An ASIC Primer

Table of Contents
Preface - Preface ..........................................................................................0-1

Chapter 1 - What is an ASIC? ...................................................................1-1

Section 1 - Uses of ASICs ............................................................................1-1

Section 2 - Types of ASICs ..........................................................................1-2

Array-Based ASICs ....................................................................1-3

Cell-Based ASICs .......................................................................1-3

Application Specific Standard Products (ASSPs) .......................1-4

Section 3 - The Characteristics of ASICs .....................................................1-4

Complexity ..................................................................................1-5

Silicon Efficiency ........................................................................1-6

NRE .............................................................................................1-6

Inventory Risks ...........................................................................1-7

Design Risks ...............................................................................1-7

Prototype Turnaround Time (TAT) .............................................1-8

EDA Tools (The old CAD / CAE Tools) ....................................1-9

Performance ................................................................................1-10

Section 4 - Array-Based vs. Cell-Based Trade-offs .....................................1-10

Why doesn’t LSI Logic do Gate Arrays Anymore? ...................1-11

 LSI Logic Corporation 1988 - 2000 i-1

An ASIC Primer

Section 5 - Standard Product ASIC Products (ASSPs) ................................1-12

Section 6 - Embedded Arrays .......................................................................1-13

Section 7 - Summary ....................................................................................1-14

Chapter 2 - Design Trade-offs ....................................................................2-1

Section 1 - Use NAND Gates Rather than NOR Gates ................................2-2

Section 2 - Proper use of Buffers ..................................................................2-4

Section 3 - On Chip memory vs Off chip memory ......................................2-9

Section 4 - The Use of a PLL .......................................................................2-12

Section 5 - Hard Coded vs. Soft Coded Logic .............................................2-13

Section 6 - Cores and Megacells ..................................................................2-15

Section 7 - Some Limitations of Synthesis Tools .........................................2-17

Section 8 - I/O Cells .....................................................................................2-19

Section 9 - ESD Cells ...................................................................................2-20

Chapter 3 - Design for Speed .....................................................................3-1

Section 1 - Reduce the Effect of Wire Delays ..............................................3-1

Section 2 - Use complex cells where appropriate ........................................3-2

Section 3 - Feed Late-Occurring Signals into the Last Logic Levels ...........3-3

Section 4 - Use Shift Counters in Place of Binary Counters ........................3-4

Section 5 - Use Duplicate Logic to Reduce Fanout and Increase Speed ......3-5

Section 6 - Use Input of Floorplanning ........................................................3-6

i-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Chapter 4 - Designing for Reliability ........................................................4-1

Section 1 - Use Synchronous Design ...........................................................4-1

Section 2 - Avoid “Glitch” Generators .........................................................4-4

Section 3 - Use Care with Gated Clocks ......................................................4-6

Section 4 - Avoid Race Conditions ...............................................................4-8

Section 5 - Avoid Floating Nodes on Internal Three-State Buses ................4-11

Section 6 - Use Bidirectional Output Buffers with Pull-Up or Pull-Down

Resistors to Increase the Noise Margin .....................................4-13

Section 7 - Avoid Excessive Fanout .............................................................4-15

Section 8 - Avoid “Dangerous” Decoding ....................................................4-16

Section 9 - Do Not Use Unbuffered Transmission Gates .............................4-19

Section 10 - Avoid Bus Contention ..............................................................4-20

Section 11 - Buried Buses: Sometimes Yes, Sometimes No ........................4-22

Section 12 - Follow I/O Guidelines Correctly ..............................................4-24

Section 13 - Power Considerations ...............................................................4-27

Speed ...........................................................................................4-28

Package Selection .......................................................................4-28

FIT rate ........................................................................................4-28

Contributing factors ....................................................................4-29

Conserving power .......................................................................4-31

 LSI Logic Corporation 1988 - 2000 i-3

An ASIC Primer

Power distribution .......................................................................4-32

Power isolation ............................................................................4-33

Power and Testability ..................................................................4-33

Chapter 5 - Designing for Testability ........................................................5-1

