BIOS configures the processor in the same way as described in Section 10.5.1 with the exception that one set of the _PCT, _PSS, and _PPC ACPI objects must be declared for each processor core. For systems with a multiple processors, or systems with Registered DIMMs, BIOS declares one set of _PCT, _PSS, and _PPC ACPI objects per processor core, and configures the processor as described in Section 10.5.2.
10.5.4
BIOS-Initiated P-State Transitions
All processors power up in the minimum P-state. BIOS can transition processors from the minimum P-state to a higher P-state during the POST routine. If BIOS performs a P-state transition it must follow the P-state transition algorithm defined in 10.5.7 on page 291. If the BIOS performs P-state transitions during POST on dual core processor, BIOS must enable core 1 by setting the Cpu1En bit (Function 0, Offset 68h) before initiating the P-state change.If the BIOS performs P-state transitions during POST in a multiprocessor system, BIOS must enable all processors by setting the ReqDis bit (Function 0, Offset 6Ch) before initiating the P-state change. The BIOS must never initiate P-state transitions after passing control to the operating system.
10.5.5
BIOS Support for Operating System/CPU Driver-Initiated P-State Transitions
Operating systems that have native control for processor P-states exclusively use the presence of BIOS provided ACPI 2.0 defined processor P-state objects to determine if processor P-states are supported in a system. If the ACPI objects do not exist, then the processor driver is not loaded. In operating systems that do not have native support for processor P-state transitions, an AMD defined PSB provided by the BIOS informs AMD PowerNow! technology software that P-state transitions are supported. Once it has been determined that a processor and system support P-state transitions, the BIOS must provide both: ACPI-2.0-defined processor P-state objects for operating systems with native P-state support. See Section 10.6 on page 303.
290
Chapter 10
32559
Rev. 3.08
July 2007
AMD-defined PSB for use with AMD PowerNow! software and operating systems that do not have native P-state support. See Section 10.8 on page 313.
10.5.6
Processor Driver Requirements
The processor driver that supports native operating system policy and control of P-state objects is required to use the ACPI 2.0 P-state objects as defined in this document. The processor driver is not permitted to use the legacy PSB tables described in section 10.8 on page 313.
10.5.7
P-State Transition Sequence
This section describes the P-state transition algorithm for AMD NPT Family 0Fh Processors. 10.5.7.1 FID to VCO Frequency Relationship
Table 66 and Table 67 provide the core frequency-to-VCO frequency relationship for AMD NPT Family 0Fh Processors. These processors do not support frequency transitions in VCO frequency steps greater than 200 MHz. The processor driver, and the AMD PowerNow! driver use these tables when making P-state transitions. Only one P-state is allowed in Table 66. The one P-state consists of a frequency from the Minimum Core Frequency column in Table 66 and is the minimum P-State supported by the processor. Table 66 additionally defines Portal Core Frequencies. The Portal Core Frequencies in Table 67 column of Table 66 defines: The core frequencies in Table 67 to which direct transitions can be made from the minimum core frequency in Table 66. The core frequencies in Table 67 that support transitions to the minimum core frequency in Table 66.
The processor supports direct transitions between the minimum core frequency in Table 66 and any of the core frequencies from Table 67 listed in the Portal Core Frequencies column associated with the minimum core frequency. If the minimum P-state from the low table (Table 66) has a VCO frequency > 1600 MHz, then the VCO/core frequency of all P- states in the high table (Table 67) must be >= the VCO frequency of the minimum P-state minus 200 MHz. Example 1: Given a processor with a minimum core frequency of 800 MHz and a maximum core frequency of 2000 MHz, to transition from 800 MHz to 2000 MHz, the processor driver will (in the order specified): 1. Transition the processor from the 800 MHz core frequency to 1800 MHz core frequency. 1800 MHz is a portal frequency to and from 800 MHz as defined in Table 66, because the VCO Chapter 10 Power and Thermal Management 291
32559
Rev. 3.08 July 2007
frequency change from an 800 MHz core frequency to an 1800 MHz core frequency is <= 200 MHz. 2. Transition the processor core frequency from 1800 MHz to 2000 MHz (VCO frequency change is <= 200 MHz according to Table 67). Note: To simplify this example, the voltage transitions and isochronous relief times are not described in this example. Example 2: Given a processor with a minimum core frequency of 1400 MHz and a maximum core frequency of 3400 MHz, to transition from 1400 MHz to 3400 MHz, the processor driver will (in the order specified): 1. Transition the processor from the core frequency of 1400 MHz to a core frequency of 3000 MHz. 3000 MHz is a portal frequency to and from 1400 MHz as defined in Table 66, because the VCO frequency change from 1400 MHz core frequency to 3000 MHz core frequency is <= 200 MHz. 2. Transition the processor core frequency from 3000 MHz to 3200 MHz (VCO frequency change is <= 200 MHz per Table 67). 3. Transition the processor core frequency from 3200 MHz to 3400 MHz (VCO frequency change is <= 200 MHz per Table 67). Note: Note: to simplify this example, the voltage transitions and isochronous relief times are not described in this example. Table 66.
FID[5:0] 000000b 000001b 000010b 000011b 000100b 000101b 000110b 000111b
Low FID Frequency Table (< 1600 MHz)

Minimum Core Frequency MHz 800 900 1000 1100 1200 1300 1400 1500 VCO Frequency MHz 1600 1800 2000 2200 2400 2600 2800 3000 Portal Core Frequencies in Table 67 1600, 1700, 1800 1600, 1700, 1800, 1900, 2000 1800, 1900, 2000, 2100, 2200 2000, 2100, 2200, 2300, 2400 2200, 2300, 2400, 2500, 2600 2400, 2500, 2600, 2700, 2800 2600, 2700, 2800, 2900, 3000 2800, 2900, 3000, 3100, 3200
Note: Odd FID values are supported in revision G and later revisions.
292
Chapter 10
32559
Rev. 3.08
July 2007
Table 67.
FID[5:0] 001000b 001001b 001010b 001011b 001100b 001101b 001110b 001111b 010000b 010001b 010010b 010011b 010100b 010101b 010110b 010111b 011000b 011001b 011010b 011011b 011100b 011101b 011110b 011111b 100000b 100001b 100010b 100011b 100100b 100101b 100110b 100111b 101000b
High FID Frequency Table (>= 1600 MHz)

Core Frequency MHz 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 2600 2700 2800 2900 3000 3100 3200 3300 3400 3500 3600 3700 3800 3900 4000 4100 4200 4300 4400 4500 4600 4700 4800 VCO Frequency MHz 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 2600 2700 2800 2900 3000 3100 3200 3300 3400 3500 3600 3700 3800 3900 4000 4100 4200 4300 4400 4500 4600 4700 4800
Chapter 10
293
32559
Rev. 3.08 July 2007
Table 67.
FID[5:0] 101001b 101010b
High FID Frequency Table (>= 1600 MHz) (Continued)

Core Frequency MHz 4900 5000 VCO Frequency MHz 4900 5000
10.5.7.2
P-state Transition Algorithm
The P-state transition algorithm has three phases. During phase 1 the processor voltage is transitioned to the level required to support frequency transitions. During phase 2 the processor frequency is transitioned to frequency associated with the OS-requested P-state. During phase 3 the processor voltage is transitioned to the voltage associated with the OS-requested P-state. Figure 8 shows the high-level P-state transition flow. Figure 9 is a high-level timing diagram depicting voltage and frequency stepping associated with each phase of a P-state transition. Figure 10 shows the core voltage transition flow for phase 1 of a P-state transition. Figure 12 shows the core frequency transition flow for phase 2 of a P-state transition. Figure 15 shows the core voltage transition flow for phase 3 of a P-state transition. The algorithm shown in Figure 8 through Figure 15 describes the 3-phases of a P-state transition in the context of Windows XP and the associated processor driver, but the algorithm also applies to the AMD PowerNow! driver and any other software that performs P-state transitions (such as a debug tool).
294
Chapter 10
32559
Rev. 3.08
July 2007
OS: determines a P-state transition is required and calls the processor driver to make the transition to a specified P-state.
Processor driver: informs the OS that the requested P-state transition is complete, and begins the transition. Phase 1: Processor driver: perform a series of VID-only transitions each increasing the voltage by the maximum voltage step (MVS) to change the voltage: To the voltage for the requested P-state plus the voltage indicated by the ramp voltage offset (RVO), if the requested P-state is greater in frequency than the current P-state. To the voltage indicated by the CurrVID plus the voltage indicated by the RVO, if the requested P-state is lower in frequency than the current P-state. Software counts off voltage stabilization time (VST) after each voltage step.
Phase 2: Processor driver: changes the processor core frequency to the frequency for the OS-requested P-state.
Phase 3: Processor driver: performs a VID-only change to the VID for the OS requested P-state.
Figure 8.
High-Level P-state Transition Flow
Chapter 10
295
Example P-State Transition Timing Diagram

1.550V 1.550V RVO removed. 1.500 V 1.525V 1.500V
VST = 100s VST = 100s VST = 100s VST = 100s

001010 1800 1800 StpGntTOCnt
Processor driver counts IRT.
Phase 1: Increase the voltage. Add RVO
Phase 2: Change frequency.
Processor driver counts IRT.
296
00011 00001 00010 00000 00010 1.475V
MVS = 25 mV VST = 100s
Figure 9.
001100 2000 2000 StpGntTOCnt Current P-State 2000 [email protected] Phase 3: Transition VID to remove RVO Not drawn to scale. This figure depicts a P-state transition from a P-state of 800 MHz at 1.45 V to a P-state of 2000 MHz at 1.50 V. For MVS see section 9.6.2.1.2. For StpGntTOCnt see section 9.5.5.2.2. For IRT see section 9.6.2.1.5. For RVO and VID see section 9.6.2.1.4. For VST see section 9.6.2.1.1.
VID[4:0]
00100
VDD
1.45V
FID[5:0]
000000
Core Frequency (MHz)
800
VCO Frequency 1600 (MHz)
Current P-State 800 [email protected]
32559
Rev. 3.08 July 2007
Chapter 10
32559
Rev. 3.08
July 2007
TargetVID = VID after RVO applied CurrentVID = VID read from FIDVID_STATUS_MSR
See Figure 11 TargetVID Calculation
Is CurrentVID TargetVID?
N Write VID-only change to FIDVID_CONTROL_MSR as indicated by MVS.
Voltage Transition Complete
Read current state from FIDVID_STATUS_MSR. N
Is Pending bit == 0 in FIDVID_STATUS_MSR?
Y Processor driver counts off voltage stabilization time. (VST)
Figure 10. Phase 1: Core Voltage Transition Flow
Chapter 10
297
32559
Rev. 3.08 July 2007
FinalFID = OS-Requested FID TargetFID = FinalFID truncated to an even FID CurrenFID = FID read from FIDVID_Status_MSR InitialFID = CurrentFID truncated to an even FID FinalVID = OS-requested VID TargetVID = VID after RVO applied CurrenVID = VID read from FIDVID_Status_MSR
TargetFID > InitialFID?
TargetVID = FinalVID RVO Is the CurrentVID > FinalVID Y
TargetVID = FinalVID RVO
TargetVID = CurrentVID RVO
Figure 11. TargetVID Calculation
298
Chapter 10
32559
Rev. 3.08
July 2007
FinalFID = OS-requested FID TargetFID = FinalFID, truncated to an even FID CurrentFID = FID read from FIDVID_STATUS_MSR
Is CurrentFID == FinalFID?
FID Transition Process
FID Transition Process To next lower even FID
Is CurrentFID an odd FID?
Write FID-only changes to FIDVID_CONTROL_MSR.
N Y Is CurrentFID < TargetFID? Y See Figure 13 Transition to a Higher Core Frequency Is Pending bit == 0 in FIDVID_STATUS_MSR? Read current state from FIDVID_STATUS_MSR.
N See Figure 14 Transition to a Lower Core Frequency
Is CurrentFID > TargetFID?
FID Transition Process To FinalFID
Is CurrentFID == FinalFID?
Processor driver counts off Isochronous Relief Time (IRT)
The processor core is now at the frequency requested by the OS.
FID Transition Complete
Figure 12. Phase 2: Core Frequency Transition Flow
Chapter 10
299
32559
Rev. 3.08 July 2007
FinalFID = OS-requested FID TargetFID = FinalFID, truncated to an even FID CurrentFID = FID read from the FIDVID_STATUS_MSR
Is CurrentFID in the High FID Frequency Table?
Is TargetFID above VCO portal of CurrentFID?
FID Transition Process To TargetFID
Y
FID Transition Process To next higher FID in High FID Frequency Table
Y
FID Transition Process To highest portal FID in High FID Frequency Table
Is CurrentFID == TargetFID?
The processor core is now at TargetFID.
Figure 13. Transition to a Higher Core Frequency
300
Chapter 10
32559
Rev. 3.08
July 2007
FinalFID = OS-requested FID TargetFID = FinalFID, truncated to an even FID CurrentFID = FID read from the FIDVID_STATUS_MSR
Is TargetFID in the High FID Frequency Table?
Is TargetFID below VCO portal of CurrentFID?
Y
FID Transition Process To next lower FID in High FID Frequency Table
Y
FID Transition Process To next lower FID in High FID Frequency Table
Is CurrentFID == TargetFID?
Is TargetFID below VCO portal of CurrentFID?
FID Transition Process ToTargetFID
The processor core is now at TargetFID.
Figure 14. Transition to a Lower Core Frequency
Chapter 10
301
32559
Rev. 3.08 July 2007
Initiate a VID-only transition directly to the VID code for the P-state requested by the OS. This transition is made by following the steps in section 10.5.7.2.1
Count off the VST (See section 10.6.2.1.1)
Figure 15. Phase 3: Core Voltage Transition Flow 10.5.7.2.1 Changing the VID Note: Software must hold the FID constant when changing the VID. To change the processor voltage: 1. Write the following values to FIDVID_CTL (MSR C001_0041h): NewVID field (bits 138) with the VID code associated with the target voltage NewFID field (bits 50) with the CurrFID value indicated in the FIDVID_STATUS MSR InitFidVid bit (bit 16) set to 1. Setting this bit initiates the VID change. Clear all other bits to 0. 2. Loop on reading the FidVidPending bit (bit 31) of FIDVID_STATUS (MSR C001_0042h) until the bit returns 0. The FidVidPending bit stays set to 1 until the new VID code has been driven to the voltage regulator. 10.5.7.2.2 Changing the FID AMD NPT Family 0Fh Processors support direct FID transitions only between FIDs that represent a VCO frequency change less then or equal to 200 MHz. To change between FIDs that represent a VCO frequency change greater than 200 MHz, software must make multiple transitions each less than or equal to a 200 MHz change in VCO frequency. For information about allowable FID transitions, see Tables 66 and 67 on page 292. Note: Software must hold the VID constant when changing the FID. To change the core frequency:
302
Chapter 10
32559
Rev. 3.08
July 2007
1. Write the following values to FIDVID_CTL (MSR C001_0041h): NewVID field (bits 138) with the CurrVID value reported in the FIDVID_STATUS MSR. NewFID field (bits 50) with the FID code associated with the target frequency. StpGntTOCnt field (bits 5132) with a value corresponding to the processor PLL lock time. The PLL lock time is specified in section 14.2.9.1 on page 411. PLL lock time is communicated to the processor driver through the PLL_LOCK_TIME field of the _PSS object. PLL lock time is in microseconds. To translate the PLL lock time to an StpGntTOCnt value, multiply PLL_LOCK_TIME by 1000 to get nanoseconds, then divide by 5 which is the clock period of the counter in ns. Therefore, StpGntTOCnt value = PLL_LOCK_TIME 1000/5. For example, a PLL lock time of 2 s results in an StpGntTOCnt value of 400 decimal and 190h. InitFidVid bit (bit 16) set to 1. Setting this bit initiates the P-state transition. Clear all other bits to 0. 2. Loop on reading the FidVidPending bit (bit 31) of FIDVID_STATUS (MSR C001_0042h) until the bit returns 0. The FidVidPending bit stays set to 1 until the new FID code is in effect. Figure 12 on page 299 illustrates the transition flow for core frequencies, including the requirement to wait the isochronous relief time between FID steps. 10.5.7.3 Dual Core P-state Transition Requirements
For dual core processors, it is the processor drivers responsibility to coordinate operating systems P-state transition requests for the individual cores. The cores in a dual core processor share a common clock grid and voltage plane, and therefore do not support independent P-states even though the operating system views the cores as independent processors, and a set of ACPI P-state objects is declared to the operating system for each core. The processor driver enforces a highest requested P-state policy. Both cores will be in the highest P-state requested for either core. The processor driver keeps track of P-state transition requests and only reduces the P-state of either core after the operating system has requested that the P-state of both cores be reduced. When the operating system requests that the P-state of either core be increased, the processor driver increases the P-state of both cores.
10.6
ACPI 2.0 Processor P-State Objects
For systems that support P-state transitions, the BIOS must provide the following ACPI 2.0 P-state objects to support operating systems that provide native support for processor P-state transitions. These subsections discuss the specific implementation of these objects. In multiprocessor systems that support P-state transitions, the BIOS must provide the ACPI 2.0 P-state objects defined in this section for each processor. For dual core processors, one set of ACPI P-state objects is declared for each core.
Chapter 10
303
32559
Rev. 3.08 July 2007
10.6.1
_PCT (Performance Control)
The _PCT object is generically described in section 8.3.3.1 of the ACPI 2.0 specification. The _PCT, _PSS, and _PPC objects (defined in the next sections) are all placed after the Processor object in the \_PR name space for operating systems that have native support for processor P-states. This section provides BIOS specifics regarding the use of _PCT. The _PCT object specifies the control register and status register used to initiate and verify P-state transitions. The processors P-state transition protocol is abstracted from the operating system by the processor driver, so this object does not declare the MSRs used for FID_Change protocol, but rather declares these two registers to be functional fixed hardware. Functional fixed hardware means that the mechanism is specific to the processor, and the processor driver knows how to initiate and verify P-state transitions because the processor driver is written by the processor vendor or written based upon documentation supplied to the operating system vendor by the processor vendor. The following is a sample _PCT object. Sections referenced are in the ACPI 2.0 specification and ACPI 2.0 Errata Document, revision 1.4.
Name (_PCT, Package (2) // Performance Control Object { // ResourceTemplate () {Register(FFixedHW, 0, 0, 0)} // Control Buffer () { 0x82, // B0-Generic Register Descriptor (Sec. 6.4.3.7) 0xC,0, // B1:2-length (from _ASI through _ADR fields) 0x7F, // B3-Address space ID, _ASI, SystemIO 0, // B4-Register Bit Width, _RBW 0, // B5-Register Bit Offset, _RBO 0, // B6-Reserved. Must be 0. 0,0,0,0,0,0,0,0, // B7:14-register address, _ADR (64bits) 0x79,0}, // B15:16-End Tag (Section 6.4.2.8) // 6.4.3.7) 0xC,0, 0x7F, 0, 0, 0, 0,0,0,0,0,0,0,0, 0x79,0}, }) // End of _PCT object // // // // // // // B1:2-length (2 bytes) B3-Address space ID, _ASI, SystemIO B4-Register Bit Width, _RBW B5-Register Bit Offset, _RBO B6-Reserved. Must be 0. B7:14-register address, _ADR (64bits) B15:16-End Tag (Section 6.4.2.8) ResourceTemplate () {Register(FFixedHW, 0, 0, 0)} // Status Buffer () { 0x82, // B0-Generic Register Descriptor (Section
10.6.2
_PSS (Performance-Supported States)
The _PSS object is generically described in Section 8.3.3.2 of the ACPI 2.0 specification. This object follows the _PCT object in the \_PR name space in support of operating systems with native support for processor P-states. This section provides the BIOS specifics regarding use of _PSS. 304 Power and Thermal Management Chapter 10
32559
Rev. 3.08
July 2007
The _PSS object specifies the performance states that are supported by the system. Based upon the maximum voltage and frequency supported by the processor and system requirements, the BIOS builds the _PSS object to specify the performance states supported by the system. The BIOS creates the _PSS object based on the valid P-states determined by the BIOS. The BIOS is required to maintain a set of Performance State Tables (PSTs) sheet for all processors capable of Pstate transitions. Only the PST entries that correspond to the processor found in the specific system by the BIOS during the POST routine are used by the BIOS to create the _PSS object for that system. The _PSS object can have as many performance states as the BIOS determines are supported by the processor. For example, the _PSS object could provide performance state definitions P0P2: P0: 2400 MHz at 1.20 V P1: 1600 MHz at 1.10 V P2: 800 MHz at 1.00 V The generic _PSS object description has the following format: Name (_PSS, Package() { // Field Name Field Type Package () // Performance State 0 Definition P0 { CoreFreq, // DWordConst Power, // DWordConst TransitionLatency, // DWordConst BusMasterLatency, // DWordConst Control, // DWordConst Status // DWordConst }, Package () // Performance State n Definition Pn { CoreFreq, // DWordConst Power, // DWordConst TransitionLatency, // DWordConst BusMasterLatency, // DWordConst Control, // DWordConst Status // DWordConst } }) // End of _PSS object Each performance state entry contains six data fields as follows: CoreFreq is the core CPU operating frequency (in MHz).
Chapter 10
305
32559
Rev. 3.08 July 2007
Power is the typical power dissipation (in milliWatts). The BIOS must fill out this field for the each performance state using the power calculated when determining which P-states are supported. TransitionLatency is the worst-case latency, in microseconds, that the CPU is unavailable during a transition from any performance state to this performance state. During a P-state transition, the CPU is unavailable for no more than 6 s for each frequency step. While the total P-state transition could take up to 2 ms, the operating system is not blocked during the transition. Therefore the recommended value for this field is 100 s. BusMasterLatency is the worst-case latency, in microseconds, that Bus Masters are prevented from accessing memory during a transition from any performance state to this performance state. This value is estimated at 7 s. Control is the value to be written to the _PSS Control Field in order to initiate a transition to the performance state. The BIOS must fill this out as described on page 306. Status is the value that the processor driver can compare to a value read from the _PSS Status Field to ensure that the transition to the performance state was successful. The BIOS must fill this out as described on page 310. _PSS Control Field
10.6.2.1
The control field of each performance state definition within the _PSS object contains the information that the processor driver uses. This includes NewVID and NewFID values that must be written into the FIDVID_CTL register (MSR C001_0041) to initiate transitions between P-states. _PSS Control field has the following format:
31 30 29 28 27 26 IRT RVO ExtType PLL_LOCK_TIME 20 19 18 17 MVS VST 11 10 NewVID 6 5 NewFID 0
Bits 3130 2928 27 2620 1918 1711 106 IRT RVO
Field Ramp Voltage Offset Extended Interface = 1.
Description Isochronous Relief Time
ExtType PLL_LOCK_TIME MVS VST NewVID
Phase-Locked Loop Time Maximum Voltage Step Voltage Stabilization Time The NewVID[4:0] value associated with the OS-requested P-state. CurrVID[5] of the _PSS Status field specifies the OS-requested NewVID[5] value. The FID value associated with the OS-requested P-state
50
NewFID
The following sections describe the sub-fields of the _PSS control field.
306
Chapter 10
32559
Rev. 3.08
July 2007
10.6.2.1.1 Voltage Stabilization Time The Voltage Stabilization Time (VST) defines the number of microseconds (in 20 s increments) that are required for the voltage to increase by the amount specified by the MVS field when the VID outputs of the processor transition. The processor driver counts VST between VID steps that increase the processors core voltage, not between steps that decrease it. Table 68 shows the actual stabilization time for several VST values. Table 68.
7'h00 7'h01 7'h02 7'h64 7'h7e 7'h7f
Sample VST Values

0 s 20 s 40 s (BIOS default) ... 2000 s 2520 s 2540 s
VST (Bits 1711) Stabilization Time
10.6.2.1.2 Maximum Voltage Step The Maximum Voltage Step (MVS) defines the maximum voltage increment the processor driver can use when changing from a lower voltage to a higher voltage. For the processor driver, when increasing core voltage, the next VID = CurrVID 2 . Table 69 on page 308 shows the values and increments for this field.
MVS
For example, if the voltage is being increased, and the current VID is 00100 (1.45 V), the OS-requested target VID is 00000 (1.55 V), and MVS is 0 (which selects 25mV voltage steps).
Next_VID = CurrVID - 2
MVS
Then, the processor driver calculates the first VID to which to transition as follows:
= 00100 - 2 = 00100 - 1 = 00011 = 1.475 V
0
Next the processor driver transitions the VID to select 1.475 V. After the VST is counted off, the processor driver calculates the next VID to which to transition as follows: Next_VID = CurrVID - 2
MVS
= 00011 - 2 = 00010 = 1.500 V

0
This process iterates until the CurrVID is equal to the OS-requested target VID minus the VID associated with the ramp voltage offset. After each VID transition the processor driver counts off the voltage stabilization time before calculating the Next_VID. Chapter 10 Power and Thermal Management 307
32559
Rev. 3.08 July 2007
Table 69.
00 01 10 11
MVS Values
25 mV (BIOS default. This is the required value that must be used.) 50 mV (Reserved for future use) 100 mV (Reserved for future use) 200 mV (Reserved for future use)
MVS (Bits 1918) Increment
10.6.2.1.3 PLL Lock Time The 7-bit binary PLL_LOCK_TIME value defines the time required for the processor PLLs to lock in microseconds. The PLL lock time is specified in section 14.2.9.1 on page 411. 10.6.2.1.4 Ramp Voltage Offset The Ramp Voltage Offset (RVO) is the offset voltage applied to the VID requested VDD supply level (VID_VDD) voltage when performing P-state transitions. The RVO is specified in the processor data sheet. Table 70 lists the RVO values and their corresponding offset voltages. Table 70. RVO Values
Offset Voltage 0 mV 25 mV 50 mV 75 mV Description Target P-state VID needs no adjustment for RVO. (BIOS default for Multiprocessor Systems only. This is the required value.) Decrement the target P-state VID by 1 to add the RVO. Decrement the target P-state VID by 2 to add the RVO. (BIOS default. This is the required value.) Decrement the target P-state VID by 3 to add the RVO.
RVO (Bits 2928) 00 01 10 11
For the processor driver, RVO is in VID increments. To increase the processor voltage the processor driver must subtract the RVO field from the VID. The MVS field dictates the increments in which RVO can be subtracted from VID. For example, consider the case when the VID is 00100b (1.45 V), RVO is 10b (50 mV), MVS is 00b (25 mV), and VST is 7h05 (100 s). Based upon these parameters, the voltage must be increased by 50 mV in two steps of 25 mV with a 100 s delay after each VID change. The steps to apply the RVO are: 1. The VID code is reduced to 00011b. This transitions the voltage from 1.45 V to 1.475 V. 2. The processor driver waits 100 s as specified by VST. 3. The VID code is reduced to 00010b. This transitions the voltage from 1.475 V to 1.5 V.
308
Chapter 10
32559
Rev. 3.08
July 2007
4. The processor driver waits 100 s as specified by VST. Table 71 lists the VID codes and their corresponding voltages for AMD NPT Family 0Fh Processors. Table 71.
VID[5:0]
VID Code Voltages

Voltage VID[5:0] Voltage VID[5:0] Voltage VID[5:0] Voltage 01_0000b 1.150 V 01_0001b 1.125 V 01_0010b 1.100 V 01_0011b 1.075 V 01_0100b 1.050 V 01_0101b 1.025 V 01_0110b 1.000 V 01_0111b 0.975 V 01_1000b 0.950 V 01_1001b 0.925 V 01_1010b 0.900 V 01_1011b 0.875 V 01_1100b 0.850 V 01_1101b 0.825 V 01_1110b 01_1111b 0.800 V 0.775 V 10_0000b 0.7625 V 10_0001b 0.7500 V 10_0010b 0.7375 V 10_0011b 0.7250 V 10_0100b 0.7125 V 10_0101b 0.7000 V 10_0110b 0.6875 V 10_0111b 0.6750 V 10_1000b 0.6625 V 10_1001b 0.6500 V 10_1010b 0.6375 V 10_1011b 0.6250 V 10_1100b 0.6125 V 10_1101b 0.6000 V 10_1110b 10_1111b 0.5875 V 0.5750 V 11_0000b 0.5625 V 11_0001b 0.5500 V 11_0010b 0.5375 V 11_0011b 0.5250 V 11_0100b 0.5125 V 11_0101b 0.5000 V 11_0110b 11_0111b 0.4875 V 0.4750 V
00_0000b 1.550 V 00_0001b 1.525 V 00_0010b 1.500 V 00_0011b 1.475 V 00_0100b 1.450 V 00_0101b 1.425 V 00_0110b 1.400 V 00_0111b 1.375 V 00_1000b 1.350 V 00_1001b 1.325 V 00_1010b 1.300 V 00_1011b 1.275 V 00_1100b 1.250 V 00_1101b 1.225 V 00_1110b 00_1111b 1.200 V 1.175 V
11_1000b 0.4625 V 11_1001b 0.4500 V 11_1010b 0.4375 V 11_1011b 11_1100b 11_1101b 11_1110b 11_1111b 0.4250 V 0.4125 V 0.4000 V 0.3875 V 0.3750 V
10.6.2.1.5 Isochronous Relief Time The Isochronous Relief Time (IRT) is the amount of time the processor driver must count after each FID change step in a given P-state transition. During each FID change step, system memory is inaccessible to bus masters. While the processor driver is counting IRT between FID change steps, bus masters have access to system memory. Table 72 lists the IRT values and their corresponding times. Table 72.
00 01 10 11
IRT Values
Time 10 s 20 s 40 s 80 s (BIOS default)
IRT (Bits 3130)
Chapter 10
309
32559
Rev. 3.08 July 2007
10.6.2.2
_PSS Status Field
The status field of each performance state definition within the _PSS object contains the information required by the processor driver to verify that the transition has completed based upon a read of the FIDVID_STATUS MSR. The _PSS Status field is 32 bits, and the BIOS must map the FIDVID_STATUS MSR CurrVID and CurrFID information to the _PSS status field as shown in Table 73 on page 310. Table 73.
Bit 50 116 2812 3129
:
_PSS Status Field

Name CurrFID CurrVID Reserved Version Description The FID value associated with the OS-requested P-state The VID value associated with the OS-requested P-state These bits must be 0 and are ignored by the processor driver. Version = 0
The CurrFID and CurrVID fields of the FIDVID_STATUS MSR are valid only when the FidVidPending bit is 0.
10.6.3
_PPC (Performance Present Capabilities)
The _PPC object is used to limit the maximum P-state that the operating system can use at any given time. This object follows the _PSS object in the \_PR name space for the operating systems that support the ACPI 2.0 processor P-state objects. _PPC is generically described in Section 8.3.3.3 of the ACPI 2.0 specification. This section provides BIOS specifics regarding use of _PPC. If the OEM and BIOS vendor do not implement a _PPC that limits the maximum P-state under certain conditions, then the BIOS must always return a value of 0 for the _PPC object. When a value of 0 is returned for the _PPC object, all P-states specified by the _PSS object are available. Operating systems with native support for processor P-state transitions and ACPI Thermal zones make use of reduced P-states for passive cooling before making use of throttling with the stop grant state. This is the case even when the _PPC object returns a value of 0. 10.6.3.1 Using _PPC to Limit the Maximum Processor P-State When battery Powered
Processor performance exceeds the demands of most typical notebook PC applications. However, there are basically idle scenarios where applications run instructions that incorrectly cause the operating system to conclude that the processor is 100% utilized. This causes the operating system to force the processor to its maximum performance state, wasting power with no benefit to the user. Also, static power consumption of the processor increases when the processor is operating in a P-state above the minimum voltage level supported by the processor. 310 Power and Thermal Management Chapter 10
32559
Rev. 3.08
July 2007
Therefore, it increases the systems battery-powered runtime to limit the maximum performance state available when battery-powered. 10.6.3.2 Implementation Overview
An ACPI-compliant General-Purpose Event (GPE) input of the chipset (typically the I/O Hub) is dedicated to causing an SCI whenever the notebook power source changes. This input is referred to as AC available (ACAV). The system can control ACAV directly based upon the presence of an AC adapter, or the embedded controller (EC) in the system can control ACAV. Whenever the AC adapter is inserted or removed from the system, the chipset logic associated with ACAV sets the status bit associated with ACAV, and if enabled causes an SCI to be asserted. Note: the control method associated with ACAV status could be implemented in the EC. A chipset GPE dedicated for AC adapter status is used for this description. The BIOS provides the following ACPI ASL control methods and objects: _PPC (Performance Present Capabilities) _PSR (Power Source) is described in Section 11.3.1 of the ACPI 2.0 specification. _PSR returns: 0x00000000 if the system is running on battery power. 0x00000001 if the system is running from an AC adapter. \_SB.AC is a BIOS-defined device object for the AC adapter. _L12 is the control method associated with GPE 12.
The flow associated with AC adapter insertion or removal is: AC adapter is inserted or removed The GPE Status bit associated with ACAV is asserted (for this example, GPE 12) An SCI (system control interrupt is issued) The OS/ACPI driver determines that GPE 12 was asserted and runs the associated control method (_L12) _L12 Issues a Notify(\_PR.CPU0, 0x80) which forces the OS to re-evaluate the _PPC object. _PPC reads the state of the ACAV pin and returns: 0 if the system is AC-powered. All P-states are available. n if the system is battery-powered, where n is the highest P-state is supported when battery-powered. Only P-states n and lower are available. For example, if a processor supports 3 P-states: P0 = 2400 MHz at 1.2 V P1 = 1600 MHz at 1.1 V P2 = 800 MHz at 1.0 V Power and Thermal Management 311
Chapter 10
32559
Rev. 3.08 July 2007
Then n = 2 limits the maximum P-state to 800 MHz at 1.0 V when the system is batterypowered. OS takes into account the available P-states and if necessary performs a P-state transition.
Issues a Notify(\_SB.AC, 0x80) where \_SB.AC is the AC adapter device object. OS runs _PSR to determine if the current power source is the AC adapter. PSR reads the state of the ACAV pin and returns: 1 if the system is AC-powered. 0 if the system is battery-powered.
Normal operation continues.
10.6.4
PSTATE_CNT
PSTATE_CNT is an entry in the Fixed ACPI Description Table (FADT). A PSTATE_CNT entry of 0 informs the operating system that the BIOS in System Management Mode (SMM) does not perform P-state transitions. The BIOS must report a value of 0 in the PSTATE_CNT field of the FADT. The only BIOS-controlled P-state transition, if any, must be performed near the beginning of the POST routine before control is passed to the operating system. All subsequent transitions are made by system software not the BIOS. System software is either the AMD PowerNow! technology software (for Microsoft operating systems prior to the Windows XP operating system) or the operating system and associated processor driver (for the Windows XP operating system and subsequent Microsoft operating systems).

Note: SMM is not used to perform P-state transitions.
10.6.5
CST_CNT
CST_CNT is an entry in the FADT. If non-zero, this field contains the value the operating system writes to the SMI_CMD register to indicate operating system support for the _CST object and C States Changed notification. The BIOS must report a value of 0 in the CST_CNT field of the FADT. AMD platforms only support ACPI 1.0b processor power state functionality. Processor power states are referred to as C states by the ACPI specifications.
10.7
ACPI 3.0 Processor P-State Objects
For systems that support P-state transitions, the BIOS must provide the following ACPI 3.0 P-state objects to support operating systems that provide native support for processor P-state transitions. The _PSD (P-State Dependency) object is generically described in section 8.4.4.5 of the ACPI 3.0b specification. The _PSD object is placed after the Processor object in the \_PR name space for
312
Chapter 10
32559
Rev. 3.08
July 2007
operating systems that have native support for processor P-states. This section provides BIOS specifics regarding the use of _PSD. The _PSD object provides P-state control information for logical processor dependencies to OSPM. For example, in a dual-core processor, each core may not operate at a unique voltage. With the _PSD object, a P-state transition on one processor core will cause OSPM to inititate a P-state transition to the same performance state on the remaining dependent core. The _PSD object should not be included for single-core processors. A _PSD object must be included for each dual-core processor in the system. Name (_PSD, Package(0x1) { // Field Name Field Type Package (0x5) { 0x5, // 5 entries 0x0, // Revision 0 Domain, // node or socket number, starting with zero 0xfd, // OSPM Coordinate, Initiate on any Proc 0x2, // 2 cores per node/socket } } ) // End of _PSD object Domain is the logical processor enumeration, starting with a value of zero, and corresponding with a processor node/socket.
10.8
BIOS Support for AMD PowerNow! Software with Legacy Operating Systems
If BIOS determines it has a processor and system that support P-state transitions, it provides a PSB that is comprised of a header and the processor specific Performance State Table (PST) that matches the processor in the system. The PSB must be 16-byte aligned in memory. If the BIOS cannot determine that the system and processor support P-state transitions as defined in 10.5.1 on page 287, then the BIOS must not provide a PSB. Changing the PSB Signature field to all 0s is an effective way of removing the PSB. Table 74 defines the PSB and PST structures. Table 74. Performance State Block Structure
Name PSB Header Signature 10 Start of the PSB - Always set this field to AMDK7PNOW! Length (Bytes) Field Purpose
Chapter 10
313
32559
Rev. 3.08 July 2007
Table 74.
Name TableVersion
Performance State Block Structure (Continued)

Length (Bytes) 1 Field Purpose The version of the table as defined by AMD. Set to 14h for version 1.4. If AMD PowerNow! software does not recognize the TableVersion, it does not attempt to make P-state transitions. All bits are reserved and written to zero. Voltage Stabilization Time. The amount of time required for the processors core voltage regulator to increase the processors core voltage the increment specified by MVS. VST is specified in 20 s increments. The default is 05h (100 s). All other values are reserved. Bits 10: Ramp Voltage Offset (RVO) has the same definition as for the _PSS ACPI object. Bits 32: Isochronous Relief Time (IRT) has the same definition as for the _PSS ACPI object. Bits 54: Maximum Voltage Step (MVS) has the same definition as for the _PSS ACPI object. Bits 76: Number of available P-states when the system is battery powered has the following definition: 00b = all P-states are available 01b = only the minimum P-state is available 10b = 2 lowest P-states are available 11b = 3 lowest P-states are available.
Flags VST
1 2
Reserved1
NumPST PST Header CPUID
1 4
The total number of PSTs in the PSB. This value must be set to 01h. CPUID (EAX of CPUID extended function 8000_0001h) of the processor that this PST belongs to. This provides processor family, extended family, model, extended model, and stepping (revision) fields. Processor PLL Lock time in microseconds. For example, PLL_LOCK_TIME of 2 s is represented as 00000010b.The PLL Lock time is specified in section 14.2.9.1 on page 411. MaxFID[5:0] as reported by the MaxFID field of the FIDVID_STATUS MSR. MaxVID[5:0] as reported by the MaxVID field of the FIDVID_STATUS MSR. The number of FID, VID combinations in the PST. This field must be >= 1h. FID[5:0] code corresponding to the frequency of the lowest performance state that the processor supports. This is defined in the processor data sheet. VID[5:0] code corresponding to the voltage of the lowest performance state that the processor supports. This is defined in the processor data sheet. FID, VID pairs are concatenated here so that the total number of pairs is equal to the NumPstates field. Each subsequent FID, VID pair is appended in ascending order (i.e., the last FID,VID pair corresponds to the maximum performances state supported by the processor).
PLL Lock Time
MaxFID MaxVID NumPStates FID VID -----------
1 1 1 1 1 ---------
314
Chapter 10
32559
Rev. 3.08
July 2007
10.9
10.9.1

System Configuration for Power Management

Chipset Configuration for Power Management
BIOS configuration of the system for power management is described in the sections that follow.
Chipset configuration is covered in chipset related documentation. The BIOS is required to: Ensure that the SMAF code to STPCLK and Stop Grant mapping for all components based on HyperTransport technology is configured as described in Table 65 on page 283. For systems using AM2 or S1g1 package processors, the HyperTransport links are tristated in response to LDTSTOP_L assertion. Chipset power management features are appropriately enabled based upon system class. LDTSTOP_L assertion for FID changes during P-state transitions: Must be at least 2s. Note: for UMA chipsets 2s is ideal and may be required in some cases. Note: Links need to be 8-bits wide or wider and 400 MHz or greater. Must be less than 6s for chipsets that support isochronous streams (e.g. UMA graphics). LDTSTOP_L assertion for link width/frequency changes during initialization is a minimum of 2 s for all processors. Note: For Links operating at 200 MHz no DMA traffic can be in progress. LDTSTOP_L assertion during FID_Change is 10 s for systems using registered DIMMs. LDTSTOP_L assertion during C3 is a minimum of 1 s for all processors.
10.9.2
Processor Configuration for Power Management
BIOS configures the processor to enable power management during the POST routine. Table 75 on page 316 lists processor memory system configuration register settings that the BIOS must initialize.
Chapter 10
315
32559
Rev. 3.08 July 2007
Table 75.
Configuration Register Settings for Power Management

Register LDT0 Link Control (function 0: offset 84h[13]) (LdtStopTriEn) LDT1 Link Control (function 0: offset A4h[13]) (LdtStopTriEn) LDT2 Link Control (function 0: offset C4h[13]) (LdtStopTriEn) Mobile Setting 1b UP Server / Desktop Setting
1 1
Functionality Enable HyperTransport links to be tristated in response to LDTSTOP_L assertions.
MP Server Setting 0b
0b
Enable HyperTransport Receivers to be tristated in response to LDTSTOP_L assertion in mobile systems.
LDT0 Link Control (function 0: offset 84h[11]) (LdtStopRcvDis) LDT1 Link Control (function 0: offset A4h[11]) (LdtStopRcvDis) LDT2 Link Control (function 0: offset C4h[11]) (LdtStopRcvDis)
1b
0b
0b
Do not enable overriding tristating of HyperTransport clocks in response to LDTSTOP_L assertion except for mobile systems with UMA Graphics.
LDT0 Link Control (function 0: offset 84h[15]) (LdtStopTriClkOvr) LDT1 Link Control (function 0: offset A4h[15]) (LdtStopTriClkOvr) LDT2 Link Control (function 0: offset C4h[15]) (LdtStopTriClkOvr)
Discrete graphics 0b UMA graphics 1b
0b
0b
Notes:
1. Single processor branded server with unbuffered DIMMs only, or any dekstop brand (with unbuffered DIMMs). 2. Optionally, a value of 2113_2113h can be used. This significantly reduces processor power during halt and may reduce system performance. To ensure that system performance loss is kept to a minimum in this state, ClkRampHys (Function 3, Offset D4h) should be programmed to 2 us (0111b.) 3. For mobile systems that enable AltVID, VSTIME should be programmed to account for the time it takes to ramp the voltage from AltVID to the VID for the 800MHz P-State. VSTIME = ((800 MHz VID) - AltVID) * 2us/mV 4. Dual-Core Revision G processors in AM2 package with unbuffered DIMMs and SMI initiated C1E only. All others use Revision F SMAF 0 and 1 values. Do not enable C1E with SO-DIMMs. Do not enable AltVID with these processors. If chipset does not support SMI inititated C1E then use Revision F SMAF 0 and 1 values.
316
Chapter 10
32559
Rev. 3.08
July 2007
Table 75.
Configuration Register Settings for Power Management (Continued)

Register Power Management Control Low (function 3: offset 80h) Mobile Setting Discrete graphics 6B07_6B61h UMA graphics 3307_3331 UP Server / Desktop Setting
1 1
Functionality Map STPCLK actions to SMAF codes: SMAF 0 = C2 SMAF 1 = C3 SMAF 2 = VID/FID Change SMAF 3 = S1
MP Server Setting 2307_0000h
Revision F processors 2307_0000h Revision G processors with Discrete graphics 2307_6361h

4
Revision G processors with UMA graphics 2307_3331h

4
Map STPCLK actions to SMAF codes: SMAF 4 = S3 SMAF 5 = Throttling SMAF 6 = S4/S5 SMAF 7 is not used for STPCLK, the corresponding field is used for HALT. BCLK and LCLK PLL frequency lock Sets Ramp VID (0 mV) and voltage regulator stabilization time (0 s). The processor driver performs this function, not hardware.
3
Power Management Control High (function 3: offset 84h)
Discrete graphics 6113_3113h UMA graphics 3113_3113h
2113_2113h
0013_2113h
Clock Power/Timing Low (function 3: offset D4h) Clock Power/Timing High (function 3:offset D8h [30:28], [19:0])
(RampVIDOf) (VSTime)
000D_A701h 000b[30:28] 000b[30:28] 0_0000h[19:0] 0_0000h[19:0]
04E2_A706h 010b[30:28] 0_2710h[19:0]
Notes:
Chapter 10
317
32559
Rev. 3.08 July 2007
Table 75.
Configuration Register Settings for Power Management (Continued)

Register Clock Power/Timing High (function 3:offset D8h [24:20]) (AltVid) Mobile Setting Discrete graphics P[Min] Voltage AltVidOffset UMA graphics N/A UP Server / Desktop Setting
1 1
Functionality Set the AltVID value for mobile systems with discrete graphics.
MP Server Setting N/A
N/A
Enable PSIVID for mobile systems. Set the PSIVID value for mobile systems. Do not enable tristating of memory clock except systems with discrete graphics.
Power Control Miscellaneous (function 3: offset A0h [7]) (PsiVidEn) Power Control Miscellaneous (function 3: offset A0h [5:0]) (PsiVid) DRAM Control (function 2: offset 78h[16])
1b
0b
0b
P[Min] VID
N/A
N/A
UMA graphics 0b discrete graphics 1b
0b
0b
Enable DRAM power down mode. Enable DRAM controller power savings in C3.
DRAM Configuration Low (function 2: offset 94h[15]) DRAM Controller Miscellaneous (function 2: offset A0h [10]) (PwrSavingsEn)
1b 1b
1b 0b
1b 0b
Notes:
10.9.2.1
Additional BIOS Support Requirements for S3
During the S3 state, power is removed from the processors CPU core, host bridge, and memory controller. It is required that the BIOS restore the state of the processors memory controller upon resume from S3 prior to having the memory controller bring system memory out of self-refresh mode. The system must provide non volatile storage for the memory controller configuration registers 318 Power and Thermal Management Chapter 10
32559
Rev. 3.08
July 2007
during the S3 state. For UP and DP systems, this storage is provided in the chipset. There are two opportunities for the BIOS to store the memory controller configuration to this nonvolatile memory. During POST as the memory controller is being initially configuredthis is preferred, since it does not require SMM support. In SMM just prior to entry to the S3 stateBIOS uses an SMI trap on writes to the ACPI PM1 control register to get control of the system just prior to entry to S3. After storing the memory controller configuration registers, I/O instruction restart is used to place the system into the S3 state. This is not preferred because it relies on SMM.
The following processor registers must be stored in nonvolatile RAM for support of resume from S3: Function 1, offsets 40h7Ch, F0h Function 2, offsets 40h6Ch, 78h94h, A0h Function 2, offset 98h indexes 0h7h, 10h, 13h, 16h, 19h, 20h27h, 30h, 33h, 36h, 39h Function 3, offset 44h and offset B0h if online spare is enabled MSRs 1Bh, 200h20Fh, 250h, 258h, 259h, 268h-26Fh, 2FFh, C001_0015hC001_001Ah, C001_001Dh, C001_001Fh, C001_0030hC001_0035h, C001_0048h, C001_0055h, C001_1004h, C001_1005h, C001_100Dh 10.9.2.2 ACPI Tables/Objects
FADT entries for C-state and throttling support and throttling is listed in Table 76 on page 319. Some legacy operating systems do not use C3 if support for C2 is not declared. Table 76.
Field PSTATE_CNT CST_CNT P_LVL2_LAT P_LVL3_LAT DUTY_WIDTH
FADT Table Entries

Mobile 0 0 > 100 if C2 is not supported 1 if C2 is supported > 1000 if C3 is not supported 10 if C3 is supported 0 if _PSV is not used 3 if _PSV is used UP Desktop 0 0 > 100 > 1000 0 if _PSV is not used 3 if _PSV is used Multiprocessor 0 0 > 100 > 1000 0
Chapter 10
319
32559
Rev. 3.08 July 2007
320
Chapter 10
32559
Rev. 3.08
July 2007
11
Performance Monitoring
AMD NPT Family 0Fh Processors support the performance-monitoring features introduced in earlier processor implementations. These features allow the selection of events to be monitored, and include a set of corresponding counter registers that track the frequency of monitored events. Software tools can use these features to identify performance bottlenecks, such as sections of code that have high cache-miss rates or frequently mis-predicted branches. This information can then be used as a guide for improving or eliminating performance problems through software optimizations or hardwaredesign improvements. For a general description of how to use these performance monitoring features, refer to the Debug and Performance Resources section in Volume 2: System Programming of the AMD64 Architecture Programmers Manual, order# 24593.
11.1
Performance Counters
The processor provides four 48-bit performance counters. Each counter can monitor a different event. The accuracy of the counters is not ensured. Some events are reserved. When a reserved event is selected, the results are undefined. Performance counters are used to count specific processor events, such as data-cache misses, or the duration of events, such as the number of clocks it takes to return data from memory after a cache miss. During event counting, the processor increments the counter when it detects an occurrence of the event. During duration measurement, the processor counts the number of processor clocks it takes to complete an event. Each performance counter can be used to count one event, or measure the duration of one event at a time. In addition to RDMSR instruction, the PerfCtrn registers can be read using a special read performance-monitoring counter instruction, RDPMC. The RDPMC instruction loads the contents of the PerfCtrn register specified by the ECX register, into the EDX register and the EAX register. Writing the performance counters can be useful if software wants to count a specific number of events, and then trigger an interrupt when that count is reached. An interrupt can be triggered when a performance counter overflows. Software should use the WRMSR instruction to load the count as a twos-complement negative number into the performance counter. This causes the counter to overflow after counting the appropriate number of times. The performance counters are not assured of producing identical measurements each time they are used to measure a particular instruction sequence, and they should not be used to take measurements of very small instruction sequences. The RDPMC instruction is not serializing, and it can be executed out-of-order with respect to other instructions around it. Even when bound by serializing instructions, the system environment at the time the instruction is executed can cause events to be counted before the counter value is loaded into EDX:EAX. Chapter 11 Performance Monitoring 321
32559
Rev. 3.08 July 2007
PerfCtr0PerfCtr3 Registers
63 reserved 31
C001_0004h, C001_0005h, C001_0006h, C001_0007h

48 47 CTR (4732) 0 CTR (310) 32
Bit 6348 470
Name Reserved CTR
Function RAZ Counter value
R/W R R
11.2
Performance Event-Select Registers
Performance Event-Select registers (PerfEvtSeli) are used to specify the events counted by the performance counters, and to control other aspects of their operation. Each performance counter supported by the implementation has a corresponding event-select register that controls its operation. This section shows the format of the PerfEvtSeln registers. To accurately start counting with the write that enables the counter, disable the counter when changing the event and then enable the counter with a second MSR write. The edge count mode increments the counter when a transition happens on the monitored event. If the event selected is changed without disabling the counter, an extra edge is falsely detected when the first event is a static 0 and the second event is a static one. To avoid this false edge detection, disable the counter when changing the event and then enable the counter with a second MSR write. The EVENT_SELECT field specifies the event or event duration in a processor unit to be counted by the corresponding PerfCtri register. The UNIT_MASK field is used to further specify or qualify a monitored event. Section 11.2.1 Performance Monitor Events shows the events and unit masks supported by the processor. The performance counter registers can be used to track events in the Northbridge. Northbridge events include all memory controller events, crossbar events, and HyperTransport interface events as documented in 11.2.1.7, 11.2.1.8, and 11.2.1.9. When a Northbridge event is selected using one of the Performance Event-Select registers in either core of a dual core processor, a Northbridge event cannot be selected in the same Performance Event-Select register of the other core. Monitoring of Northbridge events should only be performed by one core in a dual core processor. Care must be taken when measuring Northbridge or other non-processor-specific events under conditions where the processor may go into halt mode during the measurement period. For instance, one may wish to monitor DRAM traffic due to DMA activity from a disk or graphics adaptor. This entails running some event counter monitoring code on the processor, where such code accesses the counters at the beginning and end of the measurement period, or may even sample them periodically 322 Performance Monitoring Chapter 11
32559
Rev. 3.08
July 2007
throughout the measurement period. Such code typically gives up the processor during each measurement interval. If there is nothing else for the OS to run on that particular processor at that time, it may halt the processor until it is needed. Under these circumstances, the clock for the counter logic will be stopped, hence the counters will not count the events of interest. To prevent this, simply run a low-priority background process that will keep the processor busy during the period of interest. PerfEvtSel0PerfEvtSel3 Registers
63 reserved 31 CNT_MASK
)
C001_0000h, C001_0001h, C001_0002h, C001_0003h