Section 1 - Testing Redundant Logic ...........................................................5-1

Section 2 - Use Combinational Logic ..........................................................5-2

Section 3 - Add Test Lines to Sequential Logic ...........................................5-3

Section 4 - Use Reset Signals to Initialize Storage Devices .........................5-6

Section 5 - Make Logic Observable by Multiplexing Output Pins ..............5-8

Section 6 - Decode Input Pins ......................................................................5-8

Section 7 - Break Up Long Sequential Chains .............................................5-10

Section 8 - Use Level-Sensitive Design .......................................................5-11

Section 9 - Partition Logic into Blocks ........................................................5-13

Section 10 - Do Not Design Your Own Storage Devices .............................5-14

Logic Considerations ..................................................................5-16

Section 11 - Use a Serial Scan Approach .....................................................5-17

Section 12 - Multiplex Output Pins to Make Scan Chain Observable .........5-19

Section 13 - Implementing the Vil/Vih (NAND) Tree .................................5-20

What is the NAND Tree? ............................................................5-20

Why is it important to have a combinational path? ....................5-20

i-4  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

How is the NAND tree sequenced? ............................................5-21

How is the NAND tree tested? ....................................................5-22

Subtleties .....................................................................................5-23

The PROCess MONitor ..............................................................5-25

The PROC_DRV .........................................................................5-26

The IIDDTN ................................................................................5-27

The PROCMON2A .....................................................................5-28

lsints ............................................................................................5-29

Chapter 6 - Automatic Tester Requirements ............................................6-1

Section 1 - Overview ....................................................................................6-1

Section 2 - Test Flow ....................................................................................6-2

Section 3 - Automatic Tester Configuration .................................................6-3

Section 4 - Test Period Length .....................................................................6-7

Section 5 - Input Timesets ............................................................................6-11

Section 6 - Examples of Valid and Invalid Waveforms ................................6-17

Section 7 - Output Strobes ............................................................................6-19

Section 8 - Rules for Creating Time Blocks .................................................6-22

Chapter 7 - Automatic Tester Requirements ............................................7-1

Section 1 - Overview ....................................................................................7-1

Section 2 - Prepare Simulation Test Patterns ................................................7-2

 LSI Logic Corporation 1988 - 2000 i-5

An ASIC Primer

Prepare System Simulation Input Patterns ..................................7-2

Prepare Tester Functional Simulation Input Patterns ..................7-3

Prepare Tester VIL/VIH Simulation Input Patterns ....................7-4

The Required VIL/VIH NAND Tree Test Circuit ......................7-4

Prepare Test Patterns for the Chip IDD Measurement ................7-9

Prepare Three-State Test Pattern .................................................7-10

Section 3 - Determine the Output Strobe Timing .........................................7-11

Section 4 - Avoid Voltage Spikes and Bus Contention .................................7-13

Avoid Voltage Spikes ..................................................................7-13

Avoid Bus Contention .................................................................7-15

Chapter 8 - The Beginning .........................................................................8-1

More on the Testing Issue ...........................................................8-2

Training Classes ..........................................................................8-2

Questions about this Book ..........................................................8-3

Summary .....................................................................................8-3

i-6  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

List of Figures
Figure 2.1 - Comparison of 2-input NAND and NOR Gates ........................2-2
Figure 2.2 - Crowbar Current ......................................................................2-4
Figure 2.3 - I/O Slot Used as Internal Buffer ...............................................2-5
Figure 2.4 - Repeater Buffers .......................................................................2-7
Figure 2.5 - Split Fanout ..............................................................................2-8
Figure 2.6 - A Bad Place for a Memory .......................................................2-12
Figure 3.1 - Capacitance Due to Cells vs. Capacitance Due to Wires ........3-2
Figure 3.2 - Comparison of 2-input NAND and NOR Gates ........................3-3
Figure 3.3 - Comparison of 2-input NAND and NOR Gates ........................3-3
Figure 3.4 - 3-stage Johnson Counter with Count Sequence .......................3-4
Figure 3.5 - Duplicate Logic to Increase Speed ...........................................3-5
Figure 3.6 - Better Wire Models After Cell Placement ................................3-7
Figure 4.1 - Classic Prop-Delay Based Pulse (Glitch) Generator ..............4-2
Figure 4.2 - Synchronous Pulse Generator ..................................................4-3
Figure 4.3 - Synchronous Pulse Generator Timing Diagram ......................4-3
Figure 4.4 - Glitch Behavior ........................................................................4-4
Figure 4.5 - Glitch Generating Mux .............................................................4-4
Figure 4.6 - Timing Diagram of Mux Output Glitch ....................................4-5
Figure 4.7 - The Glitch Doesn’t Matter Here ..............................................4-6
Figure 4.8 - A Gated Clock ..........................................................................4-7
Figure 4.9 - Controlling Data Instead of Clock is Safer ..............................4-8
Figure 4.10 - A Race Condition ....................................................................4-9
Figure 4.11 - Waveforms for Figure 4.10 ....................................................4-9

 LSI Logic Corporation 1988 - 2000 ii-1

An ASIC Primer

Figure 4.12 - Another Race Condition .........................................................4-10

Figure 4.13 - Waveforms for Figure 4.12 ....................................................4-11
Figure 4.14 - Avoid Floating Nodes .............................................................4-12
Figure 4.15 - Floating Node Avoided w/ Extra Buffer/Decode ....................4-12
Figure 4.16 - Bidirects with Pullup and Pulldown .......................................4-14
Figure 4.17 - Excessive Power Consumed Due to Excessive Fanout ..........4-15
Figure 4.18 - Dangerous Decoding of Terminal Count ...............................4-16
Figure 4.19 - Decoding a Johnson Counter .................................................4-18
Figure 4.20 - Alternate Glitch-Free Terminal Count Decode ......................4-19
Figure 4.21 - Faulty Use of Transmission Gate ...........................................4-20
Figure 4.22 - Avoid Bus Contention .............................................................4-21
Figure 4.23 - Two Types of Buses ................................................................4-23
Figure 4.24 - Too Many Outputs Between Power/Ground Pads .................4-24
Figure 4.25 - Enough Ground for the Outputs .............................................4-25
Figure 4.26 - Poor Placement of a Clock .....................................................4-26
Figure 4.27 - Proper Placement of a Clock .................................................4-27
Figure 4.28 - CMOS Power Consumption ...................................................4-30
Figure 5.1 - Untestable Redundant Logic ....................................................5-1
Figure 5.2 - Testable Redundant Logic ........................................................5-2
Figure 5.3 - Combinational Logic is Easy to Test ........................................5-3
Figure 5.4 - Untestable Sequential Logic .....................................................5-4
Figure 5.5 - Testability Added by “Test” Input (Breaking the Loop) ..........5-5
Figure 5.6 - Use Reset Lines to Initialize Storage Elements ........................5-7
Figure 5.7 - Multiplex Output Pins ..............................................................5-8
Figure 5.8 - Decode Input Pins ....................................................................5-9

ii-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Figure 5.9 - Long Counter is Difficult to Test ..............................................5-10