32
24 23 22 21 20 19 18 17 16 15 reserved USR INV INT OS EN PC UNIT_MASK E
7 EVENT_SELECT
Bit 6332 3124 23 22 21 20 19 18 17 16 158 70
Name reserved CNT_MASK INV EN reserved INT PC E OS USR UNIT_MASK EVENT_MASK
Function RAZ Counter Mask Invert Counter Mask Enable Counter SBZ Enable APIC Interrupt Pin Control Edge Detect Operating-System Mode User Mode Event Qualification Unit and Event Selection
Field Descriptions Event Selection (EVENT_SELECT)Bits 70. The EVENT_SELECT field specifies the event or event duration in a processor unit to be counted by the corresponding PerfCtri register. Event Qualification (UNIT_MASK)Bits 158. Each UNIT_MASK bit further specifies or qualifies the event specified by EVENT_MASK. All events selected by UNIT_MASK are simultaneously monitored. Unless otherwise stated, the UNIT_MASK values shown may be combined (logically ORed) to select any desired combination of the sub-events for a given event. In some cases, certain combinations can result in misleading counts, or the UNIT_MASK value is an ordinal rather than a bit mask. These situations are described where applicable, or should be obvious from the event descriptions. For events where no UNIT_MASK table is shown, the UNIT_MASK is not applicable and may be set to zeros. Chapter 11 Performance Monitoring 323
32559
Rev. 3.08 July 2007
User Mode (USR)Bit 16. Count events in user mode (at CPL > 0). Operating-System Mode (OS)Bit 17. Count events in operating-system mode (at CPL = 0). Edge Detect (E)Bit 18. 1=Edge detection. 0=Level detection. Pin Control (PC)Bit 19. With a value of 1, toggles the performance monitor pin PMi when PerfCtri register overflows. With a value of 0, toggles the performance monitor pin PMi each time PerfCtri register is incremented. Enable APIC Interrupt (INT)Bit 20. Enables APIC interrupt when PerfCtri register overflows. Enable Counter (EN)Bit 22. Enables performance monitor counter. Invert Counter Mask (INV)Bit 23. See CNT_MASK. Counter Mask (CNT_MASK)Bits 3124. a) When CNT_MASK is 0, the corresponding PerfCtri register is incremented by the number of events occurring in a clock cycle. Maximum number of events in one cycle is 3. b) When CNT_MASK values are from 1 to 3 and INV is 0, the corresponding PerfCtri register is incremented by 1, if the number of events occurring in a clock cycle is greater than or equal to CNT_MASK. When CNT_MASK values are from 1 to 3 and INV is 1, the corresponding PerfCtri register is incremented by 1, if the number of events occurring in a clock cycle is less than CNT_MASK. c) CNT_MASK values from 4 to 255 are reserved.
11.2.1
Performance Monitor Events
Note: Speculative vs. Retired: Several events may include speculative activity, meaning they may be associated with false-path instructions that are ultimately discarded due to a branch misprediction. Events associated with Retire reflect actual program execution. For events where the distinction may matter, these are explicitly labeled as one or the other. 11.2.1.1 Floating Point Unit Events
Event Select: 00h - Dispatched FPU Operations The number of operations (uops) dispatched to the FPU execution pipelines. This event reflects how busy the FPU pipelines are. This includes all operations done by x87, MMX and SSE instructions, including moves. Each increment represents a one-cycle dispatch event; packed 128-bit SSE operations count as two ops; scalar operations count as one. Speculative. (See also event CBh). Note: Since this event includes non-numeric operations it is not suitable for measuring MFLOPs.
UNIT_MASK 01h 02h Add pipe ops Multiply pipe ops
324
Chapter 11
32559
Rev. 3.08
July 2007
UNIT_MASK 04h 08h 10h 20h Store pipe ops Add pipe load ops Multiply pipe load ops Store pipe load ops
Event Select: 01h - Cycles with no FPU Ops Retired The number of cycles in which no FPU operations were retired. Invert this (set the Invert control bit in the event select MSR) to count cycles in which at least one FPU operation was retired. Event Select: 02h - Dispatched Fast Flag FPU Operations The number of FPU operations that use the fast flag interface (e.g. FCOMI, COMISS, COMISD, UCOMISS, UCOMISD). Speculative. 11.2.1.2 Load/Store Unit and TLB Events
Event Select: 20h - Segment Register Loads The number of segment register loads performed.
UNIT_MASK 01h 02h 04h 08h 10h 20h 40h ES CS SS DS FS GS HS
Event Select: 21h - Pipeline restart due to self-modifying code The number of pipeline restarts that were caused by self-modifying code (a store that hits any instruction that's been fetched for execution beyond the instruction doing the store). Event Select: 22h - Pipeline restart due to probe hit The number of pipeline restarts caused by an invalidating probe hitting on a speculative out-of-order load.
Chapter 11
325
32559
Rev. 3.08 July 2007
Event Select: 23h - LS Buffer 2 Full The number of cycles that the LS2 buffer is full. This buffer holds stores waiting to retire as well as requests that missed the data cache and are waiting on a refill. This condition will stall further data cache accesses, although such stalls may be overlapped by independent instruction execution. Event Select: 24h - Locked Operations This event covers locked operations performed and their execution time. The execution time represented by the cycle counts is typically overlapped to a large extent with other instructions. The non-speculative cycles event is suitable for event-based profiling of lock operations that tend to miss in the cache.
UNIT_MASK 01h 02h 04h The number of locked instructions executed The number of cycles spent in speculative phase The number of cycles spent in non-speculative phase (including cache miss penalty)
Event Select: 65h - Memory Requests by Type These events reflect accesses to uncachable (UC) or write-combining (WC) memory regions (as defined by MTRR or PAT settings) and Streaming Store activity to WB memory. Both the WC and Streaming Store events reflect Write Combining buffer flushes, not individual store instructions. WC buffer flushes which typically consist of one 64-byte write to the system for each flush (assuming software typically fills a buffer before it gets flushed). A partially-filled buffer will require two or more smaller writes to the system. The WC event reflects flushes of WC buffers that were filled by stores to WC memory or streaming stores to WB memory. The Streaming Store event reflects only flushes due to streaming stores (which are typically only to WB memory). The difference between counts of these two events reflects the true amount of write events to WC memory.
UNIT_MASK 01h 02h 80h Requests to non-cacheable (UC) memory Requests to write-combining (WC) memory or WC buffer flushes to WB memory Streaming store (SS) requests
11.2.1.3
Data Cache Events
Event Select: 40h - Data Cache Accesses The number of accesses to the data cache for load and store references. This may include certain microcode scratchpad accesses, although these are generally rare. Each increment represents an eightbyte access, although the instruction may only be accessing a portion of that. Speculative.
326
Chapter 11
32559
Rev. 3.08
July 2007
Event Select: 41h - Data Cache Misses The number of data cache references which missed in the data cache. Speculative. Except in the case of streaming stores, only the first miss for a given line is included - access attempts by other instructions while the refill is still pending are not included in this event. So in the absence of streaming stores, each event reflects one 64-byte cache line refill, and counts of this event are the same as, or very close to, the combined count for event 42h. Streaming stores however will cause this event for every such store, since the target memory is not refilled into the cache. Hence this event should not be used as an indication of data cache refill activity - event 42h should be used for such measurements. (See event 65h for an indication of streaming store activity.) A large difference between events 41h (with all UNIT_MASK bits set) and 42h would be due mainly to streaming store activity. Event Select: 42h - Data Cache Refills from L2 or System The number of data cache refills satisfied from the L2 cache (and/or the system), per the UNIT_MASK. UNIT_MASK bits 4:1 allow a breakdown of refills from the L2 by coherency state. UNIT_MASK bit 0 reflects refills which missed in the L2, and provides the same measure as the combined sub-events of event 43h. Each increment reflects a 64-byte transfer. Speculative.
UNIT_MASK 01h 02h 04h 08h 10h Refill from System Shared-state line from L2 Exclusive-state line from L2 Owned-state line from L2 Modified-state line from L2
Event Select: 43h - Data Cache Refills from System The number of L1 cache refills satisfied from the system (system memory or another cache), as opposed to the L2. The UNIT_MASK selects lines in one or more specific coherency states. Each increment reflects a 64-byte transfer. Speculative.
UNIT_MASK 01h 02h 04h 08h 10h Invalid Shared Exclusive Owned Modified
Chapter 11
327
32559
Rev. 3.08 July 2007
Event Select: 44h - Data Cache Lines Evicted The number of L1 data cache lines written to the L2 cache or system memory, having been displaced by L1 refills. The UNIT_MASK may be used to count only victims in specific coherency states. Each increment represents a 64-byte transfer. Speculative. In most cases, L1 victims are moved to the L2 cache, displacing an older cache line there. Lines brought into the data cache by PrefetchNTA instructions, however, are evicted directly to system memory (if dirty) or invalidated (if clean). There is no provision for measuring this component by itself. The Invalid case (UNIT_MASK value 01h) reflects the replacement of lines that would have been invalidated by probes for write operations from another processor or DMA activity.
UNIT_MASK 01h 02h 04h 08h 10h Invalid Shared Exclusive Owned Modified
Event Select: 45h - L1 DTLB Miss and L2 DTLB Hit The number of data cache accesses that miss in the L1 DTLB and hit in the L2 DTLB. Speculative. Event Select: 46h - L1 DTLB and L2 DTLB Miss The number of data cache accesses that miss in both the L1 and L2 DTLBs. Speculative. Event Select: 47h - Misaligned Accesses The number of data cache accesses that are misaligned. These are accesses which cross an eight-byte boundary. They incur an extra cache access (reflected in event 40h), and an extra cycle of latency on reads. Speculative. Event Select: 48h - Microarchitectural Late Cancel of an Access Event Select: 49h - Microarchitectural Early Cancel of an Access Event Select: 4Ah - Single-bit ECC Errors Recorded by Scrubber The number of single-bit errors corrected by either of the error detection/correction mechanisms in the data cache.
UNIT_MASK 01h 02h Scrubber error Piggyback scrubber errors
328
Chapter 11
32559
Rev. 3.08
July 2007
Event Select: 4Bh - Prefetch Instructions Dispatched The number of prefetch instructions dispatched by the decoder. Speculative. Such instructions may or may not cause a cache line transfer. Any Dcache and L2 accesses, hits and misses by prefetch instructions are included in these types of events.
UNIT_MASK 01h 02h 04h Load (Prefetch, PrefetchT0/T1/T2) Store (PrefetchW) NTA (PrefetchNTA)
Event Select: 4Ch - DCACHE Misses by Locked Instructions The number of data cache misses incurred by locked instructions. (The total number of locked instructions may be obtained from event 24h.) Such misses may be satisfied from the L2 or system memory, but there is no provision for distinguishing between the two. When used for event-based profiling, this event will tend to occur very close to the offending instructions. (See also event 24h.) This event is also included in the basic Dcache miss event (41h).
UNIT_MASK 02h Data cache misses by locked instructions
11.2.1.4
L2 Cache and System Interface Events
Event Select: 67h - Data Prefetcher These events reflect requests made by the data prefetcher. UNIT_MASK bit 1 counts total prefetch requests, while bit 0 counts requests where the target block is found in the L2 or data cache. The difference between the two represents actual data read (in units of 64-byte cache lines) from the system by the prefetcher. This is also included in the count of event 7Fh, UNIT_MASK bit 0 (combined with other L2 fill events).
UNIT_MASK 01h 02h Cancelled prefetches Prefetch attempts
Event Select: 6Ch - System Read Responses by Coherency State The number of responses from the system for cache refill requests. The UNIT_MASK may be used to select specific cache coherency states. Each increment represents one 64-byte cache line transferred from the system (DRAM or another cache, including another core on the same node) to the data cache, instruction cache or L2 cache (for data prefetcher and TLB table walks). Modified-state Chapter 11 Performance Monitoring 329
32559
Rev. 3.08 July 2007
responses may be for Dcache store miss refills, PrefetchW software prefetches, hardware prefetches for a store-miss stream, or Change-to-Dirty requests that get a dirty (Owned) probe hit in another cache. Exclusive responses may be for any Icache refill, Dcache load miss refill, other software prefetches, hardware prefetches for a load-miss stream, or TLB table walks that miss in the L2 cache; Shared responses may be for any of those that hit a clean line in another cache.
UNIT_MASK 01h 02h 04h Exclusive Modified Shared
Event Select: 6Dh - Quadwords Written to System The number of quadword (8-byte) data transfers from the processor to the system. These may be part of a 64-byte cache line writeback or a 64-byte dirty probe hit response, each of which would cause eight increments; or a partial or complete Write Combining buffer flush (Sized Write), which could cause from one to eight increments.
UNIT_MASK 01h Quadword write transfer
Event Select: 7Dh - Requests to L2 Cache The number of requests to the L2 cache for Icache or Dcache fills, or page table lookups for the TLB. These events reflect only read requests to the L2; writes to the L2 are indicated by event 7Fh. These include some amount of retries associated with address or resource conflicts. Such retries tend to occur more as the L2 gets busier, and in certain extreme cases (such as large block moves that overflow the L2) these extra requests can dominate the event count. These extra requests are not a direct indication of performance impact - they simply reflect opportunistic accesses that don't complete. But because of this, they are not a good indication of actual cache line movement. The Icache and Dcache miss and refill events (81h, 82h, 83h, 41h, 42h, 43h) provide a more accurate indication of this, and are the preferred way to measure such traffic.
UNIT_MASK 01h 02h 04h 08h 10h IC fill DC fill TLB fill (page table walks) Tag snoop request Cancelled request
330
Chapter 11
32559
Rev. 3.08
July 2007
Event Select: 7Eh - L2 Cache Misses The number of requests that miss in the L2 cache. This may include some amount of speculative activity, as well as some amount of retried requests as described in event 7Dh. The IC-fill-miss and DC-fill-miss events tend to mirror the Icache and Dcache refill-from-system events (83h and 43h, respectively), and tend to include more speculative activity than those events.
UNIT_MASK 01h 02h 04h IC fill DC fill (includes possible replays, whereas event 41h does not) TLB page table walk
Event Select: 7Fh - L2 Fill/Writeback The number of lines written into the L2 cache due to victim writebacks from the Icache or Dcache, TLB page table walks and the hardware data prefetcher (UNIT_MASK bit 0); or writebacks of dirty lines from the L2 to the system (UNIT_MASK bit 1). Each increment represents a 64-byte cache line transfer. Note: Victim writebacks from the Dcache may be measured separately using event 44h. However this is not quite the same as the Dcache component of event 7Fh, the main difference being PrefetchNTA lines. When these are evicted from the Dcache due to replacement, they are written out to system memory (if dirty) or simply invalidated (if clean), rather than being moved to the L2 cache.
UNIT_MASK 01h 02h L2 fills (victims from L1 caches, TLB page table walks and data prefetches) L2 Writebacks to system.
11.2.1.5
Instruction Cache Events
All instruction cache events are speculative. Event Select: 80h - Instruction Cache Fetches The number of instruction cache accesses by the instruction fetcher. Each access is an aligned 16 byte read, from which a varying number of instructions may be decoded. Event Select: 81h - Instruction Cache Misses The number of instruction fetches that miss in the instruction cache. This is typically equal to or very close to the sum of events 82h and 83h. Each miss results in a 64-byte cache line refill.
Chapter 11
331
32559
Rev. 3.08 July 2007
Event Select: 82h - Instruction Cache Refills from L2 The number of instruction cache refills satisfied from the L2 cache. Each increment represents one 64-byte cache line transfer. Event Select: 83h - Instruction Cache Refills from System The number of instruction cache refills from system memory (or another cache). Each increment represents one 64-byte cache line transfer. Event Select: 84h - L1 ITLB Miss, L2 ITLB Hit The number of instruction fetches that miss in the L1 ITLB but hit in the L2 ITLB. Event Select: 85h - L1 ITLB Miss, L2 ITLB Miss The number of instruction fetches that miss in both the L1 and L2 ITLBs. Event Select: 86h - Pipeline Restart Due to Instruction Stream Probe The number of pipeline restarts caused by invalidating probes that hit on the instruction stream currently being executed. This would happen if the active instruction stream was being modified by another processor in an MP system - typically a highly unlikely event. Event Select: 87h - Instruction Fetch Stall The number of cycles the instruction fetcher is stalled. This may be for a variety of reasons such as branch predictor updates, unconditional branch bubbles, far jumps and cache misses, among others. May be overlapped by instruction dispatch stalls or instruction execution, such that these stalls don't necessarily impact performance. Event Select: 88h - Return Stack Hits The number of near return instructions (RET or RET Iw) that get their return address from the return address stack (i.e. where the stack has not gone empty). This may include cases where the address is incorrect (return mispredicts). This may also include speculatively executed false-path returns. Return mispredicts are typically caused by the return address stack underflowing, however they may also be caused by an imbalance in calls vs. returns, such as doing a call but then popping the return address off the stack. Note: This event cannot be reliably compared with events C9h and CAh (such as to calculate percentage of return mispredicts due to an empty return address stack), since it may include speculatively executed false-path returns that are not included in those retire-time events. Event Select: 89h - Return Stack Overflows The number of (near) call instructions that cause the return address stack to overflow. When this happens, the oldest entry is discarded. This count may include speculatively executed calls. 332 Performance Monitoring Chapter 11
32559
Rev. 3.08
July 2007
11.2.1.6
Execution Unit Events
Event Select: 26h - Retired CLFLUSH Instructions The number of CLFLUSH instructions retired. Event Select: 27h - Retired CPUID Instructions The number of CPUID instructions retired. Event Select: 76h - CPU Clocks not Halted The number of clocks that the CPU is not in a halted state (due to STPCLK or a HALT instruction). Note: this event allows system idle time to be automatically factored out from IPC (or CPI) measurements, providing the OS halts the CPU when going idle. If the OS goes into an idle loop rather than halting, such calculations will be influenced by the IPC of the idle loop. Event Select: C0h - Retired Instructions The number of instructions retired (execution completed and architectural state updated). This count includes exceptions and interrupts - each exception or interrupt is counted as one instruction. Event Select: C1h - Retired uops The number of micro-ops retired. This includes all processor activity (instructions, exceptions, interrupts, microcode assists, etc.). Event Select: C2h - Retired Branch Instructions The number of branch instructions retired. This includes all types of architectural control flow changes, including exceptions and interrupts. Event Select: C3h - Retired Mispredicted Branch Instructions The number of branch instructions retired, of any type, that were not correctly predicted. This includes
those for which prediction is not attempted (far control transfers, exceptions and interrupts).
Event Select: C4h - Retired Taken Branch Instructions The number of taken branches that were retired. This includes all types of architectural control flow changes, including exceptions and interrupts. Event Select: C5h - Retired Taken Branch Instructions Mispredicted The number of retired taken branch instructions that were mispredicted.
Chapter 11
333
32559
Rev. 3.08 July 2007
Event Select: C6h - Retired Far Control Transfers The number of far control transfers retired including far call/jump/return, IRET, SYSCALL and SYSRET, plus exceptions and interrupts. Far control transfers are not subject to branch prediction. Event Select: C7h - Retired Branch Resyncs The number of resync branches. These reflect pipeline restarts due to certain microcode assists and events such as writes to the active instruction stream, among other things. Each occurrence reflects a restart penalty similar to a branch mispredict. Relatively rare. Event Select: C8h - Retired Near Returns The number of near return instructions (RET or RET Iw) retired. Event Select: C9h - Retired Near Returns Mispredicted The number of near returns retired that were not correctly predicted by the return address predictor. Each such mispredict incurs the same penalty as a mispredicted conditional branch instruction. Event Select: CAh - Retired Indirect Branches Mispredicted The number of indirect branch instructions retired where the target address was not correctly predicted. Event Select: CBh - Retired MMX/FP Instructions The number of MMX, SSE or X87 instructions retired. The UNIT_MASK allows the selection of the individual classes of instructions as given in the table. Each increment represents one complete instruction. Note: Since this event includes non-numeric instructions it is not suitable for measuring MFLOPS.
UNIT_MASK 01h 02h 04h 08h x87 instructions MMX and 3DNow! instructions Packed SSE and SSE2 instructions Scalar SSE and SSE2 instructions
334
Chapter 11
32559
Rev. 3.08
July 2007
Event Select: CCh - Retired Fastpath Double Op Instructions

UNIT_MASK 01h 02h 04h With low op in position 0 With low op in position 1 With low op in position 2
Event Select: CDh - Interrupts-Masked Cycles The number of processor cycles where interrupts are masked (EFLAGS.IF = 0). Using edge-counting with this event will give the number of times IF is cleared; dividing the cycle-count value by this value gives the average length of time that interrupts are disabled on each instance. Compare the edge count with event CFh to determine how often interrupts are disabled for interrupt handling vs. other reasons (e.g. critical sections). Event Select: CEh - Interrupts-Masked Cycles with Interrupt Pending The number of processor cycles where interrupts are masked (EFLAGS.IF = 0) and an interrupt is pending. Using edge-counting with this event and comparing the resulting count with the edge count for event CDh gives the proportion of interrupts for which handling is delayed due to prior interrupts being serviced, critical sections, etc. The cycle count value gives the total amount of time for such delays. The cycle count divided by the edge count gives the average length of each such delay. Event Select: CFh - Interrupts Taken The number of hardware interrupts taken. This does not include software interrupts (INT n instruction). Event Select: D0h - Decoder Empty The number of processor cycles where the decoder has nothing to dispatch (typically waiting on an instruction fetch that missed the Icache, or for the target fetch after a branch mispredict). Event Select: D1h - Dispatch Stalls The number of processor cycles where the decoder is stalled for any reason (has one or more instructions ready but can't dispatch them due to resource limitations in execution). This is the combined effect of events D2h - DAh, some of which may overlap; this event reflects the net stall cycles. The more common stall conditions (events D5h, D6h, D7h, D8h, and to a lesser extent D2) may overlap considerably. The occurrence of these stalls is highly dependent on the nature of the code being executed (instruction mix, memory reference patterns, etc.).
Chapter 11
335
32559
Rev. 3.08 July 2007
Event Select: D2h - Dispatch Stall for Branch Abort to Retire The number of processor cycles the decoder is stalled waiting for the pipe to drain after a mispredicted branch. This stall occurs if the corrected target instruction reaches the dispatch stage before the pipe has emptied. See also event D1h. Event Select: D3h - Dispatch Stall for Serialization The number of processor cycles the decoder is stalled due to a serializing operation, which waits for the execution pipeline to drain. Relatively rare; mainly associated with system instructions. See also event D1h. Event Select: D4h - Dispatch Stall for Segment Load The number of processor cycles the decoder is stalled due to a segment load instruction being encountered while execution of a previous segment load operation is still pending. Relatively rare except in 16-bit code. See also event D1h. Event Select: D5h - Dispatch Stall for Reorder Buffer Full The number of processor cycles the decoder is stalled because the reorder buffer is full. May occur simultaneously with certain other stall conditions; see event D1h. Event Select: D6h - Dispatch Stall for Reservation Station Full The number of processor cycles the decoder is stalled because a required integer unit reservation stations is full. May occur simultaneously with certain other stall conditions; see event D1h. Event Select: D7h - Dispatch Stall for FPU Full The number of processor cycles the decoder is stalled because the scheduler for the Floating Point Unit is full. This condition can be caused by a lack of parallelism in FP-intensive code, or by cache misses on FP operand loads (which could also show up as event D8h instead, depending on the nature of the instruction sequences). May occur simultaneously with certain other stall conditions; see event D1h Event Select: D8h - Dispatch Stall for LS Full The number of processor cycles the decoder is stalled because the Load/Store Unit is full. This generally occurs due to heavy cache miss activity. May occur simultaneously with certain other stall conditions; see event D1h. Event Select: D9h - Dispatch Stall Waiting for All Quiet The number of processor cycles the decoder is stalled waiting for all outstanding requests to the system to be resolved. Relatively rare; associated with certain system instructions and types of interrupts. May partially overlap certain other stall conditions; see event D1h.
336
Chapter 11
32559
Rev. 3.08
July 2007
Event Select: DAh - Dispatch Stall for Far Transfer or Resync to Retire The number of processor cycles the decoder is stalled waiting for the execution pipeline to drain before dispatching the target instructions of a far control transfer or a Resync (an instruction stream restart associated with certain microcode assists). Relatively rare; does not overlap with other stall conditions. See also event D1h. Event Select: DBh - FPU Exceptions The number of floating point unit exceptions for microcode assists. The UNIT_MASK may be used to isolate specific types of exceptions.
UNIT_MASK 01h 02h 04h 08h x87 reclass microfaults SSE retype microfaults SSE reclass microfaults SSE and x87 microtraps
Event Select: DCh - DR0 Breakpoint Matches The number of matches on the address in breakpoint register DR0, per the breakpoint type specified in DR7. The breakpoint does not have to be enabled. Each instruction breakpoint match incurs an overhead of about 120 cycles; load/store breakpoint matches do not incur any overhead. Event Select: DDh - DR1 Breakpoint Matches The number of matches on the address in breakpoint register DR1. See notes for event DCh. Event Select: DEh - DR2 Breakpoint Matches The number of matches on the address in breakpoint register DR2. See notes for event DCh. Event Select: DFh - DR3 Breakpoint Matches The number of matches on the address in breakpoint register DR3. See notes for event DCh. 11.2.1.7 Memory Controller Events
Event Select: E0h - DRAM Accesses The number of memory accesses performed by the local DRAM controller. The UNIT_MASK may be used to isolate the different DRAM page access cases. Page miss cases incur an extra latency to open a page; page conflict cases incur both a page-close as well as page-open penalties. These penalties may be overlapped by DRAM accesses for other requests and don't necessarily represent lost DRAM bandwidth. The associated penalties are as follows: Page miss: Chapter 11 Trcd (DRAM RAS-to-CAS delay) Performance Monitoring 337
32559
Rev. 3.08 July 2007
Page conflict:
Trp + Trcd (DRAM row-precharge time plus RAS-to-CAS delay)
Each DRAM access represents one 64-byte block of data transferred if the DRAM is configured for 64-byte granularity, or one 32-byte block if the DRAM is configured for 32-byte granularity. (The latter is only applicable to single-channel DRAM systems, which may be configured either way.)
UNIT_MASK 01h 02h 04h Page hit Page Miss Page Conflict
Event Select: E1 - Memory Controller Page Table Overflows The number of page table overflows in the local DRAM controller. This table maintains information about which DRAM pages are open. An overflow occurs when a request for a new page arrives when the maximum number of pages are already open. Each occurrence reflects an access latency penalty equivalent to a page conflict. Event Select: E3 - Memory Controller Turnarounds The number of turnarounds on the local DRAM data bus. The UNIT_MASK may be used to isolate the different cases. These represent lost DRAM bandwidth, which may be calculated as follows (in bytes per occurrence): DIMM turnaround: DRAM_width_in_bytes * 2 edges_per_memclk * 2 R/W turnaround: R/W turnaround: DRAM_width_in_bytes * 2 edges_per_memclk * 1 DRAM_width_in_bytes * 2 edges_per_memclk * (Tcl-1)
where DRAM_width_in_bytes is 8 or 16 (for single- or dual-channel systems), and Tcl is the CAS latency of the DRAM in memory system clock cycles (where the memory clock for DDR-400, or PC3200 DIMMS, for example, would be 200 MHz).
UNIT_MASK 01h 02h 04h DIMM (chip select) turnaround Read to write turnaround Write to read turnaround
338
Chapter 11
32559
Rev. 3.08
July 2007
Event Select: E4h - Memory Controller Bypass Counter Saturation

UNIT_MASK 01h 02h 04h 08h Memory controller high priority bypass Memory controller low priority bypass DRAM controller interface bypass DRAM controller queue bypass
Event Select: E5 - Sized Blocks Sized Read/Write activity: The Sized Read/Write events reflect 32- or 64-byte transfers (as opposed to other sizes which could be anywhere between 1 and 64 bytes), from either the processor or the Hostbridge (on any node in an MP system). Such accesses from the processor would be due only to write combining buffer flushes, where 32-byte accesses would reflect flushes of partially-filled buffers. Event 65h provides a count of sized write requests associated with WC buffer flushes; comparing that with counts for these events (providing there is very little Hostbridge activity at the same time) will give an indication of how efficiently the write combining buffers are being used. Event 65h may also be useful in factoring out WC flushes when comparing these events with the Upstream Requests component of event ECh.
UNIT_MASK 04h 08h 10h 20h 32-byte Sized Writes 64-byte Sized Writes 32-byte Sized Reads 64-byte Sized Reads
Event Select: E8h - Thermal Status and ECC Errors

UNIT_MASK 01h 02h 04h 08h 80h Number of clocks CPU is active when HTC is active Number of clocks CPU clock is inactive when HTC is active Number of clocks when die temperature is higher than the software high temperature threshold Number of clocks when high temperature threshold was exceeded Number of correctable and Uncorrectable DRAM ECC errors
Event Select: E9h - CPU/IO Requests to Memory/IO These events reflect request flow between units and nodes, as selected by the UNIT_MASK. The UNIT_MASK is divided into two fields: request type (CPU or I/O access to I/O or Memory) and source/target location (local vs. remote). One or more requests types must be enabled via bits 3:0, and
Chapter 11
339
32559
Rev. 3.08 July 2007
at least one source and one target location must be selected via bits 7:4. Each event reflects a request of the selected type(s) going from the selected source(s) to the selected target(s). Not all possible paths are supported. The following table shows the UNIT_MASK values that are valid for each request type:
Source/Target Local -> Local Local -> Remote Remote -> Local Remote -> Remote
CPU to Mem A8h 98h -
CPU to I/O A4h 94h 64h -
I/O to Mem A2h 92h -
I/O to I/O A1h 91h 61h -
Any of the mask values shown may be logically ORed to combine the events. For instance, local CPU requests to both local and remote nodes would be A8h | 98h = B8h. Any CPU to any I/O would be A4h | 94h | 64h = F4h (but remote CPU to remote I/O requests would not be included). Note: It is not possible to tell from these events how much data is going in which direction, as there is no distinction between reads and writes. Also, particularly for I/O, the requests may be for varying amounts of data, anywhere from one to sixty-four bytes. Event E5h provides an indication of 32- and 64-byte read and write transfers for such requests (although from the target point of view). For a direct measure of the amount and direction of data flowing between nodes, use events F6h, F7h and F8h.
UNIT_MASK 01h 02h 04h 08h 10h 20h 40h 80h I/O to I/O I/O to Mem CPU to I/O CPU to Mem To remote node To local node From remote node From local node
Event Select: EAh - Cache Block Commands The number of requests made to the system for cache line transfers or coherency state changes, by request type. Each increment represents one cache line transfer, except for Change-to-Dirty. If a Change-to-Dirty request hits on a line in another processor's cache that's in the Owned state, it will
340
Chapter 11
32559
Rev. 3.08
July 2007
cause a cache line transfer, otherwise there is no data transfer associated with Change-to-Dirty requests.
UNIT_MASK 01h 04h 08h 10h 20h Victim Block (Writeback) Read Block (Dcache load miss refill) Read Block Shared (Icache refill) Read Block Modified (Dcache store miss refill) Change to Dirty (first store to clean block already in cache)
Event Select: EBh - Sized Commands The number of Sized Read/Write commands handled by the System Request Interface (local processor and hostbridge interface to the system). These commands may originate from the processor or hostbridge. Typical uses of the various Sized Read/Write commands are given in the UNIT_MASK table. See also event E5h, which covers commonly-used block sizes for these requests, and event ECh, which provides a separate measure of Hostbridge accesses.
UNIT_MASK 01h 02h 04h NonPosted SzWr Byte (1-32 bytes) NonPosted SzWr Dword (1-16 dwords) Posted SzWr Byte (1-32 bytes) Typical Usage Legacy or mapped I/O, typically 1-4 bytes Legacy or mapped I/O, typically 1 dword Sub-cache-line DMA writes, size varies; also flushes of partially-filled Write Combining buffer Block-oriented DMA writes, often cache-line sized; also processor Write Combining buffer flushes Legacy or mapped I/O Block-oriented DMA reads, typically cacheline size
08h
Posted SzWr Dword (1-16 dwords)
10h 20h 40h
SzRd Byte (4 bytes) SzRd Dword (1-16 dwords) RdModWr
Event Select: ECh - Probe Responses and Upstream Requests This covers two unrelated sets of events: cache probe results, and requests received by the Hostbridge from devices on non-coherent links. Probe results: These events reflect the results of probes sent from a memory controller to local caches. They provide an indication of the degree data and code is shared between processors (or moved between processors due to process migration). The dirty-hit events indicate the transfer of a 64-byte cache line to the requestor (for a read or cache refill) or the target memory (for a write). The system bandwidth used by these, in terms of bytes per unit of time, may be calculated as 64 times the event count, divided by the elapsed time. Sized writes to memory that cover a full cache line do not Chapter 11 Performance Monitoring 341
32559
Rev. 3.08 July 2007
incur this cache line transfer -- they simply invalidate the line and are reported as clean hits. Cache line transfers will occur for Change2Dirty requests that hit cache lines in the Owned state. (Such cache lines are counted as Modified-state refills for event 6Ch, System Read Responses.) Upstream requests: The upstream read and write events reflect requests originating from a device on a local non-coherent HyperTransport link. The two read events allow display refresh traffic in a UMA system to be measured separately from other DMA activity. Display refresh traffic will typically be dominated by 64-byte transfers. Non-display-related DMA accesses may be anywhere from 1 to 64 bytes in size, but may be dominated by a particular size such as 32 or 64 bytes, depending on the nature of the devices. Event E5h can provide a measure of 32- and 64-byte accesses by the hostbridge (possibly combined with write combining buffer flush activity from the processor, although that can be factored out via event 65h).
UNIT_MASK 01h 02h 04h 08h 10h 20h 40h Probe miss Probe hit clean Probe hit dirty without memory cancel (probed by Sized Write or Change2Dirty) Probe hit dirty with memory cancel (probed by DMA read or cache refill request) Upstream display refresh reads Upstream non-display refresh reads Upstream writes
Event Select: EEh - GART Events These events reflect GART activity, and in particular allow one to calculate the GART TLB miss ratio as GART_miss_count divided by GART_aperture_hit_count. GART aperture accesses are typically from I/O devices as opposed to the processor, and generally from a 3D graphics accelerator, but can be from other devices when the GART is used as an IO MMU).
UNIT_MASK 01h 02h 04h GART aperture hit on access from CPU GART aperture hit on access from I/O GART miss
342
Chapter 11
32559
Rev. 3.08
July 2007
11.2.1.8 11.2.1.9
Crossbar Events HyperTransport Interface Events
Event Select: F6h - HyperTransport Link 0 Transmit Bandwidth F7h - HyperTransport Link 1 Transmit Bandwidth F8h - HyperTransport Link 2 Transmit Bandwidth The number of dwords transmitted (or unused, in the case of Nops) on the outgoing side of the HyperTransport links. The sum of all four subevents (all four UNIT_MASK bits set) directly reflects the maximum transmission rate of the link. Link utilization may be calculated by dividing the combined Command, Data and Buffer Release count (UNIT_MASK 07h) by that value plus the Nop count (UNIT_MASK 08h). Bandwidth in terms of bytes per unit time for any one component or combination of components is calculated by multiplying the count by four and dividing by elapsed time. The Data event provides a direct indication of the flow of data around the system. Translating this link-based view into a source/target node based view requires knowledge of the system layout (i.e. which links connect to which nodes).
UNIT_MASK 01h 02h 04h 08h Command dword sent Data dword sent Buffer release dword sent Nop dword sent (idle)
Chapter 11
343
32559
Rev. 3.08 July 2007
344
Chapter 11
32559
Rev. 3.08
July 2007
12
CPUID Function Registers
Processor feature capabilities and configuration information are provided through the CPUID instruction. Different information is accessed by (1) setting EAX as an index to the registers to be read, (2) executing the CPUID instruction, and (3) reading the results in EAX, EBX, ECX, and EDX. The phrase CPUID function X or CPUID FnX refers to the CPUID instruction when EAX is preloaded with X. Undefined function numbers return 0s in all 4 registers. The following provides AMD Family 0Fh specific details about CPUID. See the CPUID Specification for further information. CPUID Fn[8000_0000,0000_0000] AMD Authentic Identifier
Register EAX Bits 31:0 Description Function 0000_0000h returns the largest CPUID standard-function input value supported by the processor implementation: 0000_0005h. Function 8000_0000h returns the largest CPUID extended-function input value supported by the processor implementation: 8000_0019h. EBX, ECX, EDX 31:0 The 12 8-bit ASCII character codes to create the string AuthenticAMD. EBX=6874_7541h h t u A, ECX=444D_4163h D M A c, EDX=6974_6E65h i t n e
CPUID Fn[8000_0001,0000_0001]_EAX Family, Model, Feature Identifiers This register provides identical information to Function 3, Offset FCh. Family is an 8-bit value and is defined as: Family[7:0] = ({0000b,BaseFamily[3:0]} + ExtendedFamily[7:0]). E.g. If BaseFamily[3:0]=Fh and ExtendedFamily[7:0]=01h, then Family[7:0]=10h. This document applies only to family 0Fh processors. Model is an 8-bit value and is defined as: Model[7:0] = {ExtendedModel[3:0], BaseModel[3:0]}. E.g. If ExtendedModel[3:0]=Eh and BaseModel[3:0]=8h, then Model[7:0] = E8h. Model numbers vary with product.
Bits 31:28 27:20 19:16 15:12 11:8 7:4 3:0 Description Reserved. ExtendedFamily: 00h. ExtendedModel. Reserved. BaseFamily: Fh BaseModel. Stepping: processor stepping (revision) for a specific model.
Chapter 12
345
32559
Rev. 3.08 July 2007
CPUID Fn[0000_0001]_EBX LocalApicId, LogicalProcessorCount, CLFlush, 8BitBrandId

Bits 31:24 Description LocalApicId: Provides the initial APICID (APIC Offset 20h) value in the three LSBs. After NodeId (Function 0: Offset 60h) has been initialized, changes to APICID do not effect the value of this register. LogicalProcessorCount: If CPUID Fn[8000_0001, 0000_0001]_EDX[HTT] = 1, then this field indicates the number of CPU cores in the processor as CPUID Fn8000_0008[NC] + 1. Otherwise, this field is reserved. CLFlush: CLFLUSH size in quadwords = 08h. 8BitBrandId: 8 bit brand ID = 00h. Indicates that the brand ID is in CPUID Fn8000_0001_EBX.
23:16
15:8 7:0
CPUID Fn[8000_0001]_EBX BrandId Identifier