Figure 5.10 - Two Short Counters are Easier to Test ..................................5-11
Figure 5.11 - Level Sensitive Design ............................................................5-12
Figure 5.12 - Safely Gated Enable Line .......................................................5-12
Figure 5.13 - Block Level Boundary Scan ....................................................5-13
Figure 5.14 - Block Partitioning Using Muxes ............................................5-14
Figure 5.15 - Customer-Designed Storage Device ......................................5-15
Figure 5.16 - Dangerous Use of a Transmission Gate Mux ........................5-16
Figure 5.17 - Scan Design ............................................................................5-18
Figure 5.18 - Sharing the Scan Output .........................................................5-19
Figure 5.19 - NAND Tree Structure .............................................................5-21
Figure 5.20 - NAND Tree Sequence .............................................................5-22
Figure 5.21 - NAND Tree Test .....................................................................5-23
Figure 5.22 - A Complete NAND Tree Circuit .............................................5-26
Figure 6.1 - Test Flow ..................................................................................6-2
Figure 6.2 - Fundamentals of Tester Operation ...........................................6-3
Figure 6.3 - Automatic Tester Equivalent Diagram .....................................6-5
Figure 6.4 - Example Showing Five Timesets ..............................................6-7
Figure 6.5 - Maximum Propagation Delay Path ..........................................6-8
Figure 6.6 - Defining a Shorter Test Period ................................................6-9
Figure 6.7 - Minimum Cycle Time ................................................................6-9
Figure 6.8 - Determining the Minimum Test Period ....................................6-11
Figure 6.9 - Four Types of Input Waveforms ...............................................6-14
Figure 6.10 - Idealized Input Test Patterns ..................................................6-15
Figure 6.11 - Legal Timeset Initialization ....................................................6-15

 LSI Logic Corporation 1988 - 2000 ii-3

An ASIC Primer

Figure 6.12 - Input Test Patterns Using NRZ Waveforms Only ...................6-17
Figure 6.13 - Adding RZ to NRZ Reduces the Number of Patterns ..............6-17
Figure 6.14 - Simulation Inputs Replicated by Timing Generators .............6-18
Figure 6.15 - Simulation Inputs that Can’t be Replicated by Timing
Generators ...............................................................................6-19
Figure 6.16 - Tester Tolerance .....................................................................6-21
Figure 7.1 - The Vil/Vih NAND Tree ............................................................7-5
Figure 7.2 - Example NAND Tree Input Patterns ........................................7-7
Figure 7.3 - A NAND Tree with Bidirects ....................................................7-8
Figure 7.4 - Test Patterns for NAND Tree with Bidirects ............................7-9
Figure 7.5 - Using Output Strobe #2 ............................................................7-11
Figure 7.6 - Long Path Delay, Dominated by Chip Circuitry ......................7-12
Figure 7.7 - Short Path Delay, Dominated by Tester Capacitance .............7-12
Figure 7.8 - NOR Gate Output Edge Asymmetry .........................................7-13
Figure 7.9 - NOR Output Asymmetry Generating a Spike/Glitch ................7-14
Figure 7.10 - Avoiding the Glitch .................................................................7-15
Figure 7.11 - Avoiding Bus Contention ........................................................7-16

List of Tables
Table 1.1 - Technology Comparisons ...........................................................1-11

ii-4  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Preface
Preface

This book has been written to serve as a general guide for the ASIC
design engineer. It covers a number of important fundamentals related
to design methodology relevant to application specific integrated
circuits (ASICs). Although different silicon technologies are
mentioned, this book does not focus on any one technology type.

It is recommended that you as an LSI Logic customer read this book

thoroughly. In this book we present some of the potential
considerations that may be necessary when designing your ASIC
system. If this is your first ASIC design, you will most likely benefit
from suggestions and comparisons made throughout this book. They
should help you decide what trade-offs to make in choosing particular
logic configurations for both performance and logic integrity.

In addition, test issues are discussed in order to emphasize the

growing need to design testability into your logic at the beginning of
your design and not as an afterthought.

 LSI Logic Corporation 1988 - 2000 0-1

Quit Table of Contents History Go Back I

An ASIC Primer

The following issues are discussed in Chapters 1 through 7:

Chapter 1: an overview of ASICs - their uses, types, and

characteristics

Chapter 2: some general design considerations - the use of

cores and memories

Chapter 3: designing for speed

Chapter 4: designing for reliability

Chapter 5: designing for testability

Chapter 6: automatic tester requirements

Chapter 7: preparing test patterns for simulation

0-2  LSI Logic Corporation 1988 - 2000

Quit Table of Contents History Go Back I

 An ASIC Primer

1
What is an ASIC?

An ASIC (Application Specific Integrated Circuit) is a semiconductor

device designed especially for a particular customer (versus a
Standard Product, which is designed for general use by any
customer). The two major categories of ASIC Technology are Array-
Based and Cell-Based (also referred to as Standard Cell). Array-
Based ASICs configure a customer's design at the metal layers,
whereas Cell-Based ASICs are uniquely fabricated at all layers of the
silicon process including the diffusion layers.

CHAPTER 1 - SECTION 1
Uses of ASICs
To save chip area, ASIC technology integrates the logic and much of
the memory formerly distributed among multiple ICs, thus improving
reliability, optimizing PC Board space, and reducing component
costs. In addition, the higher integration and smaller size results in
significantly better system performance. ASICs were originally used
solely to replace or consolidate TTL “glue” logic and consisted of
relatively low complexity logic. Improvements in design tools and
implementation software, in process technology, and in large pin
count packages, now integrate much more of the logic formerly

 LSI Logic Corporation 1988 - 2000 1-1

An ASIC Primer

distributed among numerous ICs onto a single, very large scale

System-On-a-Chip (SOC)TM.
CHAPTER 1 - SECTION 2
Types of ASICs
There are two general categories of ASIC technologies: mask
programmable ASICs and field programable ASICs. This book will
focus on the mask programmable ASICs. In general, field
programmable devices are reserved for smaller designs than are
mask programmable devices. Although the turn around time for
field programmable devices is less than that of mask programmable
devices, the programable devices tend to offer this improved turn
around time at the expense of performance. The metal lines and
logic that are part of the field programmable devices are still present
even after the device has been configured. So it is not uncommon to
“breadboard” a design using field programmable logic to get the
system design working, and then convert to a mask programmable
ASIC for production purposes. The overhead in terms of both
performance and die size is usually too high a price to pay to go
into high volume production with field programmable devices.
Some of the most popular field programmable technologies are:

- FPGA - PAL - PLA - HAL

The two most popular types of mask programmable ASICs that are
being used in a great many System on-a-Chip designs today are:

- Array-Based -Cell-Based

A new emerging category of ASICs is the Application Specific

Standard Product, or ASSP. This type of ASIC is discussed in more
detail in Section 1.5. Don't confuse the ASSP with the term “full
custom”, which refers to a chip designed from scratch, but without
the hooks needed to be used as an ASIC.

1-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Array-Based ASICs

Array-Based ASICs offer the fastest customized implementation of

a customer's logic onto a piece of silicon because of the small
number of masks required to define the function of the chip. The
diffusion layers have already been completed by the silicon
foundry, and wafers are sitting in inventory waiting for the
personalization provided by the customer. Most Gate Array
Product Families have several “masterslice” sizes available to allow
for some reasonable selection of die size that comes closest to the
size needed to implement customer logic. Gate Arrays will always
contain more gates than will be used to implement the custom
design simply because they have been built as general purpose pre-
constructed pieces of silicon, specifically so that turn around time
can be fast, since only the top few layers of metal must be
processed to complete a design definition. This concept is ideal for
designs that will either have relatively low production volume or for
designs that may need to be redone because of design changes or
enhancements. Many designers will build designs in a Gate Array
technology as proof of concept, and then redo the design in a Cell-
Based technology before going to volume production.

Cell-Based ASICs

Cell-Based designs offer better die size optimization than Array-

Based designs. A Cell-Based design consists of optimized building
blocks, and approximates the performance limits and silicon
efficiency of a full custom device. In Cell-Based designs, transistor
geometries come in variable sizes. Components are not pre-
processed (as in Array-Based), therefore cells can be located
anywhere on the die. We will see later that cells are actually
restricted to “cell columns” or “cell rows” soas to minimize the
amount of bends that the Power and Ground metal must make. This
also allows for easier “stacking” of cells if cells are placed in

 LSI Logic Corporation 1988 - 2000 1-3

An ASIC Primer

regular column or row structures rather than in a totally random

fashion.