Bits 31:16 15:0 Description Reserved. BrandId: brand ID.
CPUID Fn0000_0001_ECX Feature Identifiers These fields are encoded as 1=Feature is supported by the processor; 0=Feature is not supported by the processor.
Bits 31:14 13 12:1 0 Description Reserved. CMPXCHG16B: CMPXCHG16B instruction = 1. Reserved. SSE3: SSE3 extensions = 1.
CPUID Fn8000_0001_ECX Feature Identifiers These fields are encoded as 1=Feature is supported by the processor; 0=Feature is not supported by the processor.
Bits 31:9 8 7:5 4 3 2 Description Reserved. 3DNowPrefetch: Prefetch and PrefetchW instructions = 1. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. Reserved. LockMovCr0: LOCK MOV CR0 means MOV CR8 = 1. ExtApicSpace: extended APIC register space = 1. SVM: Secure Virtual Mode feature (setting varies by product).
346
Chapter 12
32559
Rev. 3.08
July 2007
Bits 1 0
Description CmpLegacy: core multi-processing legacy mode (setting varies by product). LahfSahf: LAHF/SAHF instructions =1.
CPUID Fn[8000_0001, 0000_0001]_EDX Feature Identifiers These fields are encoded as 1=Feature is supported by the processor; 0=Feature is not supported by the processor. The value returned in EDX may be identical or different for Fn0000_0001 and Fn8000_0001, as indicated.
Bits 31 30 29 Function 0000_0001 8000_0001 0000_0001 8000_0001 0000_0001 8000_0001 0000_0001 28 8000_0001 27 26 25 24 23 22 21 20 19 18 17 16 0000_0001 8000_0001 0000_0001 8000_0001 0000_0001 8000_0001 both both 0000_0001 8000_0001 both 0000_0001 8000_0001 0000_0001 8000_0001 both both both Description Reserved. 3DNow: 3DNow! instructions = 1. Reserved. 3DNowExt: AMD extensions to 3DNow! instructions = 1. Reserved. LM: long mode (may vary by product). HTT: hyper-threading technology. This bit qualifies the meaning of CPUID Fn0000_0001_EBX[LogicalProcessorCount]. Reserved. Reserved. RDTSCP: RDTSCP instruction = 1. SSE2: SSE2 extensions = 1. Reserved. SSE: SSE extensions = 1. FFXSR: FXSAVE and FXRSTOR instruction optimizations = 1. FXSR: FXSAVE and FXRSTOR instructions = 1. MMX: MMX instructions = 1. Reserved. MmxExt: AMD extensions to MMX instructions = 1. Reserved. Reserved. NX: no-execute page protection = 1. CLFSH: CLFLUSH instruction = 1. Reserved. Reserved. PSE36: page-size extensions = 1. PAT: page attribute table = 1.
Chapter 12
347
32559
Rev. 3.08 July 2007
Bits 15 14 13 12 11 10 9
Function both both both both 0000_0001 8000_0001 both both
Description CMOV: conditional move instructions, CMOV, FCOMI, FCMOV = 1. MCA: machine check architecture, MCG_CAP = 1. PGE: page global extension, CR4.PGE = 1. MTRR: memory-type range registers = 1. SysEnterSysExit: SYSENTER and SYSEXIT instructions = 1. SysCallSysRet: SYSCALL and SYSRET instructions = 1. Reserved. APIC: advanced programmable interrupt controller (APIC) exists and is enabled. This bit reflects the state of the APIC Base Address register (APIC_BAR) MSR 0000_001B[ApicEn]. CMPXCHG8B: CMPXCHG8B instruction = 1. MCE: machine check exception, CR4.MCE = 1. PAE: physical-address extensions (PAE) = 1. MSR: AMD model-specific registers (MSRs), with RDMSR and WRMSR instructions = 1. TSE: time stamp counter, RDTSC/RDTSCP instructions, CR4.TSD = 1. PSE: page-size extensions (4 MB pages) = 1. DE: debugging extensions, IO breakpoints, CR4.DE = 1. VME: virtual-mode enhancements = 1. FPU: x87 floating point unit on-chip = 1.
8 7 6 5 4 3 2 1 0
both both both both both both both both both
CPUID Fn8000_000[4, 3, 2] Processor Name String Identifier These return the ASCII string corresponding to the processor name, stored in MSR C001_00[35:30]. The MSRs are mapped to these registers as follows: Function 8000_0002: {EDX, ECX, EBX, EAX} == {MSRC001_0031, MSRC001_0030}; Function 8000_0003: {EDX, ECX, EBX, EAX} == {MSRC001_0033, MSRC001_0032}; Function 8000_0004: {EDX, ECX, EBX, EAX} == {MSRC001_0035, MSRC001_0034}; CPUID Fn8000_0005 TLB and L1 Cache Identifiers This provides the processors first level cache and TLB characteristics for each CPU core. The associativity fields returned are encoded as follows: 00h Reserved. 01h Direct mapped. 02h - FEh Specifies the associativity; e.g., 04h would indicate a 4-way associativity. FFh Fully associative.
Register EAX Bits 31:24 Description Data TLB associativity for 2 MB and 4 MB pages = FFh.
348
Chapter 12
32559
Rev. 3.08
July 2007
Register EAX
Bits 23:16
Description Data TLB number of entries for 2 MB and 4 MB pages = 8. The value returned is for the number of entries available for the 2 MB page size; 4 MB pages require two 2 MB entries, so the number of entries available for the 4 MB page size is one-half the returned value. Instruction TLB associativity for 2 MB and 4 MB pages = FFh. Instruction TLB number of entries for 2 MB and 4 MB pages = 8. The value returned is for the number of entries available for the 2 MB page size; 4 MB pages require two 2 MB entries, so the number of entries available for the 4 MB page size is one-half the returned value. Data TLB associativity for 4 KB pages = FFh. Data TLB number of entries for 4 KB pages = 32. Instruction TLB associativity for 4 KB pages = FFh. Instruction TLB number of entries for 4 KB pages = 32. L1 data cache size in KB = 64. L1 data cache associativity = 2. L1 data cache lines per tag = 1. L1 data cache line size in bytes = 64. L1 instruction cache size in KB = 64. L1 instruction cache associativity = 2. L1 instruction cache lines per tag = 1. L1 instruction cache line size in bytes = 64.
EAX EAX
15:8 7:0
EBX EBX EBX EBX ECX ECX ECX ECX EDX EDX EDX EDX
31:24 23:16 15:8 7:0 31:24 23:16 15:8 7:0 31:24 23:16 15:8 7:0
Chapter 12
349
32559
Rev. 3.08 July 2007
CPUID Fn8000_0006 L2 Cache Identifiers This provides the processors second level cache and TLB characteristics for each CPU core. The presence of a unified L2 TLB is indicated by a value of 0000h in the upper 16 bits of the EAX and EBX registers. The unified L2 TLB information is contained in the lower 16 bits of these registers. The associativity fields are encoded as follows: 0h: The L2 cache or TLB is disabled. 1h: Direct mapped. 2h: 2-way associative. 4h: 4-way associative. All other encodings are reserved.
Register EAX EBX EBX EBX EBX ECX ECX ECX ECX EDX Bits 31:0 31:24 23:16 15:8 7:0 31:24 23:16 15:8 7:0 31:0 Description The L2 TLB has no entries for 2 MByte and 4 MByte pages. 0000_000h L2 data TLB associativity for 4 KB pages = 4. L2 data TLB number of entries for 4 KB pages = 512. L2 instruction TLB associativity for 4 KB pages = 4. L2 instruction TLB number of entries for 4 KB pages = 512. L2 cache size in KB (varies with product). May be one of 128, 256, 512, or 1024. L2 cache associativity = 8. L2 cache lines per tag = 1. L2 cache line size in bytes = 64. Reserved.
6h: 8-way associative. 8h: 16-way associative. Fh: Fully associative.
CPUID Fn8000_0007 Advanced Power Management Information This provides advanced power management feature identifiers. Unless otherwise specified, these bits are encoded as 1=Feature is supported by the processor; 0=Feature is not supported by the processor.
Register EAX, EBX, ECX EDX EDX Bits 31:0 Description Reserved.
31:7 6
Reserved. 100MhzSteps: 100 Mhz multiplier Control = 1. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. STC: software thermal control (STC) is supported (support may vary by product). TM: hardware thermal control (HTC) is supported (support may vary by product). TTP: THERMTRIP is supported = 1. VID: Voltage ID control is supported.
EDX EDX EDX EDX
5 4 3 2
350
Chapter 12
32559
Rev. 3.08
July 2007
Register EDX EDX
Bits 1 0
Description FID: Frequency ID control is supported. TS: Temperature sensor = 1.
CPUID Fn8000_0008 Address Size And Physical Core Count Information This provides information about the number of physical CPU cores and the maximum physical and linear address width supported by the processor.
Register EAX EAX Bits 31:16 15:8 Description Reserved. LinAddrSize: Maximum linear byte address size in bits. If the processor supports long mode (see CPUID Fn[8000_0001, 0000_0001]_EDX[LM]) then this is 30h; else this is 20h. PhysAddrSize: Maximum physical byte address size in bits = 28h. Reserved. Reserved. NC: number of physical CPU cores - 1. The number of CPU cores in the processor is NC+1 (e.g., if NC=0, then there is one CPU core). Reserved.
EAX EBX ECX ECX EDX
7:0 31:0 31:8 7:0 3:0
CPUID Fn8000_0009 Reserved CPUID Fn8000_000A SVM Revision and Feature Identification This provides SVM revision and feature information. If CPUID Fn8000_0001_ECX[SVM]=0 then CPUID Fn8000_000A is reserved.
Register EAX EAX EBX ECX EDX EDX EDX Bits 31:8 7:0 31:0 31:0 31:2 1 0 Description Reserved. SvmRev: SVM revision = 01h. NASID: number of address space identifiers (ASID) = 40h. Reserved. Reserved. LbrVirtulization: LBR Virtulazation = 1. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. Reserved.
Chapter 12
351
32559
Rev. 3.08 July 2007
352
Chapter 12
32559
Rev. 3.08
July 2007
13
BIOS Checklist
The checklist in this chapter reviews information that is described elsewhere and is required by BIOS developers to properly incorporate AMD NPT Family 0Fh Processors into systems.
13.1

CPUID
Use the CPUID instruction to properly identify the processor. Determine the processor type, stepping, and available features by using functions 0000_0001h and 8000_0001h of the CPUID instruction. Boot-up display - The processor name should be displayed according to the processor name string defined in AMD Revision Guide for NPT Processors, order# 33610
For more information on the CPUID instruction, see Chapter 12, CPUID Function Registers, the AMD Processor Recognition Application Note, order# 20734 and the CPUID Guide for the AMD Athlon 64 and AMD Opteron Processors, order# 25481.
13.2
CPU Speed Detection
The BIOS can use the time-stamp counter (TSC) to clock a timed operation and compare the result to the Real-Time Clock (RTC) to determine the operating frequency. See the example of frequencydetermination assembler code available on the AMD web site.
13.3
HyperTransport Link Detection
After any reset, the BIOS should check every HyperTransport link on the processor to detect whether the link is connected by executing the following steps: 1. Wait until the LinkConPend bit in Function 0, Offset 98h, B8h, or D8h is cleared. 2. Check LinkCon bit in Function 0, Offset 98h, B8h, or D8h. If LinkCon bit has a value of 0, the link is not connected. If LinkCon bit has a value of 1, the link is connected. After a warm reset, if the link is connected, the BIOS should additionally check LinkFail bit in Function 0, Offset 84h, A4h, or C4h to determine if sync flood packets were detected by the link receivers before the warm reset assertion. The LinkFail bit should then be cleared.
Chapter 13
BIOS Checklist
353
32559
Rev. 3.08 July 2007
13.4
HyperTransport Link Frequency Selection
The BIOS should determine the set of frequencies that a processor is capable of for every HyperTransport link connected to the processor. Each frequency in the set must be defined by the corresponding link frequency capability bit in the LnkFreqCap field, Function 0, Offsets 88h, A8h, C8h, and it must be equal to or less than the maximum frequency values specified in the processor data sheets. If a HyperTransport link is non-coherent, then the BIOS should initialize the HyperTransport link frequency (Freq field, Function 0, Offsets 88h, A8h, C8h) with the highest value from the set of frequencies of which the processor is capable and which is less than or equal to the maximum frequency values specified in the chipset data sheets, and the maximum frequency supported by the platform. If a HyperTransport link is coherent, then the BIOS should initialize the HyperTransport link frequencies (Freq field, Function 0, Offsets 88h, A8h, C8h) of both processors with the highest common frequency value from the sets of frequencies of which each processor is capable that is less than or equal to the maximum frequency supported by the platform.
13.5
Multiprocessing Capability Detection
The multiprocessing capability of the AMD Opteron processor is determined by the MPCap and BigMPCap bits in the Northbridge Capability Register (Function 3, Offset E8h).
Multiprocessing Capability MPCap BigMPCap
UP Capable DP Capable MP Capable
0 1 1
0 0 1
During POST, the BIOS checks the multiprocessing capability of AMD Opteron processors, and configures the system accordingly. MP setup by the BIOS is also required in a UP where dual-cores are present. Multiprocessing capability detection is not required in a UP system. All processors must be DP capable or MP capable in a DP system. If any processor is only UP capable, the BIOS must configure the BSP as a UP processor, and must not initialize the AP. All processors must be MP capable in an MP system. If any processor is not MP capable, the BIOS must configure the BSP as a UP processor, and must not initialize APs. If all processors do not have adequate multiprocessing capability for a DP or an MP system, the BIOS must display the following message:
*********** Warning: non-MP Processor ***********
354
BIOS Checklist
Chapter 13
32559
Rev. 3.08
July 2007
The processor(s) installed in your system are not multiprocessing capable. Now your system will halt.
If all processors have adequate multiprocessing capability for a DP or an MP system, but have different model numbers or operate at different frequencies, the BIOS can do one of the following: configure the BSP as a UP processor, not initialize APs, and display a message indicating that processors with different model numbers or maximum frequencies have been installed, or implement the algorithm defined in Processors With Different Revisions or Maximum Frequencies in Multiprocessor Systems.
13.6
Setup for Dual Core Processors
To ensure that dual core processors work correctly in existing operating systems, the BIOS must follow several steps. These steps must only be performed on a dual core processor when both cores are enabled. 1. Re-direct all MC4 accesses and error logging to core 0. This is done by setting the NbMcaToMstCpuEn bit (Function 3, Offset 44h). Setting this bit is required to ensure that current operating systems do not become confused with conflicting programming of this shared resource between two cores. 2. Ensure that the core number is the least significant bit of the initial APIC ID. This is done by setting the InitApicIdCpuIdLo bit (MSR C001_001Fh, bit 54) in each core. Setting this bit is required because current operating systems expect the core ID bits to be in the least significant bit positions of the APIC ID. The operating system makes scheduling decisions based on this information. 3. The firmware must properly report dual-core processors in a number of tables required by the Advanced Configuration and Power Interface (ACPI) specification: The Static Resource Affinity Table (SRAT) must contain a "Processor Local APIC/SAPIC Affinity Structure" for each core rather than for each processor. Cores within the same processor must have the same proximity domain. The CPU objects in the Differentiated System Description Table (DSDT) must contain the correct number of cores. The Multiple APIC Description Table (MADT) must contain a "Processor Local APIC" record and an "ACPI Processor object" for each core rather than for each processor. 4. Each core, rather than each processor, should have a processor entry in the MP configuration table as defined in the MultiProcessor Specification ver, 1.4. 5. Each core, rather than each processor, should have a unique SMM Base Address specified in MSR (C001_0111h).
Chapter 13
BIOS Checklist
355
32559
Rev. 3.08 July 2007
6. Software can always calculate the number of physical processors in a system by taking the total number of processor entries/records in the MADT or MP configuration tables and dividing by the number of cores per processor returned by CPUID Fn8000_0008. 7. The order of the processor entries/records in the MADT and the MP configuration table are not specified by AMD. However, some firmware vendors may choose to work around the licensing restrictions of some legacy operating systems by listing the entries/records of the first core of each processor (core 0) before listing the entries/records of the remaining cores: processor 0, core 0 ... processor N, core 0 processor 0, core 1 ... processor N, core 1
13.7
APIC ID Assignment Requirements
The arbitration protocol used when sending data over the original serial APIC bus required that IOAPIC IDs be unique and not overlap the processor APIC IDs (i.e. mapped into the same linear ID space such as APIC IDs from 0 to N and IOAPIC IDs from N+1 to N+M). A number of operating systems may still have this requirement even though AMD NPT processors do not require it. Most of these operating systems also support numbering the IOAPIC IDs first before numbering the processor APIC IDs. Firmware should support this model for the broadest operating system support. If the total number of IOAPICs plus the total number of processor APICs (number of processor * number of cores per processor) is less than 16: APIC_ID[node=i] = i for a single-core system APIC_ID[node=i, core=j] = i*2 + j for a dual-core system APIC_ID for IOAPIC starts from the next available number after the last APIC_ID assigned to the processor.
If the total number of processor APICs is equal to or greater than 16: APIC_ID for IOAPIC starts from 0 APIC_ID[node=i] = i + IOAPIC_OFFSET for a single-core system APIC_ID[node=i, core=j] = (i*2 + j) + IOAPIC_OFFSET for a dual-core system. IOAPIC_OFFSET = total number of IOAPIC, or total number of IOAPIC + 1 if it is an odd number.
356
BIOS Checklist
Chapter 13
32559
Rev. 3.08
July 2007
13.8
Processors With Different Revisions or Maximum Frequencies in Multiprocessor Systems
AMD Opteron processors with different revisions or maximum frequencies can be mixed in a multiprocessor system. Refer to the Revision Guide for AMD NPT Family 0Fh Processors, order# 33610 for information about processors that can be mixed in a multiprocessor system. The following requirements must be met when processors of different revisions or maximum frequencies are mixed: 1. The system must be configured as a DP or an MP system. See section Multiprocessing Capability Detection on page 354 for more information. 2. All processors must properly initialize the Freq field in the corresponding Function 0, Offset 88h, A8h, or C8h register. See HyperTransport Link Frequency Selection on page 354 for more information. 3. The BIOS must set the FID of all processors in the system to the same value. That value should be the highest common FID for all processors in the system. The set of FID values of which each processor is capable is determined as follows:
Execute CPUID instruction with EAX=80000007h; if(EDX[1] == 1) // Processor supports FID changes. Read MSR C001_0042h; // Get FIDVID Status register. MaxFID = EAX[21:16]; // MaxFID is the maximum FID of which the processor is capable. // The processor is capable of all FIDs from MaxFID down to the FID that // represents 1600 MHz, and if any FID is a portal core frequency, then // the corresponding minimum core frequency is also a FID of which the // processor is capable. See Table 66. else // Processor does not supports FID changes. Read MSR C001_0015h; // Get HWCR register. START_FID = EAX[29:24]; // This is the only FID of which the processor is capable.
The BIOS should determine whether a highest common FID value for all processors in the system exists. If such a value exists, then FID values for all processors must be set to that value. See Chapter 10, Power and Thermal Management for more information on FID changes. If there is not a common FID value for all processors in the system, then the BIOS should not change the FID on any processor, and it should display a warning message. Power management registers defined in Table 75 must be programmed before changing the processor FID value. 4. The BIOS must follow P-state change rules specified in Chapter 10, Power and Thermal Management, however, VID changes do not have to be made. 5. All processors must have the same L2 cache size, returned in ECX[31:16] when a CPUID instruction is executed with EAX=80000006h. If all processors in the system do not have the Chapter 13 BIOS Checklist 357
32559
Rev. 3.08 July 2007
same L2 cache size, then the BIOS should display a warning message. See CPUID Guide for the AMD Athlon 64 and AMD Opteron Processors, order# 25481 for more information. 6. The BIOS must ensure that all CPUs present the same feature set to the operating system. For instance, if a CPU in the system does not support SSE3 instructions, then all CPUs in the system must be configured to not report the SSE3 capability to OS or application software. To accomplish this, BIOS must first determine the feature set that is common to all CPUs in the system. Then, for all CPUs, BIOS must disable any features that are not supported by all CPUs. The following registers contain the feature set bits: CPUIDFeatures (MSR C001_1004h) CPUIDExtFeatures (MSR C001_1005h) The bits in these registers identify individual features. A value of 1 indicates that the feature is supported and a value of 0 indicates that the feature is not supported. All of these bits are read/write, thus, if a bit has a value of 1, it can be written to a 0 to make it appear to the OS that the feature is not supported. 7. The errata, workarounds, and microcode updates must be properly applied to each processor in the system. See Revision Guide for AMD NPT Family 0F Processors, order# 33610 for more information. 8. If S3 power state is implemented in the system, all steps listed above must be executed on a resume from S3.
13.9
Model-Specific Registers (MSRs)
Access only the MSRs that are implemented in the processor. Follow the processor MSR programming sequence described in this document. This includes setting the HWCR, SYSCFG, MTRRs, IORRs, clock control MSR, Top of Memory, and other registers.
13.9.1
Software Considerations for Accessing Northbridge MSRs in Dual Core Processors
MSRs that control northbridge functions are shared between both cores in a dual core processor. If control of northbridge functions is shared between software on both core, software must ensure that only one core at a time is allowed to access the shared MSR. Table 77.
Address 0410h 0411h
Northbridge MSRs
Register Name MC4_CTL MC4_STATUS Description page 228 page 229
358
BIOS Checklist
Chapter 13
32559
Rev. 3.08
July 2007
Table 77.
Address 0412h 0413h
Northbridge MSRs (Continued)

Register Name MC4_ADDR MC4_MISC NB_CFG HTC FIDVID_CTL FIDVID_STATUS MCi_CTL_MASK Description page 229 page 229 page 401 page 403 page 411 page 413 page 229
C001_001Fh C001_003Eh C001_0041h C001_0042h C001_0048h
13.10 Machine Check Architecture (MCA)

The processor supports a machine check architecture that allows reporting and handling of system errors through a consistent interface. The global MCA Status register must be cleared (all machine check features disabled). MC0_CTL must be set to all 1s because some do not set these bits. In addition, the local MCA status and address registers for the following five reporting blocks must also be cleared (all error and parity bits disabled): BUBus Unit ICInstruction Cache Unit DCData Cache Unit LSLoad Store Unit NBNorthbridge Unit
After reset, the BIOS needs to determine if the cause of the reset was a power-on or soft reset. When power-on reset occurs, the BIOS must clear the MCA registers. Otherwise, the BIOS should check for any MCA information that may be related to the soft reset.
For more information on MCA, see Chapter 6, Machine Check Architecture.
13.10.1 GART Table Walk Error Reporting

The GART table walk error reporting function is enabled by setting bit 10 of the MC4_CTL register (MSR 410h). This results in a machine check exception when an AGP aperture access has no matching translation in the GART. This error is typically caused by a software graphics driver that improperly reserves or allocates aperture pages in the GART, resulting in benign visual artifacts Chapter 13 BIOS Checklist 359
32559
Rev. 3.08 July 2007
which are often undetected on other platforms. Setting MC4_CTL[10] allows software developers to debug this error; the resulting benign machine check errors can, however, confuse an end user. For this reason, AMD recommends that the BIOS developers disable this function by setting bit 10 of MC4_CTL_MASK register (MSR C001_0048h) to a value of 1. This bit must be set before MC4_CTL[10] bit is set. AMD also recommends adding a setup option to the BIOS setup menu. The following should be displayed in the setup option:
Gart Table Walk Error MC reporting: Disabled/Enabled.
The default setting is disabled. The device driver developer may enable this function for implementation and testing purposes. Also, a help message should be added with this setup option. An example of the help message is:
This option should remain disabled for normal operation.
13.11 GART Register Restoration After S3 Resume

After a resume from S3, GartEn (Function 3, Offset 90h) must be restored after all GART related and AGP device registers are restored.
13.12 Register Settings in UMA Systems

AMD recommends that UMA graphics systems use the following register field values:
Register and Field Position Function 0, Offset 68h, [28] Function 0, Offset 68h, [11] Function 2, Offset A0h, [5] Function 2, Offset 90h, [10] Function 2, OffsetA0h, [3:2] MSR C001_001Fh, [36] MSR C001_001Fh, [9] Field Name EnCpuRdHiPri, enable high priority CPU reads RspPassPW, response pass posted writes DCC_EN, dynamic idle cycle counter enable BurstLength32, enable 32-byte granularity RdWrQByp, read/write queue bypass count DisDatMsk, disable data mask DisUseRefFreeBuf, enable display refresh to use free list buffers Value 0b 11b 1b 1b 111b 1b 11b 0b 0b
Function 0, Offset 68h, [27:26] HiPriBypCnt, high-priority bypass count
Function 2, Offset 94h, [27:24] DcqBypMax, bypass maximum
Refer to Table 35 for information on recommended flow control buffer allocation in UMA graphics systems.
360
BIOS Checklist
Chapter 13
32559
Rev. 3.08
July 2007
13.13 Memory Map

13.13.1 I/O and Memory Type and Range Registers (IORRs, MTRRs)
The memory-type and range registers control the access and cacheability of memory regions in the processor. Follow the processor MTRR and IORR programming sequence that is described in this document. To initialize MTTRs in BIOS, set the default memory cache type for all addresses to uncacheable, then use a variable range MTRR to set all physical memory as cacheable (writeback), and finally use the fixed-range MTRRs to handle the special cases in the 640-Kbyte1-Mbyte region. MTRRs should be modified while cache is disabled. Modifying MTRRs while cache is enabled results in undefined behavior. The GART driver (miniport) should program the AGP aperture. The BIOS does not map the cacheability of the video frame buffer and AGP aperture space; it is done by the operating system, video, and AGP drivers.
13.13.2 Memory Map Registers (MMRs)

The processor contains several memory map registers to support the appropriate software view of memory. Program the correct values in the MSRs, such as Top of Memory (TOP_MEM).
13.14 Access Routing And Type Determination

13.14.1 Memory Access to the Physical Address Space
All memory accesses to the physical address space from a CPU core are sent to its associated northbridge (NB). All memory accesses from an I/O link are routed through the NB. An I/O link access to physical address space indicates to the NB the cache attribute (Coherent or NonCoherent). A CPU core access to physical address space has two attributes that the CPU must determine before issuing the access to the NB: the cache attribute (e.g., WB, WC, UC) and the access destination (DRAM or MMIO). 13.14.1.1 Determining The Cache Attribute 1. The CPU translates the logical address to a physical address. In that process it determines the initial cache attribute based on the settings of the Page Table Entry PAT bits, the MTRR Default Memory Type Register (MSR 02FF), the Variable-Size MTRRs (MSRs 02[0F:00]), the Fixed-
Chapter 13
BIOS Checklist
361
32559
Rev. 3.08 July 2007
Size MTRRs (MSRs 02[6F:68, 59, 58, 50]), and the SYSCFG bit Tom2ForceMemTypeWB (MSR C001_0010h bit 22). 2. The ASeg and TSeg SMM mechanisms are then checked in parallel to determine if the initial cache attribute should be overridden (see MSR C001_0112 and MSR C001_0113). If the address falls within an enabled ASeg/TSeg region, then the final cache attribute is determined as specified in MSR C001_0113. This mechanism is managed by the BIOS and does not require any setup or changes by system software. 13.14.1.2 Determining The Access Destination for CPU Accesses The access destination is based on the highest priority of the following ranges that the access falls in: 1. (Lowest priority) Compare against the top-of memory registers (MSR C001_001A and MSR C001_001D). 2. The Fixed-Size MTRRs (MSRs 02[6F:68, 59, 58, 50]). 3. The IORRs (MSRs C001_00[18, 16] and MSRs C001_00[19, 17]). 4. TSEG & ASEG (MSR C001_0112 and MSR C001_0113). 5. (Highest priority) NB AGP aperture range registers. To determine the access destination, the following steps are taken: 1. The CPU compares the address against the Top Of Memory Register (MSR C001_001A), and the Top Of Memory 2 Register (MSR C001_001D), to determine if the default access destination is DRAM or MMIO space. 2. For addresses below 1M byte, the address is then compared against the appropriate Fixed MTRRs (MSRs 02[6F:68, 59, 58, 50]) to override the default access destination. Each fixed MTRR includes two bits, RdDram and WrDram, that determine the destination based on the access type. 3. The CPU then compares the address against the IORRs (MSRs C001_00[18, 16] and MSRs C001_00[19, 17]); if it matches, the default access destination is overridden as specified by the IORRs. BIOS can use the IORRs to create an I/O hole within a range of addresses that would normally be mapped to DRAM. It can also use the IORRs to re-assert a DRAM destination for a range of addresses that fall within a bigger I/O hole that overlays DRAM. Some key points to consider: a. Operating system software never needs to program IORRs to re-map addresses that naturally target DRAM; any such programming is done by the BIOS. b. The IORRs should not cover the range used for the AGP aperture if the GART logic in the NB is enabled. c. The IORRs should be programmed to cover the AGP aperture if the aperture/GART translation is handled by an I/O device (e.g., the chipset).
362
BIOS Checklist
Chapter 13
32559
Rev. 3.08
July 2007
4. The ASeg and TSeg SMM mechanisms are then checked in parallel to determine if the destination should be overridden (see MSR C001_0112 and MSR C001_0113). If the address falls within an enabled ASeg/TSeg region, then the destination is determined as specified in MSR C001_0113. This mechanism is managed by the BIOS and does not require any setup or changes by system software.
13.14.2 Northbridge Processing of Accesses To Physical Address Space

The physical address, its cacheability type, its access type and its access destination is presented to the processor NB for further processing. Processing of an access is performed by NB in the following manner: 1. Regardless of the access destination, if supplied, the physical address is checked against the NBs AGP-aperture range-registers, if enabled; if the address matches, the NB translates the physical address through the AGP GART. A match in the AGP aperture overrides any match to the NB's DRAM Base/Limit Registers, Function 1, Offsets [7C:40], and Memory Mapped I/O Base/Limit Registers, Function 1, Offsets [BC:80]. 2. For accesses from I/O devices, the cacheability attribute from the GART entry (GartPteCoh bit) is applied; for accesses from a CPU, the attribute already applied by the CPU is used and the GartPteCoh bit is ignored. (System software should ensure that the cacheability attribute assigned to an AGP aperture matches the GartPteCoh bit in the matching GART entry.) 3. Accesses that do not match the AGP aperture and post-GART translated addresses are compared against the NB DRAM Base/Limit Registers, Function 1, Offsets [7C:40], and Memory Mapped I/O Base/Limit Registers, Function 1, Offsets [BC:80], to determine the destination link of the access. 4. For accesses from a CPU, the access destination determines whether to compare the address against the DRAM Base/Limit Registers, Function 1, Offsets [7C:40] or the Memory Mapped I/O Base/Limit Registers, Function 1, Offsets [BC:80]. For accesses from an I/O device both sets of registers are checked; if an access matches against both, the MMIO registers take precedence.
13.15 Error Handling

Error detection and signaling capabilities of all system components: processors, chipsets, BIOS and operating system must be considered when selecting error handling mechanisms.
13.15.1 Error Handling Mechanisms

Sync flooding and machine check exception handling mechanisms are described. 13.15.1.1 Sync Flooding Sync flooding is a method used to stop data propagation in the case of a serious error. The device that detects the error initiates sync flood, and all devices in the system that detect sync flood packets on its Chapter 13 BIOS Checklist 363
32559
Rev. 3.08 July 2007
HyperTransport receivers cease normal operation and start transmitting sync flood packets. Sync flood packets reach the I/O hub. If the I/O hub is programmed to generate a warm reset in the case of sync flood detection, the BIOS can analyze all error bits in the system that were not cleared by the warm reset and detect the error source. If the I/O hub is not programmed to generate a warm reset, the system hangs. 13.15.1.2 Machine Check Exceptions Machine check exceptions are generated when uncorrectable error conditions and certain correctable error conditions are detected by AMD NPT processors. Detected error conditions are caused by the processor or an external I/O device. Machine check exception architecture can be programmed to generate sync flood instead of a machine check exception for some types of uncorrectable errors. See Chapter 6, Machine Check Architecture for more information.
13.15.2 Errors and Recommended Error Handling

13.15.2.1 CRC Errors on HyperTransport Links AMD NPT Family 0Fh Processors can be programmed to generate sync flood when a CRC error is detected on a HyperTransport link. Sync flooding is enabled with Function 0, Offset 84h, A4h, or C4h, CRCFloodEn bit. 13.15.2.2 I/O Errors Target aborts on reads and nonposted writes are indicated by ERROR=1 and NXA=0 in their response packets in the HyperTransport protocol. If detected by the processor, the recommended error handling mechanism is machine check exception. The BIOS should set Function 3, Offset 40h, TgtAbrtEn bit. Master aborts on reads and nonposted writes are indicated by ERROR=1 and NXA=1 in their response packets in the HyperTransport protocol. If detected by the processor, the recommended error handling mechanism is machine check exception. The BIOS should set Function 3, Offset 40h, MstrAbrtEn bit. It is also recommended to set DisPciCfgCpuErrRsp and CpuErrDis bits in Function 3, Offset 44h. Error status is not returned to the processor for target and master aborts on posted writes from the processor to I/O devices, and the I/O device should respond. Error handling is chipset dependent in the case of target and master aborts on posted writes, PERR and SERR. Chipsets can implement sync flooding, interrupts, and other methods for error handling. 13.15.2.3 Machine Check Architecture Errors It is recommended that all uncorrectable errors detected by the MCA banks in AMD NPT Family 0Fh Processors with the exception of uncorrectable ECC errors detected by the NB are handled by the machine check exceptions. 364 BIOS Checklist Chapter 13
32559
Rev. 3.08
July 2007
13.15.2.4 Uncorrectable ECC Errors Detected by the NB In AMD Opteron multiprocessor system, sync flooding is a recommended response to uncorrectable ECC errors detected by the NB on reads from the DRAM. BIOS must take the following steps: 1. Initialize DRAM on all processors in the system (see Chapter 5, DRAM Configuration). 2. Enable ECC checking in all DRAM controllers (Function 3, Offset 44h, bit EccEn), and initialize every DRAM location. ECC errors generated during DRAM initialization should be ignored. 3. Enable sync flooding on uncorrectable ECC errors (Function 3, Offset 44h, bit SyncOnUcEccEn). 4. Enable uncorrectable ECC reporting (Function 3, Offset 40h, bit UnCorrEccEn). 5. Enable the I/O hub to generate a warm reset when it detects sync flood packets on the HyperTransport link. 6. Determine the type of reset shortly after boot (Function 0, Offset 6Ch, ColdRstDet bit). If the reset was warm, check the LinkFail bit (Function 0, Offset 84h, A4h, or C4h) for every connected link (see HyperTransport Link Detection on page 353). If the LinkFail bit is set, sync flood packets were detected by the link before the warm reset. The LinkFail bit should be cleared by writing a 1, and the BIOS should determine the error source. Uncorrectable ECC errors detected on DRAM reads are reported in the MCA NB reporting bank. MCA NB status and address registers (Function 3, Offsets 48h, 4Ch, 50h, and 54h) should be analyzed to determine the DIMM where the error originated. 13.15.2.5 Watchdog Timeout Errors In AMD Opteron and AMD Athlon 64 Processor systems sync flooding is the recommend response to watchdog timeout errors. BIOS must take the following steps: 1. Enable sync flooding on watchdog timeout errors (Function 3, Offset 44h, bit SyncOnWdogEn). 2. Enable watchdog timeout reporting (Function 3, Offset 40h, bit WchDogTmrEn). 3. Enable the I/O hub to generate a warm reset when it detects sync flood packets on the HyperTransport link. 4. Determine the type of reset shortly after boot (Function 0, Offset 6Ch, ColdRstDet bit). If the reset was warm, check the LinkFail bit (Function 0, Offset 84h, A4h, or C4h) for every connected link (see HyperTransport Link Detection on page 353). If the LinkFail bit is set, sync flood packets were detected by the link before the warm reset. The LinkFail bit should be cleared by writing a 1, and the BIOS should determine the error source. Watchdog timeout errors are reported in the MCA NB reporting bank. BIOS should not clear the MCi_Status registers when a watchdog timeout is detected. 13.15.2.6 Correctable ECC Errors Detected by the NB System software may optionally detect and log large numbers of correctable ECC errors detected by the NB on reads from the DRAM, with minimal overhead. A programmable error threshold counter causes an interrupt when the counter overflows, indicating a measure of general DRAM health to the Chapter 13 BIOS Checklist 365
32559
Rev. 3.08 July 2007
system software. See MC4_MISCDRAM Errors Threshold Register on page 229. Note, the counter also counts uncorrectable ECC errors detected by the NB on reads from the DRAM. The count threshold interrupt may be superceded by an MCA exception if MCA exceptions are enabled for the same errors. The system software must take the following steps to initialize thresholding: 1. Check the Locked bit (bit 61) in MC4_MISC to determine if the error thresholding mechanism is already in use. 2. Initialize the error threshold count and interrupt type in MC4_MISC. 3. If APIC interrupts are used to notify system software then program APIC register offset 500h accordingly. See Threshold Count Interrupt 0 Local Vector Table Entry on page 257.
13.16 Cache Initialization For General Storage During Boot

Prior to initializing the DRAM controller for system memory, the processor cache subsystem may be initialized for use by BIOS as general storage. The following steps should be taken to set up the cache in the boot core of the BSP to supply 64K bytes of memory, including a stack: Initialize the DRAM Base/Limit Registers (Function 1, Offsets [7C:40]) to a temporary value that allocates DRAM to each of the processors. BASE=The temporary base address of the DRAM range of the BSP. Clear all the MTRRs. Set MSRC001_0010 [MtrrFixDramEn, MtrrFixDramModEn, and MtrrVarDramEn]. Enable cache through CR0. Temporarily disable cache fill probes on the BSP by programming DisFillP = 1 (Function 0, offset 68h). Use MTRRs to specify the write-back attribute (WB) to a 64 kilobyte portion of the memory map, starting at BASE. Read exactly 64 kilobytes of memory starting at BASE, using the REP MOV instruction. Restore DisFillP (Function 0, offset 68h). Set SS=BASE/16. Set SP=BASE+64K-4.
If DRAM training is used in conjunction with the cache for general storage then the storage size must be reduced and the address space of the storage chosen to meet the special requirements of the DRAM training software. The general requirement for DDR training is that 256 cache lines, corresponding to L1 cache tag indexes of 00h - FFh, are used and must be reserved for this purpose. See AMD64 Architecture Programmers Manual, Volume 2: System Programming, for more details 366 BIOS Checklist Chapter 13
32559
Rev. 3.08
July 2007
on L1 cache organization and function. One possible implementation example that meets this requirement is as follows: A general storage size of 48KB is used. The address space for the storage is from C4000h to CFFFFh.
Note: When using cache as general storage the BIOS must ensure that no data lines are evicted from the cache prior to the DRAM controller being initialized. The CLFLUSH and WBINVD must not be used before DRAM initialization when using cache as general storage.
13.17 Cache Testing and Programming

BIOS must correctly maintain the tag parity and data ECC enable bits.
13.18 Memory System Configuration Registers

Node configuration space contains registers that define RAM, PCI memory, and x86 I/O address paths. Address mapping registers are contained in function 1 of each node. Cold boot and reset sends code fetch and other addresses to the compatibility link of Node0. At cold boot and reset, these mapping registers are disabled. Once enabled, these registers distribute addresses according to register mappings, which include node and link routing. See Chapter 9, HyperTransport Technology Configuration and Enumeration. Thereafter, carefully set registers to reflect current needs for code fetch and memory access, i.e., do not map BIOS E0000h and F0000h space to DRAM until BIOS exits execution from the ROM chip. Otherwise, the system will hang. Likewise, memory-mapped I/O space should not be mapped above 1 Mbyte until BIOS exits execution from the ROM chip, or else the system will hang. Memory-mapped I/O (MMIO) and PCI I/O mappings must include address space needed to move option ROM code to RAM, as well as the required run-time resources. This could require setting and resetting the MMIO mappings during enumeration and device initialization. Configuration space access of a device with or without an option ROM is defined by the HyperTransport technology configuration mapping. Multihost bus multiprocessor systems require that MMIO mappings direct relevant addresses to the buses of devices through the node/link path to the bus. Application processors and the bootstrap processor in multiprocessor systems must be mapped separately with paths to target address space.
Chapter 13
BIOS Checklist
367
32559
Rev. 3.08 July 2007
13.19 Processor ID
ACPI 2.0 specifies that each processor is required to have a unique ProcessorID, that although arbitrary, must match its corresponding ACPI Processor ID field in the Processor Local APIC table.
13.20 XSDT Table

The Extended System Description Table (XSDT) table defined in ACPI 2.0 must be implemented in the BIOS to support 64-bit operating systems for AMD NPT Family 0Fh Processor based systems.
13.21 Detect Target Operating Mode Callback

The operating system notifies the BIOS what the expected operating mode is with the Detect Target Operating Mode callback (INT 15, function EC00h). Based on the target operating mode, the BIOS can enable or disable mode specific performance and functional optimizations that are not visible to system software. This callback does not change the operating mode; it only declares the target mode to the BIOS. It should be executed only once by the BSP before the first transition into long mode. The default operating mode assumed by the BIOS is Legacy Mode Target Only. If this is not the target operating mode, system software must execute this callback to change it before transitioning to long mode for the first time. If the target operating mode is Legacy Mode Target Only, the callback does not need to be executed. The Detect Target Operating Mode callback inputs are stored in the AX and BL registers. AX has a value of EC00h, selecting the Detect Target Operating Mode function. One of the following values in the BL register selects the operating mode: 01h Legacy Mode Target Only. All enabled processors will operate in legacy mode only. 02h Long Mode Target Only. All enabled processor will switch into long mode once. 03h Mixed Mode Target. Processors may switch between legacy mode and long mode, or the preferred mode for system software is unknown. This value instructs the BIOS to use settings that are valid in all modes. All other values are reserved. The Detect Target Operating Mode callback outputs are stored in the AH register and CF (carry flag in the EFLAGS register), and the values of other registers are not modified. The following output values are possible: AH = 00h and CF = 0, if the callback is implemented and the value in BL is supported. AH = 00h and CF = 1, if the callback is implemented and the value in BL is reserved. This indicates an error; the target operating mode is set to Legacy Mode Target Only.
368
BIOS Checklist
Chapter 13
32559
Rev. 3.08
July 2007
AH = 86h and CF = 1, if the callback is not supported.
13.22 SMM Issues

The processor includes support for system management mode (SMM). The functionality of the AMD NPT Family 0Fh Processor SMM is a superset of the Pentium processor functionality. AMD NPT Family 0Fh Processors implement the SMM remapping and control registers as model-specific registers on the CPU. Implement the SMM state-save area in a manner compatible with the description of SMM found in Chapter 8, System Management Mode (SMM).
Program the model-specific registers to set the SMM memory base, local address, destination address, memory type, size and control, etc. See Chapter 8, System Management Mode (SMM).
Chapter 13
BIOS Checklist
369
32559
Rev. 3.08 July 2007
370
BIOS Checklist
Chapter 13
32559
Rev. 3.08
July 2007
14
Processor Configuration Registers
This chapter includes descriptions of two types of model-specific registers, as follows: General model-specific registers (see page 371) AMD NPT Family 0Fh Processor model-specific registers (see page 394)
14.1
General Model-Specific Registers
Table 78 is a listing of the general model-specific registers supported by the processor, presented in ascending hexadecimal address order. Register descriptions follow the table, organized according to the following functions: System software (see page 373) Memory typing (see page 375) APIC (see page 392) Machine check architecture (see page 206) Software debug (see page 393) Performance monitoring (see page 394) General MSRs
Register Name TSC APIC_BASE EBL_CR_POWERON MTRRcap SYSENTER_CS SYSENTER_ESP SYSENTER_EIP MCG_CAP MCG_STATUS MCG_CTL Description page 394 page 392 page 392 page 375 page 373 page 374 page 374 page 207 page 207 page 208
Table 78.
Address 0010h 001Bh 002Ah 00FEh 0174h 0175h 0176h 0179h 017Ah 017Bh
Chapter 14
371
32559
Rev. 3.08 July 2007
Table 78.
Address 01D9h 01DBh 01DCh 01DDh 01DEh
General MSRs (Continued)

Register Name DebugCtl LastBranchFromIP LastBranchToIP LastExceptionFromIP LastExceptionToIP MTRRphysBase[7:0] MTRRphysMask[7:0] MTRRfix64K_00000 MTRRfix16K_80000 MTRRfix16K_A0000 MTRRfix4K_C0000 MTRRfix4K_C8000 MTRRfix4K_D0000 MTRRfix4K_D8000 MTRRfix4K_E0000 MTRRfix4K_E8000 MTRRfix4K_F0000 MTRRfix4K_F8000 PAT MTRRdefType MC0_CTL MC0_STATUS MC0_ADDR MC0_MISC MC1_CTL MC1_STATUS MC1_ADDR MC1_MISC Description page 393 page 393 page 393 page 393 page 393 page 376 page 376 page 377 page 378 page 379 page 381 page 382 page 383 page 384 page 385 page 386 page 388 page 389 page 390 page 391 page 214 page 215 page 218 page 213 page 219 page 220 page 220 page 213
0200020Eh Even 0201020Fh Odd 0250h 0258h 0259h 0268h 0269h 026Ah 026Bh 026Ch 026Dh 026Eh 026Fh 0277h 02FFh 0400h 0401h 0402h 0403h, 0404h 0405h 0406h 0407h
372
Chapter 14
32559
Rev. 3.08
July 2007
Table 78.
Address 0408h 0409h 040Ah 040Bh 040Ch 040Dh 040Eh 040Fh 0410h 0411h 0412h 0413h
General MSRs (Continued)

Register Name MC2_CTL MC2_STATUS MC2_ADDR MC2_MISC MC3_CTL MC3_STATUS MC3_ADDR MC3_MISC MC4_CTL MC4_STATUS MC4_ADDR MC4_MISC Description page 222 page 225 page 226 page 213 page 227 page 228 page 228 page 213 page 228 page 229 page 229 page 229
14.1.1
14.1.1.1
System Software Registers

SYSENTER_CS Register
This register contains the code segment selector used by the SYSENTER and SYSEXIT instructions. See the SYSENTER and SYSEXIT section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. SYSENTER_CS Register
63 reserved 31 reserved 16 15 SYSENTER_CS 0
MSR 0174h
32
Bit 6332 3116 150
Mnemonic reserved reserved SYSENTER_CS
Function RAZ SBZ SYSENTER/SYSEXIT code segment selector
R/W R R/W R/W
Reset 0 0 0
Chapter 14
373
32559
Rev. 3.08 July 2007
Field Descriptions SYSENTER/SYSEXIT Code Segment Selector (SYSENTER_CS)Bits 150. 14.1.1.2 SYSENTER_ESP Register
This register contains the stack pointer used by the SYSENTER and SYSEXIT instructions. See the SYSENTER and SYSEXIT section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. SYSENTER_ESP Register
63 reserved 31 SYSENTER_ESP 0
MSR 0175h
32
Bit 6332 310
Mnemonic reserved SYSENTER_ESP
Function RAZ SYSENTER/SYSEXIT stack pointer
R/W R R/W
Reset 0 0
Field Descriptions SYSENTER/SYSEXIT Stack Pointer (SYSENTER_ESP)Bits 310. 14.1.1.3 SYSENTER_EIP Register
This register contains the instruction pointer used by the SYSENTER and SYSEXIT instructions. See the SYSENTER and SYSEXIT section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. SYSENTER_EIP Register
63 reserved 31 SYSENTER_EIP 0
MSR 0176h
32
374
Chapter 14
32559
Rev. 3.08
July 2007
Bit 6332 310
Mnemonic reserved SYSENTER_EIP
Function RAZ SYSENTER/SYSEXIT instruction pointer
R/W R R/W
Reset 0 0
Field Descriptions SYSENTER/SYSEXIT Instruction Pointer (SYSENTER_EIP)Bits 310.
14.1.2
14.1.2.1
Memory Typing Registers

MTRRcap Register
This is a read-only register that returns information about the processors MTRR capabilities. See the Using MTRRs section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. The MTRRcap register is a read-only status register. Attempting to modify this register results in a #GP(0). MTRRcap Register
63 reserved 31 reserved 11 10 9 MtrrCapWc reserved 8 MtrrCapFix 7 MtrrCapVCnt 0
MSR 00FEh
32
Bit 6311 10 9 8 70
Mnemonic reserved MtrrCapWc reserved MtrrCapFix MtrrCapVCnt
Function RAZ Write-combining memory type RAZ Fixed range register Variable range registers count
R/W R R R R R
Field Descriptions Variable Range Registers Count (MtrrCapVCnt)Bits 70. Indicates number of variable range registers. Fixed Range Registers (MtrrCapFix)Bit 8. Indicates fixed range register capability.
Chapter 14
375
32559
Rev. 3.08 July 2007
Write-Combining Memory Type (MtrrCapWc)Bit 10. Indicates write-combining memory type capability. 14.1.2.2 MTRRphysBasei Registers
These registers define the base address and memory type for each of the variable MTRRs. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Attempting to modify reserved bits in MTRRphysBasei results in a #GP(0). MTRRphysBase07 Registers
63 reserved 31 PhyBase 190 12 11 reserved 8 7 Type
MSRs 0200h, 0202h, 0204h, 0206h, 0208h, 020Ah, 020Ch, 020Eh

40 39 PhyBase 2720 0 32
Bit 6340 3912 118 70
Mnemonic reserved PhyBase reserved Type
Function MBZ Base address MBZ Memory type
R/W R/W R/W
Reset 0 U 0 U
Field Descriptions Memory Type (Type)Bits 70. Specifies memory type for this memory range. Base Address (PhysBase)Bits 3912. Specifies base address for this memory range 14.1.2.3 MTRRphysMaski Registers
These registers define the address mask and valid bit for each of the variable MTRRs. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Attempting to modify reserved bits in MTRRphysMaski results in a #GP(0).
376
Chapter 14
32559
Rev. 3.08
July 2007
MTRRphysMask07 Registers
63 reserved 31 PhyMask 190
MSRs 0201h, 0203h, 0205, 0207h, 0209h, 020Bh, 020Dh, 020Fh