Application Specific Standard Products (ASSPs)

Full custom designs afford ultimate optimization, but at a very large

cost in design time (with corresponding development costs). In
both Array-Based and Cell-Based technologies, cells are pre-
characterized - i.e., simulation models and layouts exist for each
cell. A full custom design, costly and time consuming as it is, has
advantages of speed and area over it's Array-Based and Cell-Based
counterparts. But full custom may not always be a practical option
for today's systems, which have short lifetimes and where time-to-
market can be a significantly more important consideration.
Application Specific Standard Products are indeed full custom
chips but are available off the shelf. They have been designed by the
ASIC vendor but have been defined to fit specific system
requirements. They are not programmable devices.

CHAPTER 1 - SECTION 3
The Characteristics of ASICs
The remarks that follow further discuss some trade-offs of ASICs
with respect to the following categories:

- Complexity - Design Risks

- Silicon Efficiency - Prototype Turnaround

- NRE - CAD / CAE Support

- Inventory Risk - Performance

1-4  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Complexity

Complexity here means the number of transistors (or the amount of

logic and/or memory) per given amount of area, plus the associated
interconnect capability. Current Array-Based and Cell-Based chips
accommodate as many as 20,000,000 usable logic gates on a single
die. The term gate refers to a two-input NAND gate that consists of
4 CMOS transistors, two n-channel and two p-channel transistors.
The standard “gate” is also a regular drive gate. This means that the
gate can drive a typical fanout of about 16 unit loads for the given
technology. You will note as we proceed that many new terms are
likely to start creeping into our discussion. A standard unit load is
equivalent to the amount of capacitance and current that one gate
would have to drive while driving one input of another one of our
basic “gates”. Array-Based technologies, such as the LSI Logic
LCA500K Channel-Free Array, could implement up to about 2
million gates of random logic, up to 2 Megabits of mutliport RAM,
and up to 6 Megabits of ROM. It is true that full custom designs
can accommodate more gates, however Array-Based designs -
especially in a Channel-Free Array technology - are capable of
realizing functions that represent actual system building blocks and
incorporate system memory functions on the same die. The Array-
Based memories do tend to be about 5 times less dense than Cell-
Based memories because they are constructed out of the gates on
the masterslice. And full custom memories would provide much
higher densities than do Array-Based memories. But in fact many
designers who are using the Array-Based technologies to get fast
turn around tend to be using very small “scratch pad” or “cache”
types of memories which fit very well into the ASIC concept.

The newer Cell-Based ASICs, such as the LCBG12P Product, can

incorporate about 20 million usable random logic gates, 12
Megabits of RAM, and 32 Megabits of ROM.

 LSI Logic Corporation 1988 - 2000 1-5

An ASIC Primer

Silicon Efficiency

Array-Based technologies focus on fast implementation of logic

integration onto a single chip, rather than on absolute highest
density. Cell-Based designs allow you to get more logic onto a chip
in a given area. Cell-Based designs feature transistors and routing
tracks whose gradations of size are finer than those in Array-Based
products. Thus Cell-Based designs use silicon more efficiently than
Array-Based designs.

NRE

NRE (“Non-Recurring Engineering”) charges are the costs

associated with developing an ASIC. NRE is based on a number of
factors. Some of these factors are: the complexity of the design, the
technology chosen (# of masks required), the amount of work to be
done by the customer and by the silicon vendor, the need for special
cells or procedures, the type of package required, the schedule, the
number of layers of metal, and so on. So NRE is not something
which can be selected out of an ASIC “bluebook!” Normally, the
more work the silicon vendor does and the more special the
requirements, the higher will be the NRE. The more work the
customer does, the lower the NRE. If the silicon vendor takes a
customer design at the RTL level and does all of the synthesis, test
insertion, pattern creation, simulation, static analysis, bonding
diagram work, floorplanning, and static timing analysis, then the
customer is likely to pay a significant NRE charge. This work is
engineering intensive. If the customer does most of the
aforementioned work and the silicon vendor completes the layout,
makes the masks, and FABs the silicon, then the NRE will be lower
since more of the engineering workload resides with the customer.
Since much of the manufacturing cost is pure cost of materials
(masks, wafers, test hardware, packages, etc.), this cost will be
included in the NRE charge. The development of special cells,

1-6  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

custom memory compilers, new packages, and other special

requirements will all add incrementally to the NRE charges.

Array-Based designs require the fewest number of design-specific

masks and therefore offer the lowest NRE to prototypes. Cell-
Based designs require all masks to be generated for the chosen
process and therefore the NRE charge will be higher for a Cell-
Based design than for an Array-Based design.

Inventory Risks

A non-ASIC standard product, inasmuch as it can be used for a

variety of purposes, has the least inventory risk. ASICs are more
unique in their architecture and tend to be more focused in their
uses - more customer specific. The more application specific your
silicon is to your design, the more risk there is associated with
building large inventories (wafers or dies). Array-Based chips have
less inventory risk than Cell- based chips, at least for the vendor,
because the masterslices can be used for many different customer
designs. In Cell-Based technologies, all masks are customized per
design and are therefore unable to be used for multiple customers.
This means that if you over order silicon wafers for a Cell-Based
design, you will pay the cost of all of the wafers you order, whether
or not you use them.

Design Risks

The penalty for discovering a design error is higher for a Cell-

Based ASIC than for an Array-Based ASIC. Mistakes after
prototype fabrication in Array-Based designs usually only require
that the metal mask layers be redone. On the other hand, design
changes for a Cell-Based design may require that all masks be
redone. Since masks cost over $4000 per layer in today's deep sub-
micron world, the cost of materials will be much higher if a

 LSI Logic Corporation 1988 - 2000 1-7

An ASIC Primer

redesign is required in a Cell-Based technology. This must be

balanced, however, against the tremendous advantages that Cell-
Based technology brings to the ultimate die size (silicon area cost)
and performance (much higher than for the Array-Based
counterpart). Ironically today, as processes go more and more
towards more layers of metal, a gate array design may begin to
require a significant number of masks. For a 4 or 5 layer metal
process, at least 10 masks would be required for the 5 metal layers
and 5 via layers. In this type of process, the cost of a mask set for a
Gate Array begins to increase to a point where one of the main
advantages of Gate Arrays over Cell-Based designs starts to
disappear.

Prototype Turnaround Time (TAT)

Designs that require a complete mask set (Cell-Based) will always

require more time to manufacture than designs which use a basic
set of diffusion masks and only require customization at the metal
layers (Array-Based). This difference in time could be anywhere
from one week to 4 weeks depending on how fast the silicon vendor
can get masks from the mask shop and depending on how long the
FAB cycle is for a given process. Obviously, the more layers of
metal (more FAB steps), the longer the process will take. And
although multiple layers of metal can help to ease the interconnect
difficulties, each process step by definition will result in some yield
loss. So more layers of metal does not automatically mean
“better!” Also, in many cases, the silicon vendor will identify the
location of the cells far enough in advance of final signoff, that the
diffusion processing can actually start before the final signoff. This
helps reduce the FAB time required for the diffusion layers.

1-8  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

EDA Tools (The old CAD / CAE Tools)

The evolution of EDA tools over the last 20 years has been mind
boggling! In the 70s, vendors were struggling with the idea of
automated layout. You may have seen things like high speed color
plotters, digitizing tables, light tables for checking the electronic
version against the hand drawings against the plots, and so on. And
of course, companies were “cutting rubies” to send to mask
vendors. Well, most of the layout tasks have been automated with
very sophisticated place and route software, automated checking
tools, and automated flows between the silicon manufacturers and
the mask vendors.

Probably the biggest impact on IC design though has come as a

result of the development of true design software rather than just
software that translated, checked, and drew pretty pictures.
Although designers were impressed with the graphical tools of the
80s such as schematic editors, waveform editors, graphical
synthesis tools, and the like, it soon became obvious that to build
large systems these graphical tools were too slow. The synthesis
tools that have evolved, along with the ATPG tools for test, the high
speed simulation tools, the static timing analysis tools, and the
formal verification tools are now considered to be standard design
tools.