40 39 PhyMask 2720 12 11 10 Valid reserved 0 32
Bit 6340 3912 11 100
Mnemonic reserved PhysMask Valid reserved
Function MBZ Address mask MTRR is valid MBZ
R/W R/W R/W
Reset 0 U 0 0
Field Descriptions MTRR is Valid (Valid)Bit 11. Indicated MTRR is valid. Address Mask (PhysMask)Bits 3912. Specifies the address mask for this memory range. 14.1.2.4 MTRRfix64K_00000 Register
This register controls the memory types for the first 512 Kbyte of physical memory. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). MTRRfix64K_00000 Register
63 56 55 48 47 40 39
MSR 0250h
32
MtrrFix64k7xxxxMemType 31
MtrrFix64k6xxxxMemType
MtrrFix64k5xxxxMemType 8
MtrrFix64k4xxxxMemType 7 0
24 23
16 15
Bit 6356 5548 4740
Mnemonic MtrrFix64k7xxxxMemType MtrrFix64k6xxxxMemType MtrrFix64k5xxxxMemType
Function Memory type address 700007FFFFh Memory type address 600006FFFFh Memory type address 500005FFFFh
R/W R/W R/W R/W
Reset U U U
* Memory types must be set to values consistent with system hardware.
Chapter 14
377
32559
Rev. 3.08 July 2007
Bit 3932 3124 2316 158 70
Mnemonic MtrrFix64k4xxxxMemType MtrrFix64k3xxxxMemType MtrrFix64k2xxxxMemType MtrrFix64k1xxxxMemType MtrrFix64k0xxxxMemType
Function Memory type address 400004FFFFh Memory type address 300003FFFFh Memory type address 200002FFFFh Memory type address 100001FFFFh Memory type address 000000FFFFh
Reset U U U U U
Field Descriptions Memory Type Address 000000FFFFh (MtrrFix64k0xxxxMemType)Bits 70. Memory type for physical address 000000FFFFh. Memory Type Address 100001FFFFh (MtrrFix64k1xxxxMemType)Bits 158. Memory type for physical address 100001FFFFh. Memory Type Address 200002FFFFh (MtrrFix64k2xxxxMemType)Bits 2316. Memory type for physical address 200002FFFFh. Memory Type Address 300003FFFFh (MtrrFix64k3xxxxMemType)Bits 3124. Memory type for physical address 300003FFFFh. Memory Type Address 400004FFFFh (MtrrFix64k4xxxxMemType)Bits 3932. Memory type for physical address 400004FFFFh. Memory Type Address 500005FFFFh (MtrrFix64k5xxxxMemType)Bits 4740. Memory type for physical address 500005FFFFh. Memory Type Address 600006FFFFh (MtrrFix64k6xxxxMemType)Bits 5548. Memory type for physical address 600006FFFFh. Memory Type Address 700007FFFFh (MtrrFix64k7xxxxMemType)Bits 6356. Memory type for physical address 700007FFFFh. 14.1.2.5 MTRRfix16K_80000 Register
This register controls the memory types for physical memory addresses 800009FFFFh. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). MTRRfix16K_80000 Register
63 56 55 48 47 40 39
MSR 0258h
32
MtrrFix16k9CxxxMemType
MtrrFix16k98xxxMemType
378
Chapter 14
32559
Rev. 3.08
July 2007
16 15 8 7 0
31
24 23
MtrrFix16k8CxxxMemType
Bit 6356 5548 4740 3932 3124 2316 158 70
Mnemonic MtrrFix16k98xxxMemType MtrrFix16k94xxxMemType MtrrFix16k90xxxMemType MtrrFix16k88xxxMemType MtrrFix16k84xxxMemType MtrrFix16k80xxxMemType
Function Memory type address 980009BFFFh Memory type address 9400097FFFh Memory type address 9000093FFFh Memory type address 880008BFFFh Memory type address 8400087FFFh Memory type address 8000083FFFh
R/W R/W R/W R/W R/W R/W R/W R/W R/W
Reset U U U U U U U U
MtrrFix16k9CxxxMemType Memory type address 9C0009FFFFh
MtrrFix16k8CxxxMemType Memory type address 8C0008FFFFh
Field Descriptions Memory Type Address 8000083FFFh (MtrrFix16k80xxxMemType)Bits 70. Memory type for physical address 8000083FFFh. Memory Type Address 8400087FFFh (MtrrFix16k84xxxMemType)Bits 158. Memory type for physical address 8400087FFFh. Memory Type Address 880008BFFFh (MtrrFix16k88xxxMemType)Bits 2316. Memory type for physical address 880008BFFFh. Memory Type Address 8C0008FFFFh (MtrrFix16k8CxxxMemType)Bits 3124. Memory type for physical address 8C0008FFFFh. Memory Type Address 9000093FFFh (MtrrFix16k90xxxMemType)Bits 3932. Memory type for physical address 9000093FFFh. Memory Type Address 9400097FFFh (MtrrFix16k94xxxMemType)Bits 4740. Memory type for physical address 9400097FFFh. Memory Type Address 980009BFFFh (MtrrFix16k98xxxMemType)Bits 5548. Memory type for physical address 980009BFFFh. Memory Type Address 9C0009FFFFh (MtrrFix16k9CxxxMemType)Bits 6356. Memory type for physical address 9C0009FFFFh. 14.1.2.6 MTRRfix16K_A0000 Register
This register controls the memory types for physical memory addresses A0000BFFFFh. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). Chapter 14 Processor Configuration Registers 379
32559
Rev. 3.08 July 2007
MTRRfix16K_A0000 Register
63 56 55 48 47 40 39
MSR 0259h
32
MtrrFix16kBCxxxMemType 31
MtrrFix16kB8xxxMemType
MtrrFix16kB4xxxMemType 8
MtrrFix16kB0xxxMemType 7 0
24 23
16 15
MtrrFix16kACxxxMemType
MtrrFix16kA8xxxMemType
Bit 6356 5548 4740 3932 3124 2316 158 70
Mnemonic
Function
MtrrFix16kBCxxxMemType Memory type address BC000BFFFFh MtrrFix16kB8xxxMemType Memory type address B8000BBFFFh MtrrFix16kB4xxxMemType Memory type address B4000B7FFFh MtrrFix16kB0xxxMemType Memory type address B0000B3FFFh MtrrFix16kACxxxMemType Memory type address AC000AFFFFh MtrrFix16kA8xxxMemType Memory type address A8000ABFFFh MtrrFix16kA4xxxMemType Memory type address A4000A7FFFh MtrrFix16kA0xxxMemType Memory type address A0000A3FFFh
Field Descriptions Memory Type Address A0000A3FFFh (MtrrFix16kA0xxxMemType)Bits 70. Memory type for physical address A0000A3FFFh. Memory Type Address A4000A7FFFh (MtrrFix16kA4xxxMemType)Bits 158. Memory type for physical address A4000A7FFFh. Memory Type Address A8000ABFFFh (MtrrFix16kA8xxxMemType)Bits 2316. Memory type for physical address A8000ABFFFh. Memory Type Address AC000AFFFFh (MtrrFix16kACxxxMemType)Bits 3124. Memory type for physical address AC000AFFFFh. Memory Type Address B0000B3FFFh (MtrrFix16kB0xxxMemType)Bits 3932. Memory type for physical address B0000B3FFFh. Memory Type Address B4000B7FFFh (MtrrFix16kB4xxxMemType)Bits 4740. Memory type for physical address B4000B7FFFh. Memory Type Address B8000BBFFFh (MtrrFix16kB8xxxMemType)Bits 5548. Memory type for physical address B8000BBFFFh. Memory Type Address BC000BFFFFh (MtrrFix16kBCxxxMemType)Bits 6356. Memory type for physical address BC000BFFFFh.
380
Chapter 14
32559
Rev. 3.08
July 2007
14.1.2.7
MTRRfix4K_C0000 Register
This register controls the memory types for physical memory addresses C0000C7FFFh. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). MTRRfix4K_C0000 Register
63 56 55 48 47 40 39
MSR 0268h
32
MtrrFix4kC7xxxMemType 31
MtrrFix4kC6xxxMemType
MtrrFix4kC5xxxMemType 8
MtrrFix4kC4xxxMemType 7 0
24 23
16 15
Bit 6356 5548 4740 3932 3124 2316 158 70
Mnemonic MtrrFix4kC7xxxMemType MtrrFix4kC6xxxMemType MtrrFix4kC5xxxMemType MtrrFix4kC4xxxMemType MtrrFix4kC3xxxMemType MtrrFix4kC2xxxMemType MtrrFix4kC1xxxMemType MtrrFix4kC0xxxMemType
Function Memory type address C7000C7FFFh Memory type address C6000C6FFFh Memory type address C5000C5FFFh Memory type address C4000C4FFFh Memory type address C3000C3FFFh Memory type address C2000C2FFFh Memory type address C1000C1FFFh Memory type address C0000C0FFFh
Field Descriptions Memory Type Address C0000C0FFFh (MtrrFix4kC0xxxMemType)Bits 70. Memory type for physical address C0000C0FFFh. Memory Type Address C1000C1FFFh (MtrrFix4kC1xxxMemType)Bits 158. Memory type for physical address C1000C1FFFh. Memory Type Address C2000C2FFFh (MtrrFix4kC2xxxMemType)Bits 2316. Memory type for physical address C2000C2FFFh. Memory Type Address C3000C3FFFh (MtrrFix4kC3xxxMemType)Bits 3124. Memory type for physical address C3000C3FFFh. Memory Type Address C40000C4FFFh (MtrrFix4kC4xxxMemType)Bits 3932. Memory type for physical address C4000C4FFFh.
Chapter 14
381
32559
Rev. 3.08 July 2007
Memory Type Address C5000C5FFFh (MtrrFix4kC5xxxMemType)Bits 4740. Memory type for physical address C5000C5FFFh. Memory Type Address C6000C6FFFh (MtrrFix4kC6xxxMemType)Bits 5548. Memory type for physical address C6000C6FFFh. Memory Type Address C7000C7FFFh (MtrrFix4kC7xxxMemType)Bits 6356. Memory type for physical address C7000C7FFFh. 14.1.2.8 MTRRfix4K_C8000 Register
This register controls the memory types for physical memory addresses C8000CFFFFh. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). MTRRfix4K_C8000 Register
63 56 55 48 47 40 39
MSR 0269h
32
MtrrFix4kCFxxxMemType 31
MtrrFix4kCExxxMemType
MtrrFix4kCDxxxMemType 8
MtrrFix4kCCxxxMemType 7 0
24 23
16 15
MtrrFix4kCBxxxMemType
MtrrFix4kCAxxxMemType
Bit 6356 5548 4740 3932 3124 2316 158 70
Mnemonic MtrrFix4kCFxxxMemType MtrrFix4kCExxxMemType MtrrFix4kCDxxxMemType MtrrFix4kCCxxxMemType MtrrFix4kCBxxxMemType MtrrFix4kCAxxxMemType MtrrFix4kC9xxxMemType MtrrFix4kC8xxxMemType
Function Memory type for address CF000CFFFFh Memory type for address CE000CEFFFh Memory type for address CD000CDFFFh Memory type for address CC000CCFFFh Memory type for address CB000CBFFFh Memory type for address CA000CAFFFh Memory type for address C9000C9FFFh Memory type for address C8000C8FFFh
Field Descriptions Memory Type Address C8000C8FFFh (MtrrFix4kC8xxxMemType)Bits 70. Memory type for physical address C8000C8FFFh. Memory Type Address C9000C9FFFh (MtrrFix4kC9xxxMemType)Bits 158. Memory type for physical address C9000C9FFFh.
382
Chapter 14
32559
Rev. 3.08
July 2007
Memory Type Address CA000CAFFFh (MtrrFix4kCAxxxMemType)Bits 2316. Memory type for physical address CA000CAFFFh. Memory Type Address CB000CBFFFh (MtrrFix4kCBxxxMemType)Bits 3124. Memory type for physical address CB000CBFFFh. Memory Type Address CC0000CCFFFh (MtrrFix4kCCxxxMemType)Bits 3932. Memory type for physical address CC000CCFFFh. Memory Type Address CD000CDFFFh (MtrrFix4kCDxxxMemType)Bits 4740. Memory type for physical address CD000CDFFFh. Memory Type Address CE000CEFFFh (MtrrFix4kCExxxMemType)Bits 5548. Memory type for physical address CE000CEFFFh. Memory Type Address CF000CFFFFh (MtrrFix4kCFxxxMemType)Bits 6356. Memory type for physical address CF000CFFFFh. 14.1.2.9 MTRRfix4K_D0000 Register
This register controls the memory types for physical memory addresses D0000D7FFF. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). MTRRfix4K_D0000 Register
63 56 55 48 47 40 39
MSR 026Ah
32
MtrrFix4kD7xxxMemType 31
MtrrFix4kD6xxxMemType
MtrrFix4kD5xxxMemType 8
MtrrFix4kD4xxxMemType 7 0
24 23
16 15
Bit 6356 5548 4740 3932 3124 2316 158 70
Mnemonic MtrrFix4kD7xxxMemType MtrrFix4kD6xxxMemType MtrrFix4kD5xxxMemType MtrrFix4kD4xxxMemType MtrrFix4kD3xxxMemType MtrrFix4kD2xxxMemType MtrrFix4kD1xxxMemType MtrrFix4kD0xxxMemType
Function Memory type address D7000D7FFFh Memory type address D6000D6FFFh Memory type address D5000D5FFFh Memory type address D4000D4FFFh Memory type address D3000D3FFFh Memory type address D2000D2FFFh Memory type address D1000D1FFFh Memory type address D0000D0FFFh
Chapter 14
383
32559
Rev. 3.08 July 2007
Field Descriptions Memory Type Address D0000D0FFFh (MtrrFix4kD0xxxMemType)Bits 70. Memory type for physical address D0000D0FFFh. Memory Type Address D1000D1FFFh (MtrrFix4kD1xxxMemType)Bits 158. Memory type for physical address D1000D1FFFh. Memory Type Address D2000D2FFFh (MtrrFix4kD2xxxMemType)Bits 2316. Memory type for physical address D2000D2FFFh. Memory Type Address D3000D3FFFh (MtrrFix4kD3xxxMemType)Bits 3124. Memory type for physical address D3000D3FFFh. Memory Type Address D40000D4FFFh (MtrrFix4kD4xxxMemType)Bits 3932. Memory type for physical address D4000D4FFFh. Memory Type Address D5000D5FFFh (MtrrFix4kD5xxxMemType)Bits 4740. Memory type for physical address D5000D5FFFh. Memory Type Address D6000D6FFFh (MtrrFix4kD6xxxMemType)Bits 5548. Memory type for physical address D6000D6FFFh. Memory Type Address D7000D7FFFh (MtrrFix4kD7xxxMemType)Bits 6356. Memory type for physical address D7000D7FFFh. 14.1.2.10 MTRRfix4K_D8000 Register This register controls the memory types for physical memory addresses D8000DFFFFh. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). MTRRfix4K_D8000 Register
63 56 55 48 47 40 39
MSR 026Bh
32
MtrrFix4kDFxxxMemType 31
MtrrFix4kDExxxMemType
MtrrFix4kDDxxxMemType 8
MtrrFix4kDCxxxMemType 7 0
24 23
16 15
MtrrFix4kDBxxxMemType
MtrrFix4kDAxxxMemType
Bit 6356 5548
Mnemonic MtrrFix4kDFxxxMemType MtrrFix4kDExxxMemType
Function Memory type address DF000DFFFFh Memory type address DE000DEFFFh
R/W R/W R/W
Reset U U
384
Chapter 14
32559
Rev. 3.08
July 2007
Bit 4740 3932 3124 2316 158 70
Mnemonic MtrrFix4kDDxxxMemType MtrrFix4kDCxxxMemType MtrrFix4kDBxxxMemType MtrrFix4kDAxxxMemType MtrrFix4kD9xxxMemType MtrrFix4kD8xxxMemType
Function Memory type address DD000DDFFFh Memory type address DC000DCFFFh Memory type address DB000DBFFFh Memory type address DA000DAFFFh Memory type address D9000D9FFFh Memory type address D8000D8FFFh
R/W R/W R/W R/W R/W R/W R/W
Reset U U U U U U
Field Descriptions Memory Type Address D8000D8FFFh (MtrrFix4kD8xxxMemType)Bits 70. Memory type for physical address D8000D8FFFh. Memory Type Address D9000D9FFFh (MtrrFix4kD9xxxMemType)Bits 158. Memory type for physical address D9000D9FFFh. Memory Type Address DA000DAFFFh (MtrrFix4kDAxxxMemType)Bits 2316. Memory type for physical address DA000DAFFFh. Memory Type Address DB000DBFFFh (MtrrFix4kDBxxxMemType)Bits 3124. Memory type for physical address DB000DBFFFh. Memory Type Address DC0000DCFFFh (MtrrFix4kDCxxxMemType)Bits 3932. Memory type for physical address DC000DCFFFh. Memory Type Address DD000DDFFFh (MtrrFix4kDDxxxMemType)Bits 4740. Memory type for physical address DD000DDFFFh. Memory Type Address DE000DEFFFh (MtrrFix4kDExxxMemType)Bits 5548. Memory type for physical address DE000DEFFFh. Memory Type Address DF000DFFFFh (MtrrFix4kDFxxxMemType)Bits 6356. Memory type for physical address DF000DFFFFh. 14.1.2.11 MTRRfix4K_E0000 Register This register controls the memory types for physical memory addresses E0000E7FFFh. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). MTRRfix4K_E0000 Register
63 56 55 48 47 40 39
MSR 026Ch
32
MtrrFix4kE7xxxMemType
Chapter 14
385
31 24 23 16 15 8 7
32559
Rev. 3.08 July 2007
Bit 6356 5548 4740 3932 3124 2316 158 70
Mnemonic MtrrFix4kE7xxxMemType MtrrFix4kE6xxxMemType MtrrFix4kE5xxxMemType MtrrFix4kE4xxxMemType MtrrFix4kE3xxxMemType MtrrFix4kE2xxxMemType MtrrFix4kE1xxxMemType MtrrFix4kE0xxxMemType
Function Memory type for physical addr E7000E7FFFh Memory type for physical addr E6000E6FFFh Memory type for physical addr E5000E5FFFh Memory type for physical addr E4000E4FFFh Memory type for physical addr E3000E3FFFh Memory type for physical addr E2000E2FFFh Memory type for physical addr E1000E1FFFh Memory type for physical addr E0000E0FFFh
Field Descriptions Memory Type Address E0000E0FFFh (MtrrFix4kE0xxxMemType)Bits 70. Memory type for physical address E0000E0FFFh. Memory Type Address E1000E1FFFh (MtrrFix4kE1xxxMemType)Bits 158. Memory type for physical address E1000E1FFFh. Memory Type Address E2000E2FFFh (MtrrFix4kE2xxxMemType)Bits 2316. Memory type for physical address E2000E2FFFh. Memory Type Address E3000E3FFFh (MtrrFix4kE3xxxMemType)Bits 3124. Memory type for physical address E3000E3FFFh. Memory Type Address E40000E4FFFh (MtrrFix4kE4xxxMemType)Bits 3932. Memory type for physical address E4000E4FFFh. Memory Type Address E5000E5FFFh (MtrrFix4kE5xxxMemType)Bits 4740. Memory type for physical address E5000E5FFFh. Memory Type Address E6000E6FFFh (MtrrFix4kE6xxxMemType)Bits 5548. Memory type for physical address E6000E6FFFh. Memory Type Address E7000E7FFFh (MtrrFix4kE7xxxMemType)Bits 6356. Memory type for physical address E7000E7FFFh. 14.1.2.12 MTRRfix4K_E8000 Register This register controls the memory types for physical memory addresses E8000EFFFFh. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). 386 Processor Configuration Registers Chapter 14
32559
Rev. 3.08
July 2007
MTRRfix4K_E8000 Register
63 56 55 48 47 40 39
MSR 026Dh
32
MtrrFix4kEFxxxMemType 31
MtrrFix4kEExxxMemType
MtrrFix4kEDxxxMemType 8
MtrrFix4kECxxxMemType 7 MtrrFix4kE8xxxMemType 0
24 23
16 15
MtrrFix4kEBxxxMemType
MtrrFix4kEAxxxMemType
Bit 6356 5548 4740 3932 3124 2316 158 70
Mnemonic MtrrFix4kEFxxxMemType MtrrFix4kEExxxMemType MtrrFix4kEDxxxMemType MtrrFix4kECxxxMemType MtrrFix4kEBxxxMemType MtrrFix4kEAxxxMemType MtrrFix4kE9xxxMemType MtrrFix4kE8xxxMemType
Function Memory type for physical addr EF000EFFFFh Memory type for physical addr EE000EEFFFh Memory type for physical addr ED000EDFFFh Memory type for physical addr EC000ECFFFh Memory type for physical addr EB000EBFFFh Memory type for physical addr EA000EAFFFh Memory type for physical addr E9000E9FFFh Memory type for physical addr E8000E8FFFh
Field Descriptions Memory Type Address E8000E8FFFh (MtrrFix4kE8xxxMemType)Bits 70. Memory type for physical address E8000E8FFFh. Memory Type Address E9000E9FFFh (MtrrFix4kE9xxxMemType)Bits 158. Memory type for physical address E9000E9FFFh. Memory Type Address EA000EAFFFh (MtrrFix4kEAxxxMemType)Bits 2316. Memory type for physical address EA000EAFFFh. Memory Type Address EB000EBFFFh (MtrrFix4kEBxxxMemType)Bits 3124. Memory type for physical address EB000EBFFFh. Memory Type Address EC0000ECFFFh (MtrrFix4kECxxxMemType)Bits 3932. Memory type for physical address EC000ECFFFh. Memory Type Address ED000EDFFFh (MtrrFix4kEDxxxMemType)Bits 4740. Memory type for physical address ED000EDFFFh. Memory Type Address EE000EEFFFh (MtrrFix4kEExxxMemType)Bits 5548. Memory type for physical address EE000EEFFFh. Memory Type Address EF000EFFFFh (MtrrFix4kEFxxxMemType)Bits 6356. Memory type for physical address EF000EFFFFh.
Chapter 14
387
32559
Rev. 3.08 July 2007
14.1.2.13 MTRRfix4K_F0000 Register This register controls the memory types for physical memory addresses F0000F7FFFh. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). MTRRfix4K_F0000 Register
63 56 55 48 47 40 39
MSR 026Eh
32
MtrrFix4kF7xxxMemType 31
MtrrFix4kF6xxxMemType
MtrrFix4kF5xxxMemType 8 7
MtrrFix4kF4xxxMemType 0 MtrrFix4kF0xxxMemType
24 23
16 15 MtrrFix4kF1xxxMemType
Bit 6356 5548 4740 3932 3124 2316 158 70
Mnemonic MtrrFix4kF7xxxMemType MtrrFix4kF6xxxMemType MtrrFix4kF5xxxMemType MtrrFix4kF4xxxMemType MtrrFix4kF3xxxMemType MtrrFix4kF2xxxMemType MtrrFix4kF1xxxMemType MtrrFix4kF0xxxMemType
Function Memory type address F7000F7FFFh Memory type address F6000F6FFFh Memory type address F5000F5FFFh Memory type address F4000F4FFFh Memory type address F3000F3FFFh Memory type address F2000F2FFFh Memory type address F1000F1FFFh Memory type address F0000F0FFFh
Field Descriptions Memory Type Address F0000F0FFFh (MtrrFix4kF0xxxMemType)Bits 70. Memory type for physical address F0000F0FFFh. Memory Type Address F1000F1FFFh (MtrrFix4kF1xxxMemType)Bits 158. Memory type for physical address F1000F1FFFh. Memory Type Address F2000F2FFFh (MtrrFix4kF2xxxMemType)Bits 2316. Memory type for physical address F2000F2FFFh. Memory Type Address F3000F3FFFh (MtrrFix4kF3xxxMemType)Bits 3124. Memory type for physical address F3000F3FFFh. Memory Type Address F40000F4FFFh (MtrrFix4kF4xxxMemType)Bits 3932. Memory type for physical address F4000F4FFFh.
388
Chapter 14
32559
Rev. 3.08
July 2007
Memory Type Address F5000F5FFFh (MtrrFix4kF5xxxMemType)Bits 4740. Memory type for physical address F5000F5FFFh. Memory Type Address F6000F6FFFh (MtrrFix4kF6xxxMemType)Bits 5548. Memory type for physical address F6000F6FFFh. Memory Type Address F7000F7FFFh (MtrrFix4kF7xxxMemType)Bits 6356. Memory type for physical address F7000F7FFFh. 14.1.2.14 MTRRfix4K_F8000 Register This register controls the memory types for physical memory addresses F8000FFFFFh. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). MTRRfix4K_F8000 Register
63 56 55 48 47 40 39
MSR 026Fh
32
MtrrFix4kFFxxxMemType 31
MtrrFix4kFExxxMemType
MtrrFix4kFDxxxMemType 8
MtrrFix4kFCxxxMemType 7 MtrrFix4kF8xxxMemType 0
24 23
16 15 MtrrFix4kF9xxxMemType
MtrrFix4kFBxxxMemType
MtrrFix4kFAxxxMemType
Bit 6356 5548 4740 3932 3124 2316 158 70
Mnemonic MtrrFix4kFFxxxMemType MtrrFix4kFExxxMemType MtrrFix4kFDxxxMemType MtrrFix4kFCxxxMemType MtrrFix4kFBxxxMemType MtrrFix4kFAxxxMemType MtrrFix4kF9xxxMemType MtrrFix4kF8xxxMemType
Function Memory type address FF000FFFFFh Memory type address FE000FEFFFh Memory type address FD000FDFFFh Memory type address FC000FCFFFh Memory type address FB000FBFFFh Memory type address FA000FAFFFh Memory type address F9000F9FFFh Memory type address F8000F8FFFh
Field Descriptions Memory Type Address F8000F8FFFh (MtrrFix4kF8xxxMemType)Bits 70. Memory type for physical address F8000F8FFFh. Memory Type Address F9000F9FFFh (MtrrFix4kF9xxxMemType)Bits 158. Memory type for physical address F9000F9FFFh.
Chapter 14
389
32559
Rev. 3.08 July 2007
Memory Type Address FA000FAFFFh (MtrrFix4kFAxxxMemType)Bits 2316. Memory type for physical address FA000FAFFFh. Memory Type Address FB000FBFFFh (MtrrFix4kFBxxxMemType)Bits 3124. Memory type for physical address FB000FBFFFh. Memory Type Address FC0000FCFFFh (MtrrFix4kFCxxxMemType)Bits 3932. Memory type for physical address FC000FCFFFh. Memory Type Address FD000FDFFFh (MtrrFix4kFDxxxMemType)Bits 4740. Memory type for physical address FD000FDFFFh. Memory Type Address FE000FEFFFh (MtrrFix4kFExxxMemType)Bits 5548. Memory type for physical address FE000FEFFFh. Memory Type Address FF000FFFFFh (MtrrFix4kFFxxxMemType)Bits 6356. Memory type for physical address FF000FFFFFh. 14.1.2.15 PAT Register This register contains the eight page attribute fields used for specifying memory types for pages. See the Page-Attribute Table Mechanism section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting any range to an undefined memory type results in a #GP(0). Setting any reserved bits results in a #GP(0). PAT Register
63 reserved 31 reserved 27 26 PA3 59 58 PA7 24 23 reserved 56 55 reserved 19 18 PA2 51 50 PA6 16 15 reserved 48 47 reserved 11 10 PA1 43 42 PA5 8 7 reserved 40 39 reserved 3 2 PA0
MSR 0277h
35 34 PA4 0 32
Bit 6359 5856 5551 5048 4743 4240 3935 3432 3127
Mnemonic reserved PA7 reserved PA6 reserved PA5 reserved PA4 reserved
Function MBZ Memory type for Page Attribute index 7 MBZ Memory type for Page Attribute index 6 MBZ Memory type for Page Attribute index 5 MBZ Memory type for Page Attribute index 4 MBZ
R/W R/W R/W R/W R/W
Reset 0 0 0 7 0 4 0 6 0
390
Chapter 14
32559
Rev. 3.08
July 2007
Bit 2624 2319 1816 1511 108 73 20
Mnemonic PA3 reserved PA2 reserved PA1 reserved PA0
Function Memory type for Page Attribute index 3 MBZ Memory type for Page Attribute index 2 MBZ Memory type for Page Attribute index 1 MBZ Memory type for Page Attribute index 0
R/W R/W R/W R/W R/W
Reset 0 0 7 0 4 0 6
14.1.2.16 MTRRdefType Register This register enables the MTRRs and defines the default memory type for memory not within one of the MTRR ranges. See the Memory Type Range Registers section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Setting MtrrDefMemType to an undefined memory type results in a #GP(0). Setting any reserved bits results in a #GP(0). MTRRdefType Register
63 reserved 31 12 11 10 9 MtrrDefTypeFixEn MtrrDefTypeEn 8 7 0
MSR 02FFh
32
reserved
reserved
MtrrDefMemType
Bit 6312 11 10 98 70
Mnemonic reserved MtrrDefTypeEn MtrrDefTypeFixEn reserved MtrrDefMemType
Function MBZ Enable MTRRs Enable Fixed Range MTRRs Default Memory Type
R/W R/W R/W R/W
Reset 0 0 0 0 0
Field Descriptions Default Memory Type (MtrrDefMemType)Bit 70. Enable Fixed Range MTRRs (MtrrDefTypeFixEn)Bit 10.
Chapter 14
391
32559
Rev. 3.08 July 2007
Enable MTRRs (MtrrDefTypeEn)Bit 11.
14.1.3
14.1.3.1
APIC Registers
APIC_BASE Register
This register enables APIC and defines the APIC base address. It also identifies the Boot Strap Processor (BSP). Setting any reserved bits results in a #GP(0). APIC_BASE Register
63 reserved 31 ApicBase 190 12 11 10 9 reserved ApicEn 8 BSP 7 reserved 40 39 ApicBase 2720 0
MSR 001Bh
32
Bit 6340 3912 11 109 8
Mnemonic reserved ApicBase ApicEn reserved BSP
Function MBZ APIC Base Address Enable APIC MBZ Boot Strap Processor
R/W R/W R/W R/W
Reset 0 00FEE00h 0 0 1 if uniprocessor or bsp of multiprocessor system 0
70
reserved
MBZ
* APIC configuration must be consistent with system hardware.
Field Descriptions Boot Strap Processor (Bsp)Bit 8. Enable APIC (ApicEn)Bit 11. APIC Base Address (ApicBase)Bits 3912. 14.1.3.2 EBL_CR_POWERON Register
This read-only register contains the APIC cluster ID. See the APIC Initialization section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. Attempting to write to this register results in a #GP(0). 392 Processor Configuration Registers Chapter 14
32559
Rev. 3.08
July 2007
EBL_CR_POWERON Register
63 reserved 31 18 17 16 15 ApicClusterID
MSR 002Ah
32
reserved
reserved
Bit 6318 1716 150
Mnemonic reserved ApicClusterID reserved
Function MBZ APIC Cluster ID MBZ
R/W R
Field Descriptions APIC Cluster ID (ApicClusterID)Bits 1716.
14.1.4
Software Debug Registers
AMD NPT Family 0Fh Processors incorporate extensive debug features. These include the following control, status, and control-transfer recording MSRs. DebugCtl Register (MSR 01D9h)Provides additional debug controls over control-transfer recording and single stepping, as well as external-breakpoint reporting and trace messages. LastBranchFromIP Register (MSR 01DBh)Loaded with the segment offset of the branch instruction. LastBranchToIP Register (MSR 01DCh)Holds the target rIP of the last branch that occurred before an exception or interrupt. LastExceptionFromIP Register (MSR 01DDh)Holds the source rIP of the last branch that occurred before the exception or interrupt. LastExceptionToIP Register (MSR 01DEh)Holds the target rIP of the last branch that occurred before the exception or interrupt.
For more detailed information on the use of these MSRs, see the Debug and Performance Resources section in Volume 2 of the AMD64 Architecture Programmers Manual.
Chapter 14
393
32559
Rev. 3.08 July 2007
14.1.5
14.1.5.1
Performance Monitoring Registers

TSC Register
The time-stamp counter (TSC) register maintains a running count of the number of internal processor clock cycles executed after a reset. It is incremented by 1 on each internal processor clock. When the TSC overflows its 64-bit range, it wraps around to 0. See the Time-Stamp Counter section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. TSC Register
63 PCLKS 6332 31 PCLKS 310 0
MSR 0010h
32
Bit 630
Mnemonic PCLKS
Function Running Clock Cycle Count
R/W R/W
Reset 0
Field Descriptions Processor Clock Cycles (PCLKS)Bits 630. Running count of number of internal processor clock cycles.
14.2
AMD NPT Family 0Fh Processor Model-Specific Registers
Table 79 is a listing of the model-specific registers supported by AMD NPT Family 0Fh Processors, presented in ascending hexadecimal address order. Register descriptions follow the table, organized according to the following functions: Features (see page 396) Identification (see page 403) Memory typing (see page 405) I/O range registers (see page 406) System call extension registers (see page 407) Segmentation (see page 409)
394
Chapter 14
32559
Rev. 3.08
July 2007
System management (see page 259) Power management (see page 411) AMD NPT Family 0Fh Processor MSRs
Register Name EFER STAR LSTAR CSTAR SF_MASK FS.Base GS.Base KernelGSbase PerfEvtSeli PerfCtri SYSCFG HWCR IORRBase[1:0] IORRMask[1:0] TOP_MEM TOP_MEM2 MANID NB_CFG HTC Thermal Control FIDVID_CTL FIDVID_STATUS MCi_CTL_MASK IOTRAP_ADDRi IOTRAP_CTL Description page 396 page 407 page 408 page 408 page 409 page 409 page 410 page 410 page 322 page 321 page 397 page 399 page 406 page 406 page 405 page 405 page 403 page 401 page 403 page 404 page 411 page 413 page 210 page 416 page 417
Table 79.
Address
C000_0080h C000_0081h C000_0082h C000_0083h C000_0084h C000_0100h C000_0101h C000_0102h C001_0000hC001_0003h C001_0004hC001_0007h C001_0010h C001_0015h C001_0016h, C001_0018h C001_0017h, C001_0019h C001_001Ah C001_001Dh C001_001Eh C001_001Fh C001_003Eh C001_003Fh C001_0041h C001_0042h C001_0044C001_0048h C001_0050C001_0053h C001_0054h
Chapter 14
395
32559
Rev. 3.08 July 2007
Table 79.
AMD NPT Family 0Fh Processor MSRs (Continued)

Interrupt Pending Message SMM_BASE SMM_ADDR SMM_MASK VM_CR VM_HSAVE_PA page 419 page 266 page 271 page 270 page 421 page 422
C001_0055h C001_0111h C001_0112h C001_0113h C001_0114 C001_0117
14.2.1
14.2.1.1
Feature Registers
EFER Register
This register controls which extended features are enabled. See the Extended Feature Enable Register (EFER) section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. The LMA bit is a read-only status bit. When writing EFER, the processor signals a #GP(0) if an attempt is made to change LMA from its previous value. BIOS should not modify this register unless it needs to use one of the features enabled by this register. EFER Register
63 reserved 31 reserved 15 14 13 12 11 10 9 reserved LMSLE
FFXSR
MSR C000_0080h
32
8 LME
7 reserved
0 SYSCALL
SVME
NXE
Bit 6315 14 13 12 11 10 9 8
Mnemonic reserved FFXSR LMSLE SVME NXE LMA reserved LME
Function MBZ Fast FXSAVE/FRSTOR Enable Long Mode Segment Limit Enable SVM Enable No-Execute Page Enable Long Mode Active MBZ Long Mode Enable
LMA
R/W R/W R/W R/W R/W R/W R R R/W
Reset 0 0 0 0 0 0 0 0
396
Chapter 14
32559
Rev. 3.08
July 2007
Bit 71 0
Mnemonic reserved SYSCALL
Function RAZ System Call Extension Enable
R/W R R/W
Reset 0 0
Field Descriptions System Call Extension Enable (SCE)Bit 0. Enables the system call extension. Long Mode Enable (LME)Bit 8. Enables the long mode feature. Long Mode Active (LMA)Bit 10. Indicates the long mode feature is active. No-Execute Page Enable (NXE)Bit 11. Enables the no-execute page feature. Long Mode Segment Limit Enable (LMSLE)Bit 13. Enables the long mode segment limit check mechanism. SVM Enable (SVME)Bit 12. Enables SVM features. Fast FXSAVE/FRSTOR Enable (FFXSE)Bit 14. Enables the fast FXSAVE/FRSTOR mechanism. A 64-bit operating system uses EDX bit 24 as returned by CPUID instruction standard function 1 to determine the presence of this feature before enabling it. This bit is set once by the operating system and its value is not changed afterwards. 14.2.1.2 SYSCFG Register
This register controls the system configuration. The MtrrFixDramModEn bit should be set to 1 during BIOS initialization of the fixed MTRRs, then cleared to 0 for operation. SYSCFG Register
63 reserved 31 23 22 21 20 19 18 17 16 15 Tom2ForceMemTypeWB MtrrFixDramModEn MtrrVarDramEn 12 11 10 9 8 7 5 4 0
MSR C001_0010h
32
MtrrFixDramEn
ChxToDirtyDis
SysUcLockEn
SetDirtyEnO
SetDirtyEnS
MtrrTom2En
SetDirtyEnE
reserved
reserved
SysVicLimit
SysAckLimit
Chapter 14
397
32559
Rev. 3.08 July 2007
Bit 6321 22 21 20 19 18 17 16 1511 10 9 8 75 40
Mnemonic reserved Tom2ForceMemTypeWB MtrrTom2En MtrrVarDramEn MtrrFixDramModEn MtrrFixDramEn SysUcLockEn ChxToDirtyDis reserved SetDirtyEnO SetDirtyEnS SetDirtyEnE SysVicLimit SysAckLimit
Function RAZ Top of Memory 2 Memory Type Write Back Top of Memory Address Register 2 Enable (reserved) Top of Memory Address Register and I/O Range Register Enable RdDram and WrDram Bits Modification Enable Fixed RdDram and WrDram Attributes Enable System Interface Lock Command Enable Change to Dirty Command Disable RAZ SharedToDirty Command for O->M State Transition Enable SharedToDirty Command for S->M State Transition Enable CleanToDirty Command for E->M State Transition Enable Outstanding Victim Bus Command Limit Outstanding Bus Command Limit
R/W R R/W R/W R/W R/W R/W R/W R/W R R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 1 0 0 1 1 0 000b
BIOS 0 0 1 0 1 1 0 1 1 0 000b
00001b 00001b
Field Descriptions Outstanding Bus Command Limit (SysAckLimit)Bit 40. Limits maximum number of outstanding bus commands. Outstanding Victim Bus Command Limit (SysVicLimit)Bit 75. Limits maximum number of outstanding victim bus commands. CleanToDirty Command for E->M State Transition Enable (SetDirtyEnE)Bit 8. Enables generating write probes when transitioning a cache line from Exclusive to Modified. SharedToDirty Command for S->M State Transition Enable (SetDirtyEnS)Bit 9. Enables generating write probes when transitioning a cache line from Shared to Modified. SharedToDirty Command for O->M State Transition Enable (SetDirtyEnO)Bit 10. Enables generating write probes when transitioning a cache line from Owned to Modified. Change to Dirty Command Disable (ChxToDirtyDis)Bit 16. Disables change to dirty commands, evicts line from DC instead. System Interface Lock Command Enable (SysUcLockEn)Bit 17. Enables lock commands on system interface. Fixed RdDram and WrDram Attributes Enable (MtrrFixDramEn)Bit 18. Enables fixed MTRR RdDram and WrDram attributes. RdDram and WrDram Bits Modification Enable (MtrrFixDramModEn)Bit 19. Enables modification of RdDram and WrDram bits in fixed MTRRS. 398 Processor Configuration Registers Chapter 14
32559
Rev. 3.08
July 2007
Top of Memory Address Register and I/O Range Register Enable (MtrrVarDramEn)Bit 20. Enables use of top of memory address register and the I/O range registers. Top of Memory Address Register 2 Enable (MtrrTom2En)Bit 21. Enables use of top of memory address register 2 (reserved). Top of Memory 2 Memory Type Write Back (Tom2ForceMemTypeWB)Bit 22. When this bit is set, the default memory type of memory between 4GB and TOM2 is write back instead of the memory type defined by MtrrDefMemType (MSR 02FFh bits 7-0). MtrrDefTypeEn (MSR 02FFh bit 11) must be set to one for this bit to have any effect. Both MTRR's and PAT can be used to override this memory type. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. 14.2.1.3 HWCR Register
This register controls the hardware configuration. Operating systems that maintain page tables in uncacheable memory (UC memory type) must set the TLBCACHEDIS bit to insure proper operation. BIOS and SMM handlers may use the WRAP32DIS bit to access physical memory above 4 Gbytes without switching into 64-bit mode. By setting WRAP32DIS to a 1 in conjunction with setting the expanded FS or GS base registers, BIOS can size all of physical memory from legacy mode. To do so, BIOS would write a >32 bit base to the FS or GS base register using WRMSR. Then it would address 2 Gbytes from one of those bases using normal memory reference instructions with a FS or GS override prefix. However, the INVLPG, FST, and SSE store instructions generate 32-bit addresses in legacy mode, regardless of the state of WRAP32DIS. The MCi_STATUS_WREN bit can be used to help debug Machine Check exception handlers. When the MCi_STATUS_WREN bit is set, privileged software can write non-zero values to the MCG_STATUS, MCi_STATUS, MCi_ADDR, and if implemented, MCi_MISC MSRs without generating exceptions, and then simulate a machine check using the INT 18 instruction. Setting a reserved bit in these MSRs does not generate an exception when this mode is enabled. However, setting a reserved bit may result in undefined behavior. HWCR Register
63 reserved
MSR C001_0015h
32
Chapter 14
399
31 30 29 24 23 19 18 17 16 15 14 13 12 11 MCi_STATUS_WREN RSMSPCYCDIS HLTXSPCYCEN SMISPCYCDIS WRAP32DIS 9 8 7
32559
Rev. 3.08 July 2007
4 INVD_WBINVD
3 TLBCACHEDIS
1 SLOWFENCE
IGNNE_EM
START_FID
reserved
reserved
Bit 6332 3130 2924 2322 21-20 19 18 17 16 15 14 13 12 119 8 7 6 5 4 3 2 1 0
Mnemonic reserved reserved START_FID reserved reserved reserved MCi_STATUS_WREN WRAP32DIS reserved SSEDIS RSMSPCYCDIS SMISPCYCDIS HLTXSPCYCEN reserved IGNNE_EM DISLOCK FFDIS reserved INVD_WBINVD TLBCACHEDIS reserved SLOWFENCE SMMLOCK
Function RAZ SBZ Startup FID Status SBZ RAZ SBZ MCi Status Write Enable 32-bit Address Wrap Disable SSE Instructions Disable Special Bus Cycle On RSM Disable Special Bus Cycle On SMI Disable Enable Special Bus Cycle On Exit From HLT SBZ IGNNE Port Emulation Enable Disable x86 LOCK prefix functionality TLB Flush Filter Disable SBZ INVD to WBINVD Conversion Cacheable Memory Disable SBZ Slow SFENCE Enable SMM Configuration Lock
R/W R R/W R R/W R R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Field Descriptions SMM Code Lock (SMMLOCK)Bit 0. Locks SMM configuration registers SMM_BASE, SMM_ADDR, SMM_MASK (all but SMM_MASK[3:2]), and HWCR[14:13, 0] by making them read-only. Slow SFENCE Enable (SLOWFENCE)Bit 1. Enable slow sfence. Cachable Memory Disable (TLBCACHEDIS)Bit 3. Disable performance improvement that assumes that the PML4, PDP, PDE and PTE entries are in cacheable memory. If page tables are uncachable, TLBCACHEDIS must be set to ensure correct functionality. 400 Processor Configuration Registers Chapter 14
SMMLOCK
DISLOCK
reserved
reserved
reserved
reserved
SSEDIS
FFDIS
32559
Rev. 3.08
July 2007
INVD to WBINVD Conversion (INVD_WBINVD)Bit 4. Convert INVD to WBINVD. TLB Flush Filter Disable (FFDIS)Bit 6. Disable TLB flush filter. Disable LOCK (DISLOCK)Bit 7. Disable x86 LOCK prefix functionality. IGNNE Port Emulation Enable (IGNNE_EM)Bit 8. Enable emulation of IGNNE port. Special Bus Cycle On HLT Exit Enable (HLTXSPCYCEN)Bit 12. Enables special bus cycle generation on exit from HLT. Special Bus Cycle On SMI Disable (SMISPCYCDIS)Bit 13. Disables special bus cycle on SMI. Special Bus Cycle On RSM Disable (RSMSPCYCDIS)Bit 14. Disables special bus cycle on RSM. SSE Instructions Disable (SSEDIS)Bit 15. Disables SSE instructions. 32-bit Address Wrap Disable (WRAP32DIS)Bit 17. Disable 32-bit address wrapping. MCi_STATUS Write Enable (MCi_STATUS_WREN)Bit 18. When set, writes by software to MCi_STATUS MSRs do not cause general protection faults. Such writes update all implemented bits in these registers. When clear, writing a non-zero pattern to these registers causes a general protection fault. See the description above for more details. Startup FID Status (START_FID)Bits 2924. Status of the startup FID. 14.2.1.4 NB_CFG Register
Software must perform a read-modify-write to this register to change its value. NB_CFG Register
63 55 54 53 46 45 44 43 42 DisThmlPfMonSmiinterrupts DisUsSysMgtRqToNLdt
MSR C001_001Fh
37 36 35 32
InitApicIdCpuIdLo
reserved
reserved
reserved
DisDatMsk
reserved
reserved
31 30 DisCohLdtCfg
10 9 DisRefUseFreeBuf
reserved
reserved
Chapter 14
401
32559
Rev. 3.08 July 2007
Bit 6355 54 5346 45 44 43 4237 36 3532 31 3010 9 80
Mnemonic reserved InitApicIdCpuIdLo reserved DisUsSysMgtRqToNLdt reserved
Function Initial APIC ID CPU ID Low Disable Upstream System Management Rebroadcast
R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0 0
DisThmlPfMonSmiinterrupts Disable Performance Monitor SMI reserved DisDatMsk reserved DisCohLdtCfg reserved DisRefUseFreeBuf reserved Disable Display Refresh from Using Free List Buffers Disable Coherent HyperTransport Configuration Accesses Disable Data Mask
R/W R/W
Field Descriptions Disable Display Refresh from Using Free List Buffers (DisRefUseFreeBuf)Bit 9. Disables display refresh requests from using free list buffers. Disable Coherent HyperTransport Configuration Accesses (DisCohLdtCfg)Bit 31. Disables automatic routing of PCI configuration accesses to the processor configuration registers. When set, PCI configuration space accesses which fall within the hard-coded range reserved for AMD NPT Family 0Fh Processor registers (see Memory System Configuration Registers on page 45) are instead routed via the configuration address maps. This can be used to effectively hide the configuration registers from software if they are routed to the I/O hub, where they will then get a master abort. It can also be used to provide a means for an external chip to route processor configuration accesses according to some scheme other than the hard-coded version described in Memory System Configuration Registers on page 45. When used, this bit needs to be set on all processors in a system. PCI configuration accesses should not be generated if this bit is not set on all processors. Disable Data Mask (DisDatMsk)Bit 36. Disables DRAM data masking function. For all sized write requests a DRAM read is performed before writing the data. Disable Performance Monitor SMI (DisThmlPfMonSmiIntr)Bit 43. Disables SMI generation for Performance Monitor interrupts. Disable Upstream System Management Rebroadcast (DisUsSysMgtRqToNLdt)Bit 45. Disables re-broadcast of Upstream StpClk and Legacy Input System Management commands downstream. Initial APIC ID CPU ID Low (InitApicIdCpuIdLo)Bit 54. When this bit is set, CpuId and NodeId[2:0] bit field positions are swapped in the APICID. 402 Processor Configuration Registers Chapter 14
32559
Rev. 3.08
July 2007
0: APICID = {CpuId, NodeId[2:0]} 1: APICID = {NodeId[2:0], CpuId}
14.2.2
14.2.2.1
Identification Registers
MANID Register
This status register holds the mask manufacturing identification number. MANID Register
63 reserved 31 reserved 10 9 ReticleSite 8 7 MajorRev 4 3 MinorRev 0
MSR C001_001Eh
32
Bit 6310 98 74 30
Mnemonic reserved ReticleSite MajorRev MinorRev
Function Reticle site Major mask set revision number Minor mask set revision number
R/W R R R
Field Descriptions Minor Mask Set Revision Number (MinorRev)Bits 30. Major Mask Set Revision Number (MajorRev)Bits 74. Reticle Site (ReticleSite)Bits 98.
14.2.3
HTC Register
See HTC Register for a complete description of this register. HTC Register
63 reserved
MSR C001_003Eh
32
Chapter 14
403
31 reserved
32559
Rev. 3.08 July 2007
5 HtcActSts
4 HtcAct
3 reserved
1 HtcSbcEn
0 HtcEn 32 0 StcSbcTmpHiEn
Bits 636 5 4 32 1 0
Mnemonic reserved HtcActSts HtcAct reserved HtcSbcEn HtcEn
Function HTC Active Status HTC Active HTC Special Bus Cycle Enable HTC Enable
R/W R R/W R R R/W R/W
Reset 0 0 0 0 0 0
14.2.4
Thermal Control Register

MSR C001_003Fh
See Thermal Control Register for a complete description of this register. Thermal Control Register
63 reserved 31 28 27 24 23 21 20 16 15 8 7 StcTmpLoSts
6 StcTmpHiSts
3 reserved
4 StcHtcEn
3 StcApcTmpLoEn
2 StcApcTmpHiEn
1 StcSbcTmpLoEn
reserved
StcHystLmt
reserved
StcTmpLmt
reserved
Bits 31-28 27-24 2321 20-16 158 7 6 5 4 3 2 1 0
Mnemonic reserved StcHystLmt reserved StcTmpLmt reserved StcTmpLoSts StcTmpHiSts reserved StcHtcEn StcApcTmpLoEn StcApcTmpHiEn StcSbcTmpLoEn StcSbcTmpHiEn
Function STC Hysteresis Limit STC Temperature Limit STC Temperature Low Status STC Temperature High Status STC HTC Enable STC APIC Temperature Low Interrupt Enable STC APIC Temperature High Interrupt Enable STC Special Bus Cycle Temperature Low Enable STC Special Bus Cycle Temperature High Enable
R/W R R/W R R/W R R/W R/W R R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0 0 0 0 0
404
Chapter 14
32559
Rev. 3.08
July 2007
14.2.5
14.2.5.1
Memory Typing Registers

TOP_MEM Register
This register holds the address of the top of memory. When enabled, this address indicates the first byte of I/O above DRAM. TOP_MEM Register
63 reserved 31 TOM 80 23 22 reserved 40 39 TOM 169 0
MSR C001_001Ah
32
Bit 6340 3923 220
Mnemonic reserved TOM reserved
Function RAZ Top of Memory RAZ
R/W R/W
Reset
Field Descriptions Top of Memory (TOM)Bits 3923. Address of top of memory (8-Mbyte granularity). 14.2.5.2 TOP_MEM2 Register
This register holds the second top of memory address. When enabled by setting the MtrrTom2En bit in the SYSCFG register, this address indicates the first byte of I/O above a second allocation of DRAM that starts at 4 Gbytes. Hence, the address in TOP_MEM2 should be set above 4 Gbytes. TOP_MEM2 Register
63 reserved 31 TOM2 80 23 22 reserved 40 39 TOM2 169 0
MSR C001_001Dh
32
Bit 6340
Mnemonic reserved
Function RAZ
R/W
Reset
Chapter 14
405
32559
Rev. 3.08 July 2007
Bit 3923 220
Mnemonic TOM2 reserved
Function Second Top of Memory RAZ
R/W R/W
Reset
Field Descriptions Second Top of Memory (TOM2)Bits 3923. Address of second top of memory (8-Mbyte granularity).
14.2.6
14.2.6.1
I/O Range Registers

IORRBasei Registers
These registers hold the bases of the variable I/O ranges. IORRBase01 Registers
63 reserved 31 Base 190 12 11 reserved
MSRs C001_0016h, C001_0018h

40 39 Base 2720 5 4 RdDram 3 WrDram 2 0 32
reserved
Bit 6340 3912 115 4 3 20
Mnemonic reserved Base reserved RdDram WrDram reserved
Function RAZ Base address RAZ Read from DRAM Write to DRAM RAZ
R/W R/W R/W R/W
Reset
Field Descriptions Write to DRAM (WrDram)Bit 3. If set, stores write to DRAM, otherwise I/O for this range. Read from DRAM (RdDram)Bit 4. If set, fetches read from DRAM, otherwise from I/O for this range. Base Address (Base)Bits 3912. Base address for this range. 14.2.6.2 IORRMaski Registers
These registers hold the masks of the variable I/O ranges. 406 Processor Configuration Registers Chapter 14
32559
Rev. 3.08
July 2007
IORRMask01 Registers
63 reserved 31 Mask 190
MSRs C001_0017h, C001_0019h

40 39 Mask 2720 12 11 10 V reserved 0 32
Bit 6340 3912 11 100
Mnemonic reserved Mask V reserved
Function RAZ Address mask Enables variable I/O range registers RAZ
R/W R/W R/W
Reset
Field Descriptions Variable I/O Range (V)Bit 11. Enables variable I/O range register. Address Mask (Mask)Bits 3912. Address mask.
14.2.7
System Call Extension Registers
The system call extension is enabled by setting bit 0 in the EFER register. This feature adds the SYSCALL and SYSRET instructions which can be used in flat addressed operating systems as low latency system calls and returns. See the SYSCALL and SYSRET section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. 14.2.7.1 STAR Register
This register holds the target address used by the SYSCALL instruction and the code and stack segment selector bases used by the SYSCALL and SYSRET instructions. STAR Register
63 SysRetSel 31 Target 48 47 SysCallSel 0
MSR C000_0081h
32
Chapter 14
407
32559
Rev. 3.08 July 2007
Bit 6348 4732 310
Mnemonic SysRetSel SysCallSel Target
Function SYSRET CS and SS SYSCALL CS and SS SYSCALL target address
R/W R/W R/W R/W
Reset
Field Descriptions SYSCALL Target Address (Target)Bits 310. SYSCALL CS and SS (SysCallSel)Bits 4732. SYSRET CS and SS (SysRetSel)Bits 6348. 14.2.7.2 LSTAR Register
The address stored in this register must be in canonical form (if not canonical, a #GP fault occurs). LSTAR Register
63 LSTAR 6332 31 LSTAR 310 0
MSR C000_0082h
32
Bit 630
Mnemonic LSTAR
Function Long Mode Target Address
R/W R/W
Reset
Field Descriptions Long Mode Target Address (LSTAR)Bits 630. Target address for 64-bit mode calling programs. 14.2.7.3 CSTAR Register
The address stored in this register must be in canonical form (if not canonical, a #GP fault will occur). CSTAR Register
63 CSTAR 6332
MSR C000_0083h
32
408
Chapter 14
32559
Rev. 3.08
July 2007
0 CSTAR 310
31
Bit 630
Mnemonic CSTAR
Function Compatibility mode target address
R/W R/W
Reset
Field Descriptions Compatibility Mode Target Address (CSTAR)Bits 630. Target address for compatibility mode calling programs. 14.2.7.4 SF_MASK Register
This register holds the EFLAGS mask used by the SYSCALL instruction. Each bit in this mask clears the corresponding EFLAGS bit when executing the SYSCALL instruction. SF_MASK Register
63 reserved 31 MASK 0
MSR C000_0084h
32
Bit 6332 310
Mnemonic reserved MASK
Function RAZ SYSCALL Flag Mask
R/W R/W
Reset
Field Descriptions SYSCALL Flag Mask (MASK)Bits 310.
14.2.8
14.2.8.1
Segmentation Registers
FS.Base Register
This register provides access to the expanded 64 bit FS segment base. The address stored in this register must be in canonical form (if not canonical, a #GP fault fill occur).
Chapter 14
409
32559
Rev. 3.08 July 2007
FS.Base Register
63 FS_BASE (6332) 31 FS_BASE (310)
MSR C000_0100h
32
Bit 630
Mnemonic FS_BASE
Function Expanded FS segment base
R/W R/W
Reset
Field Descriptions Expanded FS Segment Base (FS_BASE)Bits 630. 14.2.8.2 GS.Base Register
This register provides access to the expanded 64 bit GS segment base. The address stored in this register must be in canonical form (if not canonical, a #GP fault fill occur). GS.Base Register
63 GS_BASE 6332 31 GS_BASE 310 0
MSR C000_0101h
32
Bit 630
Mnemonic GS_BASE
Function Expanded GS segment base
R/W R/W
Reset
Field Descriptions Expanded GS Segment Base (GS_BASE)Bits 630. 14.2.8.3 KernelGSbase Register
This register holds the kernel data structure pointer which can be swapped with the GS_BASE register using the new 64-bit mode instruction SwapGS. See the SWAPGS Instruction section in Volume 2 of the AMD64 Architecture Programmers Manual for more information. The address stored in this register must be in canonical form (if not canonical, a #GP fault occurs).
410
Chapter 14
32559
Rev. 3.08
July 2007
KernelGSbase Register
63 KernelGSBase 6332 31 KernelGSBase 310
MSR C000_0102h
32
Bit 630
Mnemonic KernelGSBase
Function Kernel data structure pointer
R/W R/W
Reset
Field Descriptions Kernel Data Structure Pointer (KernelGSBase)Bits 630.
14.2.9
Power Management Registers
Mobile processors contain AMD PowerNow! technology hardware that allows the processor operating voltage and frequency to be dynamically controlled for power management purposes. Use CPUID function 8000_0007h (Get Advanced Power Management Feature Flags) to determine the thermal and power management capabilities of the processor. Refer to Processor Performance States on page 287 regarding the use of the FIDVID_STATUS and FIDVID_CTL MSRs. 14.2.9.1 FIDVID_CTL Register
This register holds configuration information relating to power management control. Accessing this register is allowed if the device supports either FID control or VID control as indicated by CPUID. If neither FID control nor VID control is indicated, accessing the FIDVID_CTL register causes a GP# fault. This register is used to specify new Frequency ID (FID) and/or Voltage ID (VID) codes to which to change on the next FID/VID change. This register is also used to control how long to wait following a FID/VID change before the PLL and voltage regulator are stable at the new FID/VID. This register is persistent through a warm reset. The processor core clock PLL lock time is 2 s for AMD NPT Family 0Fh Processors. This time refers to all components of PLL lock including frequency lock, phase lock, and settling time. FIDVID_CTL Register
63 reserved 52 51 StpGntTOCnt
MSR C001_0041h
32
Chapter 14
411
32559
Rev. 3.08 July 2007
31 reserved
17 16 15 InitFidVid reserved
13 NewVID
7 reserved
5 NewFID
Bit 6352 5132 3117 16 1514 138 76 50
Mnemonic reserved StpGntTOCnt reserved InitFidVid reserved NewVID reserved NewFID
Function MBZ Stop Grant Time-Out Count MBZ Initiate FID/VID Change MBZ New VID MBZ New FID
R/W R/W W R/W R R/W
Reset 0 0 0 0 0 0 0 0
Field Descriptions Stop Grant Time-Out Count (StpGntTOCnt)Bits 51-32. This field carries a count of system clock cycles (5 ns) that must elapse from the time a new FID is applied until the time that the PLL is stable at the new FID. The BIOS or processor driver must program this field with the PLL lock time, defined above, prior to a FID change. Initiate FID/VID Change (InitFidVid)Bits 16. Writing this bit to a 1 initiates a FID/VID change. In a multiprocessor system, only the bootstrap processor sees this bit written as a 1. Writing this bit to a 1 initiates the FID change special bus cycle. This is a write-only bit and always reads as 0. The status of the resulting FID/VID change can be determined from the FIDVID_STATUS register. New VID (NewVID)Bits 138. This field is the new VID to transition to. If an attempt is made to write a NewVID value that corresponds to a voltage greater than the voltage that MaxVID corresponds to in the FIDVID_STATUS register then the MaxVID value is written instead.See Table 71 on page 309 for VID values and their corresponding voltages. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. New FID (NewFID)Bits 50. This field is the new FID to transition to. If an attempt is made to write a NewFID value greater than MaxFID in the FIDVID_STATUS register then the MaxFID value is written instead. See Table 80 for FID code translations.
Table 80. FID Code Translations
Reference Clock Multiplier 4x 6-Bit FID Code 00_0000b Reference Clock Multiplier 15x 6-Bit FID Code 01_0110b
412
Chapter 14
32559
Rev. 3.08
July 2007
Table 80. FID Code Translations (Continued)

Reference Clock Multiplier 4.5x 5x 5.5x 6x 6.5x 7x 7.5x 8x 8.5x 9x 9.5x 10x 10.5x 11x 11.5x 12x 12.5x 13x 13.5x 14x 14.5x 6-Bit FID Code 00_0001b 00_0010b 00_0011b 00_0100b 00_0101b 00_0110b 00_0111b 00_1000b 00_1001b 00_1010b 00_1011b 00_1100b 00_1101b 00_1110b 00_1111b 01_0000b 01_0001b 01_0010b 01_0011b 01_0100b 01_0101b Reference Clock Multiplier 15.5x 16x 16.5x 17x 17.5x 18x 18.5x 19X 19.5X 20X 20.5X 21X 21.5X 22x 22.5x 23x 23.5x 24x 24.5x 25x 6-Bit FID Code 01_0111b 01_1000b 01_1001b 01_1010b 01_1011b 01_1100b 01_1101b 01_1110b 01_1111b 10_0000b 10_0001b 10_0010b 10_0011b 10_0100b 10_0101b 10_0110b 10_0111b 10_1000b 10_1001b 10_1010b
Note: Encodings not listed are reserved. Using reserved encodings causes unpredictable behavior.
14.2.9.2
FIDVID_STATUS Register
This register holds status of the current FID, the current VID, pending changes, and maximum values. Accessing this register is allowed if the device supports either FID control or VID control as indicated via CPUID. If neither FID control nor VID control is indicated, accessing the FIDVID_STATUS register causes a GP# fault. See Table 80 on page 412 for a complete list of FID values and their corresponding reference clock multiplier values. See Table 71 on page 309 for VID values and their corresponding voltages.
Chapter 14
413
32559
Rev. 3.08 July 2007
FIDVID_STATUS Register
63 62 61 60 59 IntPstateSup AltVidOffset reserved reserved 57 56 55 PstateStep reserved 53 48 47 reserved 45 40 39 reserved
MSR C001_0042h
37 32
MaxVID
StartVID
CurrVID
31 30 29 FidVidPending reserved
24 23 22 21 reserved
16 15 14 13 reserved
7 reserved
MaxRampVID
MaxFID
StartFID
CurrFID
Bit 6362 61 60 5957 56 5554 5348 4746 4540 3938 3732 31 30 2924 2322 2116 1514 138 76 50
Mnemonic reserved IntPstateSup reserved AltVidOffset PstateStep reserved MaxVID reserved StartVID reserved CurrVID FidVidPending reserved MaxRampVID reserved MaxFID reserved StartFID reserved CurrFID
Function RAZ Intermediate P-state support AltVID Offset P-state Voltage Step RAZ Max VID RAZ Startup VID RAZ Current VID FID/VID Change Pending RAZ Max Ramp VID RAZ Max FID RAZ Startup FID RAZ Current FID
R/W R R R R R R R R R R R R R R R R R R R R
Reset 0 X X X X 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Field Descriptions Intermediate P-State Support (IntPstateSup)Bit 61. This bit indicates if intermediate P-states are supported. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. 0b = Intermediate P-states not supported. 1b = Intermediate P-states supported.
414
Chapter 14
32559
Rev. 3.08
July 2007
AltVID Offset (AltVidOffset)Bits 5957. This field specifies the amount of voltage subtracted from the start VID to determine the AltVID voltage. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. 000b = AltVID not supported. 001b = -50mV. 010b = -100mV. 011b = -125mV. 100b = -150mV. 101b = -175mV. 110b = -200mV. 111b = -225mV. P-State Voltage Step Size (PstateStep)Bit 56. This bit indicates the voltage reduction associate with each 200MHz reduction in operating frequency. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. 0b = 25mV reduction. 1b = 50mV reduction. Max VID (MaxVID)Bits 5348. This field indicates the VID code associated with the maximum voltage software can select. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. Startup VID (StartVID)Bits 4540. This field indicates the startup VID. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. Current VID (CurrVID)Bits 3732. This field indicates the current VID. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. FID/VID Change Pending (FidVidPending)Bit 31. This bit indicates that a FID/VID change is pending. It is set to 1 by hardware when the InitFidVid bit is set in the FIDVID_CTL register. It is cleared by hardware when the FID/VID change has completed. Max Ramp VID (MaxRampVID)Bit 2924. This field indicates the max Ramp VID. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this field. Max FID (MaxFID)Bit 2116. This field indicates the max FID. Startup FID (StartFID)Bit 138. This field indicates the startup FID. Current FID (CurrFID)Bit 50. This field indicates the current FID.
Chapter 14
415
32559
Rev. 3.08 July 2007
14.2.10 I/O and Configuration Space Trapping to SMI

I/O and configuration space trapping to SMI is a mechanism for executing SMI handler if a specific I/O access to one of the specified addresses is detected. Access address and access type checking is done before I/O instruction execution. If the access address and access type match one of the specified I/O address and access types, the SMI handler is executed, the I/O instruction is not executed, and a breakpoint set on the I/O instruction is not taken. The I/O instruction can be executed inside the SMI handler. This mechanism has higher priority than the SMI interrupt. If the trap is not taken, check for SMI interrupt is done after the I/O instruction has been executed. Configuration accesses are special I/O accesses. An I/O access is defined as a configuration access when I/O instruction address bits 31-0 are CFCh, CFDh, CFEh, or CFFh. The access address for a configuration space access is the current value of doubleword I/O register CF8h. The access address for an I/O access that is not a configuration access is equivalent to the I/O instruction address, bits 31 0. The access address is compared with SmiAddr, and the instruction access type is compared with the enabled access types defined by ConfigSMI, SmiOnRdEn, and SmiOnWrEn. Access address bits 230 can be masked with SmiMask. Fields SmiAddr, SmiMask, ConfigSMI, SmiOnRdEn, and SmiOnWrEn are defined in IOTRAP_ADDRi registers (i = 0,1,2,3). An SMI handler executed as a result of I/O and configuration space trapping is by default executed only by the processor core executing the I/O instruction. The user has the ability to enable an SMI special bus cycle and RSM special bus cycle generation if SMI special bus cycles are enabled globally in the HWCR register. I/O and configuration space trapping to SMI applies only to single I/O instructions; it does not apply to string and REP I/O instructions. 14.2.10.1 IOTRAP_ADDRi Registers IOTRAP_ADDRi Registers
63 62 61 60 SmiOnRdEn SmiOnWrEn ConfigSmi 56 55
MSRs C001_0050h, C001_0051h, C001_0052h, C001_0053h

32
reserved
SmiMask
31 SmiAddr
416
Chapter 14
32559
Rev. 3.08
July 2007
Bit 63 62 61 60-56 5532 310
Mnemonic SmiOnRdEn SmiOnWrEn ConfigSmi reserved SmiMask SmiAddr
Function Enable SMI on IO Read Enable SMI on IO Write Configuration Space SMI RAZ SMI Mask SMI Address
R/W R/W R/W R/W R R/W R/W
Reset 0 0 0 0 0
Field Descriptions Enable SMI on IO Read (SmiOnRdEn)Bits 63. Enables SMI generation on a read access. Enable SMI on IO Write (SmiOnWrEn)Bit 62. Enables SMI generation on a write access. Configuration Space SMI (ConfigSmi)Bit 61. 1 = configuration access (see I/O and Configuration Space Trapping to SMI on page 416 for more information on configuration access detection); 0 = I/O access different than configuration access. SMI Mask (SmiMask)Bits 5532. SMI I/O trap mask. 0 = mask address bit; 1 = do not mask address bit. When the ConfigSmi bit is set bits 33 and 32 are ignored. SMI Address (SmiAddr)Bits 310. SMI I/O trap address. When the ConfigSmi bit set is the SMI address is internally doubleword aligned and bits 0 and 1 of SmiAddr are ignored. 14.2.10.2 IOTRAP_CTL Register IOTRAP_CTL Register
63 reserved 31 16 15 14 13 12 IoTrapCtlRsmSpcEn IoTrapCtlSmiSpcEn 8 7 6 5 4 3 2 1 0
MSR C001_0054h
32
SmiSts_3
SmiSts_2
SmiSts_1
reserved
reserved
Bit 6332 3116 15 14
Mnemonic reserved reserved IoTrapEn IoTrapCtlSmiSpcEn
Function RAZ MBZ IO Trap Enable IO Trap Control SMI Special Cycle Enable
R/W R R R/W R/W
Reset
0 0
Chapter 14
417
SmiSts_0
IoTrapEn
SmiEn_3
SmiEn_2
SmiEn_1
SmiEn_0
32559
Rev. 3.08 July 2007
Bit 13 12-8 7 6 5 4 3 2 1 0
Mnemonic IoTrapCtlRsmSpcEn reserved SmiEn_3 SmiSts_3 SmiEn_2 SmiSts_2 SmiEn_1 SmiSts_1 SmiEn_0 SmiSts_0
Function IO Trap Control RSM Special Cycle Enable MBZ SMI Enable 3 SMI Status 3 SMI Enable 2 SMI Status 2 SMI Enable 1 SMI Status 1 SMI Enable 0 SMI Status 0
R/W R/W R R/W R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0 0 0
Field Descriptions IO Trap Enable (IoTrapEn)Bit 15. Enable I/O and configuration space trapping. IO Trap Control SMI Special Cycle Enable (IoTrapCtlSmiSpcEn)Bit 14. Enable SMI special bus cycle generation when SMI handler is entered as a result of I/O and configuration space trapping. If SMISPCYCDIS (MSR C001_0015h, bit 13) is set then this bit has no effect. IO Trap Control RSM Special Cycle Enable (IoTrapCtlRsmSpcEn)Bit 13. Enable RSM special bus cycle generation when SMI handler is entered as a result of I/O and configuration space trapping. If RSMSPCYCDIS (MSR C001_0015h, bit 14) is set, then this bit has no effect. SMI Enable 3 (SmiEn_3)Bit 7. Enable SMI generation if I/O access matches access specified in MSR C001_0053h. SMI Status 3 (SmiSts_3)Bit 6. Set if I/O access matched access specified in MSR C001_0053h. SMI Enable 2 (SmiEn_2)Bit 5. Enable SMI generation if I/O access matches access specified in MSR C001_0052h. SMI Status 2 (SmiSts_2)Bit 4. Set if I/O access matched access specified in MSR C001_0052h. SMI Enable 1 (SmiEn_1)Bit 3. Enable SMI generation if I/O access matches access specified in MSR C001_0051h. SMI Status 1 (SmiSts_1)Bit 2. Set if I/O access matched access specified in MSR C001_0051h. SMI Enable 0 (SmiEn_0)Bit 1. Enable SMI generation if I/O access matches access specified in MSR C001_0050h. SMI Status 0 (SmiSts_0)Bit 0. Set if I/O access matched access specified in MSR C001_0050h.
418
Chapter 14
32559
Rev. 3.08
July 2007
14.2.11 Interrupt Pending Message Register

This register is used to ensure that the system I/O Hub can wake the processor out of the stop grant state when there is a pending interrupt. If the interrupt pending message is not specified by this register, it is possible for the system I/O Hub to place the processor into the stop grant state while an interrupt is pending in the processor. This register must be setup for platforms that support the C2 or C3 power management states. This register enables the processor to send a message to the I/O hub that results in the pending interrupt being serviced.The two message types are the I/O space message and the HyperTransport INT_PENDING message. The HyperTransport Technology INT_PENDING message is defined by the HyperTransport 1.05c specification. If the processor or the I/O hub does not support the INT_PENDING HyperTransport Technology message, the I/O space message should be selected by IntPndMsg. The I/O space message should be chosen to avoid interference with the OSPM SMI command port value. If BIOS configures a write to the SMI command port for the interrupt pending message, then it is recommended that BIOS also disable all external asyncrhonous sources of STPCLK (STPCLK throttling, e.g.). A check for a pending interrupt is performed at the end of an I/O instruction. If there is a pending interrupt and STPCLK is asserted, the processor executes a byte-size read or write to I/O space defined with IORd, IOMsgAddr, and IOMsgData (used only for I/O writes) to generate an SMI. The SMI wakes up the processor so the original pending interrupt can be serviced. The SMI handler should not take any action if the SMI is generated by this mechanism. In order to prevent SMI generation with this mechanism in the SMI handler, IntrPndMsgDis bit should be set in the SMI handler before the first I/O instruction is executed, and it should be cleared prior to resuming from SMM. If the processor and the I/O hub support the INT_PENDING HyperTransport Technology message, it should be selected by IntPndMsg. The check for a pending interrupt is performed when entering the stop grant state, if STPCLK is not initiated by Hardware Thermal Control (HTC) or Software Thermal Control (STC). If there is a pending interrupt, the processor sends an INT_PENDING HyperTransport message to the I/O hub. The other purpose for this register is to generate messages to the I/O Hub that result in SMI interrupts when all the CPU cores in the processor are in the halt state. This MSR is not shared between CPU cores and must be programmed the same on both cores. If SmiOnCmpHalt is set, then whenever a CPU core enters the halt state while all other CPU cores of the processor are also in halt, then the processor generates an I/O cycle based on the programmed values in IORd, IOMsgData and IOMsgAddr. The expectation is that this I/O cycle targets an I/O Hub register that triggers an SMI back to the processor; the SMI handler may then be used to bring the processor into the C1 enhanced mode (C1E), a low-power halt mode. Interrupt Pending Message Register
63 reserved
MSR C001_0055h
32
Chapter 14
419
31 29 28 27 26 25 24 23 C1eOnCmpHalt SmiOnCmpHalt IntrPndMsgDis IntrPndMsg 16 15
32559
Rev. 3.08 July 2007
reserved
IORd
IOMsgData
IOMsgAddr
Bit 6329 28 27 26 25 24 2316 150
Mnemonic reserved C1eOnCmpHalt SmiOnCmpHalt IORd IntrPndMsg IntrPndMsgDis IOMsgData IOMsgAddr
Function RAZ C1E on Multi-core halt SMI on Multi-core halt IO Read Interrupt Pending Message Interrupt Pending Message Disable IO Message Data IO Message Address
R/W R R/W R/W R/W R/W R/W R/W R/W
Reset 0 0 0 0 0 0 0
Field Descriptions C1E on Multi-core Halt (C1eOnCmpHalt)Bit 28. 1=When all CPU cores of the processor have entered the halt state, the processor generates an I/O cycle as specified by IORd, IOMsgData, and IOMsgAddr. When this bit is set, the I/O cycle specified should be a read to the ACPI defined P_LVL3 register. When this bit is set SmiOnCmpHalt and IntPndMsg must be 0. This bit is not supported on single core processors. SMI on Multi-core Halt (SmiOnCmpHalt)Bit 27. 1=When all CPU cores of the processor have entered the halt state, the processor generates an I/O cycle as specified by IORd, IOMsgData, and IOMsgAddr. When this bit is set, the I/O cycle specified should be a write to the system SMI command port. When this bit is set C1eOnCmpHalt and IntPndMsg must be 0. This bit is not supported on single core processors. IO Read (IORd)Bit 26. 1 = I/O read; 0 = I/O write. Interrupt Pending Message (IntrPndMsg)Bit 25. Select an interrupt pending message type. 0 = HyperTransport INT_PENDING message; 1 = I/O space message. This bit should be set after IORd, IOMsgAddr and IOMsgData fields are programmed. Interrupt Pending Message Disable (IntrPndMsgDis)Bit 24. Disable generating an interrupt pending message defined with IntrPndMsg. IO Message Data (IOMsgData)Bits 2316. I/O space message data. This field is only used when an I/O write is selected (IORd=0). IO Message Address (IOMsgAddr)Bits 150. I/O space message address.
420
Chapter 14
32559
Rev. 3.08
July 2007
14.2.12 VM_CR Register

This register provides security related controls. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. VM_CR Register
63 reserved 31 reserved 5 4 SVME_DIS 3 LOCK 2 dis_a20m 1 r_init 0 dpd
MSR C001_0114h
32
Bit 635 4 3 2 1 0
Mnemonic reserved SVME_DISABLE LOCK dis_a20m r_init dpd
Function Disable SVME Lock SVM Disable A20 Masking Intercept INIT Debug Port Disable
R/W R R/W R R/W R/W R/W
Reset 0
0 0 0
Field Descriptions Debug Port Disable (dpd)Bit 0. This bit controls if debug facilities such as JTAG and HDT have access to the processor state information. When this bit is set, all mechanism that could expose trusted code execution are disabled. 0 = Debug port enabled. 1 = Debug port disabled. Intercept INIT (r_init)Bit 1. This bit controls how INIT is delivered in host mode. 0 = INIT delivered normally. 1 = INIT translated into a SX interrupt. Disable A20 Masking (dis_a20m)Bit 2. When set, this bit disables A20 masking. Lock SVM (LOCK)Bit 3. When this bit is set writes to SVME_DISABLE (bit 4) are ignored. Disable SVME (SVME_DISABLE)Bit 4. This bit determines if the EFER.SVME bit is writeable by hypervisors. If LOCK (bit 3) is set then this bit is read-only. 0 = SVM is allowed to be enabled using EFER.SVME. 1 = SVM is not allowed. The user must change a BIOS setting to enable SVM.
Chapter 14
421
32559
Rev. 3.08 July 2007
14.2.13 VM_HSAVE_PA Register

This register contains the physical address where VMRUN saves host state and where vm-exit restores host state from. Writing this register causes a GP exception if any of the lower 12 bits are not zero or if the address written is greater than FD_000_000h. See Register Differences in Revisions of AMD NPT Family 0Fh Processors on page 21 for revision information about this register. VM_HSAVE_PA Register
63 VM_HSAVE_PA 31 VM_HSAVE_PA 5 4 3 2 1 0
MSR C001_0117h
32
Bit 630
Mnemonic VM_HSAVE_PA
Function Physical Address of Host Save Area
R/W R/w
Reset 0
422
Chapter 14
32559
Rev. 3.08
July 2007
Glossary
128-bit media instructions. Instructions that use the 128-bit XMM registers. These are a combination of SSE and SSE2 instruction sets. 64-bit media instructions. Instructions that use the 64-bit MMX registers. These are primarily a combination of MMX and 3DNow! instruction sets, with some additional instructions from the SSE and SSE2 instruction sets. 16-bit mode. Legacy mode or compatibility mode in which a 16-bit address size is active. See legacy mode and compatibility mode. 32-bit mode. Legacy mode or compatibility mode in which a 32-bit address size is active. See legacy mode and compatibility mode. 64-bit mode. A submode of long mode. In 64-bit mode, the default address size is 64 bits and new features, such as register extensions, are supported for system and application software. absolute. Said of a displacement that references the base of a code segment rather than an instruction pointer. Contrast with relative. AHDH. The high 8-bit AH, BH, CH, and DH registers. Compare ALDL. ALDL. The low 8-bit AL, BL, CL, and DL registers. Compare AHDH. ALr15B. The low 8-bit AL, BL, CL, DL, SIL, DIL, BPL, SPL, and R8BR15B registers, available in 64-bit mode. biased exponent. The sum of a floating-point values exponent and a constant bias for a particular floating-point data type. The bias makes the range of the biased exponent always positive, which allows reciprocation without overflow. byte. Eight bits. clear. To write a bit-value of 0. Compare set. compatibility mode. A submode of long mode. In compatibility mode, the default address size is 32 bits and legacy 16-bit and 32-bit applications run without modification. b. A bit, as in 1 Mb for one megabit, or lsb for least-significant bit. B. A byte, as in 1 MB for one megabyte, or LSB for least-significant byte. canonical address. AMD NPT Family 0Fh Processors implement 48-bit virtual addresses. A virtual address having all 1s or all 0s in bit 63 through the most significant implemented bit (bit 47 for AMD NPT Family 0Fh Processors) is said to be in canonical form. Accessing a virtual address that is not in canonical form results in a fault. commit. To irreversibly write, in program order, an instructions result to software-visible storage, such as a register (including flags), the data cache, an internal write buffer, or memory. CPL. Current privilege level. Chapter Glossary 423
32559
Rev. 3.08 July 2007
CRn. Control register number n. CS. Code segment register. direct. Referencing a memory location whose address is included in the instructions syntax as an immediate operand. The address may be an absolute or relative address. Compare indirect. dirty data. Data held in the processors caches or internal buffers that is more recent than the copy held in main memory. displacement. A signed value that is added to the base of a segment (absolute addressing) or an instruction pointer (relative addressing). Same as offset. doubleword. Two words, or four bytes, or 32 bits. double quadword. Eight words, or 16 bytes, or 128 bits. Also called octword. DP. Two processors or dual processors. eAXeSP. The 16-bit AX, BX, CX, DX, DI, SI, BP, SP registers or the 32-bit EAX, EBX, ECX, EDX, EDI, ESI, EBP, ESP registers. Compare rAXrSP. EFER. Extended features enable register. effective address size. The address size for the current instruction after accounting for the default address size and any address-size override prefix. effective operand size. The operand size for the current instruction after accounting for the default operand size and any operand-size override prefix. eFLAGS. 16-bit or 32-bit flags register. Compare rFLAGS. EFLAGS. 32-bit (extended) flags register. eIP. 16-bit or 32-bit instruction-pointer register. Compare rIP. EIP. 32-bit (extended) instruction-pointer register. element. See vector. exception. An abnormal condition that occurs as the result of executing an instruction. The processors response to an exception depends on the type of the exception. For all exceptions except 128bit media SIMD floating-point exceptions and x87 floating-point exceptions, control is transferred to the handler (or service routine) for that exception, as defined by the exceptions vector. For floating-point exceptions defined by the IEEE 754 standard, there are both masked and unmasked responses. When unmasked, the exception handler is called, and when masked a default response is provided instead of calling the handler. FLAGS. 16-bit flags register. flush. When referring to caches, (1) to writeback, if modified, and invalidate, as in flush the cache line; when referring to the processor pipeline, (2) to invalidate, as in flush the pipeline; or (3) to change a value, as in flush to zero. GDT. Global descriptor table. 424 Glossary Chapter
32559
Rev. 3.08
July 2007
GDTR. Global descriptor table register. GPRs. General-purpose registers. For the 16-bit data size, these are AX, BX, CX, DX, DI, SI, BP, SP. For the 32-bit data size, these are EAX, EBX, ECX, EDX, EDI, ESI, EBP, ESP. For the 64-bit data size, these include RAX, RBX, RCX, RDX, RDI, RSI, RBP, RSP, and R8R15. IDT. Interrupt descriptor table. IDTR. Interrupt descriptor table register. IGN. Ignore. Field is ignored. indirect. Referencing a memory location whose address is in a register or other memory location. The address may be an absolute or relative address. Compare direct. IP. 16-bit instruction-pointer register. IRB. The virtual-8086 mode interrupt-redirection bitmap. IST. The long-mode interrupt-stack table. IVT. The real-address mode interrupt-vector table. LDT. Local descriptor table. LDTR. Local descriptor table register. legacy x86. The legacy x86 architecture. See Related Documents on page 19 for descriptions of the legacy x86 architecture. legacy mode. An operating mode of the AMD64 architecture in which existing 16-bit and 32-bit applications and operating systems run without modification. A processor implementation of the AMD64 architecture can run in either long mode or legacy mode. Legacy mode has three submodes, real mode, protected mode, and virtual-8086 mode. long mode. An operating mode unique to the AMD64 architecture. A processor implementation of the AMD64 architecture can run in either long mode or legacy mode. Long mode has two submodes, 64-bit mode and compatibility mode. lsb. least-significant bit. LSB. least-significant byte. main memory. Physical memory, such as RAM and ROM (but not cache memory) that is installed in a particular computer system. mask. (1) A control bit that prevents the occurrence of a floating-point exception from invoking an exception handling routine. (2) A field of bits used for a control purpose. MBZ. Must be 0. If software attempts to set an MBZ bit to 1, a general-protection exception (#GP) occurs. memory. Unless otherwise specified, main memory. ModRM. A byte following an instruction opcode that specifies address calculation based on mode (mod), register (r), and memory (m) variables. Chapter Glossary 425
32559
Rev. 3.08 July 2007
moffset. A direct memory offset. I.e., a displacement that is added to the base of a code segment (for absolute addressing) or to an instruction pointer (for addressing relative to the instruction pointer, as in RIP-relative addressing). MP. More than two processors. msb. Most-significant bit. MSB. Most-significant byte. MSR. Model-specific register. multimedia instructions. A combination of 128-bit media instructions and 64-bit media instructions. octword. Same as double quadword. offset. Same as displacement. overflow. The condition in which a floating-point number is larger in magnitude than the largest, finite, positive or negative number that can be represented in the data-type format being used. packed. See vector. PAE. Physical-address extensions. physical memory. Actual memory, consisting of main memory and cache. probe. A check for an address in a processors caches or internal buffers. External probes originate outside the processor, and internal probes originate within the processor. processor. This term refers to AMD Athlon 64 processor architecture, AMD Opteron processor architecture, AMD Sempron processor architecture and AMD Turion 64 Mobile Technology processor architecture. This document covers all classes of these devices. term is used for to refer to packages that contain one or more processor cores. For details about differences between them, see the relevant processor data sheets and functional data sheets.
TM TM TM TM
protected mode. A submode of legacy mode. quadword. Four words, or eight bytes, or 64 bits. r8r15. The 8-bit R8BR15B registers, or the 16-bit R8WR15W registers, or the 32-bit R8DR15D registers, or the 64-bit R8R15 registers. rAXrSP. The 16-bit AX, BX, CX, DX, DI, SI, BP, SP registers, or the 32-bit EAX, EBX, ECX, EDX, EDI, ESI, EBP, ESP registers, or the 64-bit RAX, RBX, RCX, RDX, RDI, RSI, RBP, RSP registers. The r variable should be replaced by nothing for 16-bit size, E for 32-bit size, or R for 64-bit size. RAX. 64-bit version of EAX register. RAZ. Read as zero (0), regardless of what is written. RBP. 64-bit version of EBP register. RBX. 64-bit version of EBX register. RCX. 64-bit version of ECX register. 426 Glossary Chapter
32559
Rev. 3.08
July 2007
RDI. 64-bit version of EDI register. RDX. 64-bit version of EDX register. real-address mode. See real mode. real mode. A short name for real-address mode, a submode of legacy mode. relative. Referencing with a displacement (also called offset) from an instruction pointer rather than the base of a code segment. Contrast with absolute. reserved. Fields marked as reserved may be used at some future time. In order to preserve compatibility with future processors, reserved fields require special handling when read or written by software. Reserved fields may be further qualified as MBZ, RAZ, SBZ or IGN (see definitions). Software must not depend on the state of a reserved field, nor upon the ability of such fields to return state previously written. If a reserved field is not marked with one of the above qualifiers, software shall not change the state of that field; it must reload that field with the same values returned from a prior read. REX. An instruction prefix that specifies a 64-bit operand size and provides access to additional registers. rFLAGS. 16-bit, 32-bit, or 64-bit flags register. Compare RFLAGS. RFLAGS. 64-bit flags register. Compare rFLAGS. rIP. 16-bit, 32-bit, or 64-bit instruction-pointer register. Compare RIP. RIP-Relative Addressing. 4-bit instruction-pointer register. Addressing relative to the 64-bit RIP instruction pointer. Compare moffset. RSI. 64-bit version of ESI register. RSP. 64-bit version of ESP register. SBZ. Should be zero. If software attempts to set an SBZ bit to 1, it results in undefined behavior. set. To write a bit-value of 1. Compare clear. SIB. A byte following an instruction opcode that specifies address calculation based on scale (S), index (I), and base (B). SIMD. Single instruction, multiple data. See vector. SP. Stack pointer register. SS. Stack segment register. SSE. Streaming SIMD extensions instruction set. See 128-bit media instructions and 64-bit media instructions.
Chapter
Glossary
427
32559
Rev. 3.08 July 2007
SSE2. Extensions to the SSE instruction set. See 128-bit media instructions and 64-bit media instructions. sticky bit. A bit that is set or cleared by hardware and that remains in that state until explicitly changed by software. TOP. x87 top-of-stack pointer. TPR. Task-priority register (CR8), a new register introduced in the AMD64 architecture to speed interrupt management. TR. Task register. TSS. Task state segment. underflow. The condition in which a floating-point number is smaller in magnitude than the smallest non-zero, positive or negative number that can be represented in the data-type format being used. UP. One processor or uniprocessor. vector. (1) A set of integer or floating-point values, called elements, that are packed into a single operand. Most of the 128-bit and 64-bit media instructions use vectors as operands. Vectors are also called packed or SIMD (single-instruction multiple-data) operands. (2) An index into an interrupt descriptor table (IDT), used to for accessing exception handlers. Compare exception. virtual-8086 mode. A submode of legacy mode. word. Two bytes, or 16 bits. x86. See legacy x86.
428
Glossary
Chapter
32559
Rev. 3.08
July 2007
Index of Register Names

B
BU Machine Check Address Register (MSR 040Ah) 226 Control Mask Register (MSR C001_0046h) 225 Control Register (MSR 0408h) 222 Status Register (MSR 0409h) 225
C
Capabilities Pointer Register (Function 0: Offset 34h) 49 Class Code/Revision ID Register (Function 0: Offset 08h) 48 Class Code/Revision ID Register (Function 1: Offset 08h) 74 Class Code/Revision ID Register (Function 2: Offset 08h) 89 Class Code/Revision ID Register (Function 3: Offset 08h) 131 Clear Interrupts Register (Function 3: Offset B4h) 174 Clock Power/Timing High Register (Function 3: Offset D8h) 176 Clock Power/Timing Low Register (Function 3: Offset D4h) 174 Config Base and Limit 0 Register (Function 1: Offset E0h) 84 Base and Limit 1 Register (Function 1: Offset E4h) 84 Base and Limit 2 Register (Function 1: Offset E8h) 84 Base and Limit 3 Register (Function 1: Offset ECh) 84 Configuration Address Register 0CF8h 43
D
DC Machine Check Address Register (MSR 0402h) 218 Control Mask Register (MSR C001_0044h) 215 Control Register (MSR 0400h) 214 Status (MSR 0401h) 215 Device/Vendor ID register (Function 0: Offset 00h) 47 Device/Vendor ID Register (Function 1: Offset 00h) 74 Device/Vendor ID Register (Function 2: Offset 00h) 88 DeviceVendor ID Register (Function 3: Offset 00h) 130 Display Refresh Flow Control Buffers 165 DRAM Bank Address Mapping Register (Function 2: Offset 80h) 98 Base 0 Register (Function 1: Offset 40h) 77 Base 1 Register (Function 1: Offset 48h) 77 Base 2 Register (Function 1: Offset 50h) 77 Base 3 Register (Function 1: Offset 58h) 77 Base 4 Register (Function 1: Offset 60h) 77 Base 5 Register (Function 1: Offset 68h) 77 Base 6 Register (Function 1: Offset 70h) 77 Base 7 Register (Function 1: Offset 78h) 77 Config High Register (Function 2: Offset 94h) 111 Config Low Register (Function 2: Offset 90h) 109 CS Base 0 Register (Function 2: Offset 40h) 90 CS Base 1 Register (Function 2: Offset 44h) 90
Chapter
429
CS Base 2 Register (Function 2: Offset 48h) 90 CS Base 3 Register (Function 2: Offset 4Ch) 90 CS Base 4 Register (Function 2: Offset 50h) 90 CS Base 5 Register (Function 2: Offset 54h) 90 CS Base 6 Register (Function 2: Offset 58h) 90 CS Base 7 Register (Function 2: Offset 5Ch) 90 CS Mask 0 Register (Function 2: Offset 60h) 95 CS Mask 1 Register (Function 2: Offset 64h) 95 CS Mask 2 Register (Function 2: Offset 68h) 95 CS Mask 3 Register (Function 2: Offset 6Ch) 95 CS Mask 4 Register (Function 2: Offset 70h) 95 CS Mask 5 Register (Function 2: Offset 74h) 95 CS Mask 6 Register (Function 2: Offset 78h) 95 CS Mask 7 Register (Function 2: Offset 7Ch) 95 Limit 0 Register (Function 1: Offset 44h) 78 Limit 1 Register (Function 1: Offset 4Ch) 78 Limit 2 Register (Function 1: Offset 54h) 78 Limit 3 Register (Function 1: Offset 5Ch) 78 Limit 4 Register (Function 1: Offset 64h) 78 Limit 5 Register (Function 1: Offset 6Ch) 78 Limit 6 Register (Function 1: Offset 74h) 78 Limit 7 Register (Function 1: Offset 7Ch) 78 Scrub Address High Register (Function 3: Offset 60h 154155 Scrub Address Low Register (Function 3: Offset 5Ch) 153 Timing High Register (Function 2: Offset 8Ch) 106 Timing Low Register (Function 2: Offset 88h) 104
32559
Rev. 3.08 July 2007
G
GART Aperture Base Register (Function 3: Offset 94h) 169 Aperture Control Register (Function 3: Offset 90h) 168 Cache Control Register (Function 3: Offset 9Ch) 171 Table Base Register (Function 3: Offset 98h) 170 Global Machine Check Capabilities Register (MSR 0179h) 207 Global Machine Check Exception Reporting Control Register (MSR 017Bh) 208 Global Machine Check Processor Status Register (MSR 017Ah) 207
H
Header Type Register (Function 0: Offset 0Ch) 49 Type Register (Function 1: Offset 0Ch) 75 Type Register (Function 2: Offset 0Ch) 89 Type Register (Function 3: Offset 0Ch) 131 HyperTransport Initialization Control Register (Function 0: Offset 6Ch) 59 Read Pointer Optimization Register (Function 3: Offset DCh) 177 Transaction Control Register (Function 0: Offset 68h) 54
I
IC Machine Check Address Register (MSR 0406h) 222 Control Mask Register (MSR C001_0045h) 220 Control Register (MSR 0404h) 219
430
Chapter
32559
Rev. 3.08
July 2007
Status Register (MSR 0405h) 220
L
LDT0 Buffer Count Register (Function 0: Offset 90h) 68 Bus Number Register (Function 0: Offset 94h) 70 Capabilities Register (Function 0: Offset 80h) 60 Feature Capability Register (Function 0: Offset 8Ch) 67 Frequency/Revision Register (Function 0: Offset 88h) 66 Link Control Register (Function 0: Offset 84h) 62 Type Register (Function 0: Offset 98h) 70 LDT1 Buffer Count Register (Function 0: Offset B0h) 68 Bus Number Register (Function 0: Offset B4h) 70 Capabilities Register (Function 0: Offset A0h) 60 Feature Capability Register (Function 0: Offset ACh) 67 Frequency/Revision Register (Function 0: Offset A8h) 66 Link Control Register (Function 0: Offset A4h) 62 Type Register (Function 0: Offset B8h) 70 LDT2 Buffer Count Register (Function 0: Offset D0h) 68 Bus Number Register (Function 0: Offset D4h) 70 Capabilities Register (Function 0: Offset C0h) 60 Feature Capability Register (Function 0: Offset CCh) 67 Frequency/Revision Register (Function 0: Offset C8h) 66 Link Control Register (Function 0: Offset C4h) 62 Type Register (Function 0: Offset D8h) 70 LS Machine Check Address Register (MSR 040Eh) 228 Control Mask Register (MSR C001_0047h) 227 Control Register (MSR 040Ch) 227 Status Register (MSR 040Dh) 228
M
MC0_CTL Register (MSR 0400h) 214 MC0_STATUS Register (MSR 0401h) 215 MC1_CTL Register (MSR 0404h) 219 MC2_CTL Register (MSR 0408h) 222 MC3_CTL Register (MSR 040Ch) 227 MCA NB Address High Register (Function 3: Offset 54h) 145 NB Address Low Register (Function 3: Offset 50h) 145 NB Configuration Register (Function 3: Offset 44h) 135 NB Control Register (Function 3: Offset 40h) 133 NB Status High Register (Function 3: Offset 4Ch) 142 NB Status Low Register (Function 3: Offset 48h) 139 MCG_CAP Register (MSR 0179h) 207 MCG_CTL Register (MSR 017Bh) 208 MCG_STATUS Register (MSR 017Ah) 207 MCi_ADDR Register (MSRs 0402h, 0406h, 040Ah, 040Eh, 0412h) 212213 MCi_CTL Registers (MSRs 0400h, 0404h, 0408h, 040Ch, 0410h) 209 MCi_CTL_MASK Registers (MSRs C001_0044C001_0048h) 210 MCi_STATUS Registers (MSRs 0401h, 0405h, 0409h, 040Dh, 0411h) 210 Memory-Mapped I/O
Chapter
431
Base 0 Register (Function 1: Offset 80h) 80 Base 1 Register (Function 1: Offset 88h) 80 Base 2 Register (Function 1: Offset 90h) 80 Base 3 Register (Function 1: Offset 98h) 80 Base 4 Register (Function 1: Offset A0h) 80 Base 5 Register (Function 1: Offset A8h) 80 Base 6 Register (Function 1: Offset B0h) 80 Base 7 Register (Function 1: Offset B8h) 80 Limit 0 Register (Function 1: Offset 84h) 81 Limit 1 Register (Function 1: Offset 8Ch) 81 Limit 2 Register (Function 1: Offset 94h) 81 Limit 3 Register (Function 1: Offset 9Ch) 81 Limit 4 Register (Function 1: Offset A4h) 81 Limit 5 Register (Function 1: Offset ACh) 81 Limit 6 Register (Function 1: Offset B4h) 81 Limit 7 Register (Function 1: Offset BCh) 81
32559
Rev. 3.08 July 2007
N
NB Machine Check Address Register (MSR 0412h) 229 Control Mask Register (MSR C001_0048h) 229 Control Register (MSR 0410h) 228 Status Register (MSR 0411h) 229 NB_CFG Register (MSR C001_001Fh) 401 Node ID Register (Function 0: Offset 60h) 52 Northbridge Capabilities Register (Function 3: Offset E8h) 182
P
PCI I/O Base 0 Register (Function 1: Offset C0h) 82 Base 1 Register (Function 1: Offset C8h) 82 Base 2 Register (Function 1: Offset D0h) 82 Base 3 Register (Function 1: Offset D8h) 82 Limit 0 Register (Function 1: Offset C4h) 83 Limit 1 Register (Function 1: Offset CCh) 83 Limit 2 Register (Function 1: Offset D4h) 83 Limit 3 Register (Function 1: Offset DCh) 83 Performance Monitor 0 Register (Function 3: Offset A0h) 174 Monitor 1 Register (Function 3: Offset A4h) 174 Monitor 2 Register (Function 3: Offset A8h) 174 Monitor 3 Register (Function 3: Offset ACh) 174 Power Management Control High Register (Function 3: Offset 84h) 167 Control Low Register (Function 3: Offset 80h) 167
R
Registers General MSRs 371 Model Specific Registers 395 Power Management 316 Routing Table
432
Chapter
32559
Rev. 3.08
July 2007
5051 51 51 51 51 51 51 51
Node 0 Register (Function 0: Offset 40h) Node 1 Register (Function 0: Offset 44h) Node 2 Register (Function 0: Offset 48h) Node 3 Register (Function 0: Offset 4Ch) Node 4 Register (Function 0: Offset 50h) Node 5 Register (Function 0: Offset 54h) Node 6 Register (Function 0: Offset 58h) Node 7 Register (Function 0: Offset 5Ch)
S
Scrub Control Register (Function 3: Offset 58h) 148, 151 SMM Save State (Offset FE00FFFFh) 261 SRI-to-XBAR Buffer Count Register (Function 3: Offset 70h) 156 Status/Command Register (Function 0: Offset 04h) 48
T
Thermtrip Status Register (Function 3: Offset E4h) 179
U
Unit ID Register (Function 0: Offset 64h) 53
X
XBAR-to-SRI Buffer Count Register (Function 3: Offset 74h) 163
Chapter
433
32559
Rev. 3.08 July 2007
434
Chapter

AMD 64 BIOS and Kenrel Dev's Guide

Uploaded by

Copyright:

Available Formats

AMD 64 BIOS and Kenrel Dev's Guide

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

AMD 64 BIOS and Kenrel Dev's Guide

Uploaded by

Copyright:

Available Formats

BIOS and Kernel Developer's Guide for AMD NPT Family 0Fh Processors

Publication # 32559 Issue Date: July 2007

Rev. 3.08 July 2007

Rev. 3.08 July 2007

Rev. 3.08 July 2007

Rev. 3.08 July 2007

Rev. 3.08 July 2007

Rev. 3.08 July 2007

Rev. 3.08 July 2007

Rev. 3.08 July 2007

December 2006 3.04

Initial Public Release

Introduction and Overview

About This Guide

Rev. 3.08 July 2007

31118 31117 31731 31119 33953 33954 33952 26633

32769 25481 33610 http://www.hypertransport.org

Conventions and Definitions

Register Differences in Revisions of AMD NPT Family 0Fh Processors

Rev. 3.08 July 2007

Introduction and Overview

Introduction and Overview

Rev. 3.08 July 2007

Introduction and Overview

Processor Initialization and Configuration

Bootstrap Processor Initialization

Detecting AP and Initializing Routing Table

Rev. 3.08 July 2007

Initializing Noncoherent HyperTransport Technology Devices

Initializing Link Width and Frequency

Initializing the Memory Controller on All Processor Nodes

Initializing the Address Map Table

Application Processor Initialization

Processor Initialization and Configuration

Rev. 3.08 July 2007

BIOS Requirement for 64-Bit Operation

Sizing and Testing Memory above 4 Gbytes

Using BIOS Callbacks in 64-Bit Mode

Processor Initialization and Configuration

64-Kbyte 64-Kbyte L1 I-Cache L1 D-Cache

DDR2 SDRAM Interface

3 x 16/16 4002000 MT/s

Control PLLs and Clocks JTAG and Debug

Processor Block Diagram Functional Description 29

128-bits DDR SDRAM 200/266/333/400 MHz 16-bits ECC

Rev. 3.08 July 2007

Instruction Set Support

The processor is implemented in the following packages.

Three 16-bit Targets the server and workstation market.

Internal Cache Structures

HyperTransport Technology Overview

Rev. 3.08 July 2007

Package F(1207) AM2 S1g1

DRAM Interface Speed vs. CPU Core Clock Multiplier

Multiplier 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5

Rev. 3.08 July 2007

DRAM Interface Speed vs. CPU Core Clock Multiplier

Multiplier 10 10.5 11 11.5 12 12.5 13 13.5 14 14.5 15 15.5

Note: Fractional (half-step)multipliers are supported in revision G and later revisions.