And one of the most forward looking steps taken by many

companies today, is to recognize that one company cannot create
the best EDA tools for every part of the complex tasks that need to
be performed to make workable chips today. Thus, many silicon
vendors have begun using “3rd Party” EDA tools along with some
in house software. The in-house software is usually for proprietary
tools that perform a function unique to the silicon company's
success. These tools are also tools that EDA vendors would not
create because the market demand is not large enough.

 LSI Logic Corporation 1988 - 2000 1-9

An ASIC Primer

The difficulties today are that the design sizes are becoming so
large that the run times for many EDA tools is becoming
prohibitive. This is spurring interest in hierarchical EDA tools.

Performance

Since the beginning of IC design, the two most critical parameters

that have been used to measure the worth of new technologies have
been speed and power. From a performance standpoint, both of
these parameters are important. High power circuits are normally
fast, but the increased power requires larger power supplies and
tends to heat up the junctions on silicon chips which slows the
devices. Each technology has brought some advantage to the world
of ICs. However in today's most dominant ASIC technology -
CMOS - high power can cause accelerated junction temperatures
which can slow down speed. One way to reduce the power and still
maintain speed is to develop circuits such as differential pairs that
do not to switch from voltage rail to voltage rail. The SRC (Silicon
Research Corp.) predicts as soon as the year 2005, the average chip
voltage supply could drop to as low as 1.0 volt.

CHAPTER 1 - SECTION 4
Array-Based vs. Cell-Based Trade-offs
As a short comparison of Array-Based and Cell-Based
Technologies, we can note the following:

1-10  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Table 1.1: Technology Comparisons

Gate Arrays Cell-Based

Density low - medium high

Performance low high
Memory Capacity low medium
Cost medium - high low - medium
Turn Around Time low medium - high
Design Time ~ 6 - 9 months ~ 9 - 12 months

Why doesn’t LSI Logic do Gate Arrays Anymore?

LSI Logic has always aspired to be a high-end ASIC company.

Several years ago, when the largest feasible chips were less than
50,000 gates, design cycles could often be measured in weeks or a
few months. The two or three week savings in turn-around-time for
a Gate Array design was then a significant portion of the total
design cycle, more than enough to offset the additional die size.
Also, the smaller size of the designs and lower complexity of the
process meant that both the engineering effort and mask costs were
lower; hence, the NREs were lower, and both LSI Logic and our
customers could afford to create ASICs targeted at lower volume
systems. Lower volume means that the NRE is a large portion of the
overall cost of the project, making the NRE savings available from
a Gate Array more attractive. Lower volume also means that the
additional die size of a Gate Array is not as important.

With today’s deep-submicron processes and multi-million gate

complexities come large teams of engineers for each chip, very

 LSI Logic Corporation 1988 - 2000 1-11

An ASIC Primer

expensive mask sets, longer total design cycles, and much higher
NREs. LSI Logic and our customers are forced to consider only
higher volume designs to offset the additional NRE, so the higher
NRE of a cell-based design is no longer an impediment when
compared to the total cost of the project. Also, the longer cycle
times mean that the turn-around-time advantage of a Gate Array is
no longer very large compared to the total design cycle. The higher
volume of the design means that it is imperative to make the die
size as small possible; the die area penalty of a Gate Array is no
longer acceptable. All these factors have led LSI Logic to abandon
the Gate Array concept upon which the company was founded.

CHAPTER 1 - SECTION 5
Standard Product ASIC Products (ASSPs)
One would think that many of the Standard Product companies
would have a big edge over ASIC companies in the area of ASSPs.
And the reason would be of course because of the wealth of chips
already designed and characterized. It should be just a matter of
taking off the chip I/Os and voila - an instant Core! Unfortunately,
it's not as easy as all that. Most of the standard products that were
designed over the last 20 years were designed as hand crafted chips.
In some cases every single transistor was individually designed.
And to go along with this “customization”, trade-offs were made
that are now unable to be overcome. For example, many of the
microprocessor chips that are available were designed over a 2 to 3
year period. But as they were designed, logic models or simulation
models were not built along with them. So although the transistors
and subcircuits were run through SPICE for circuit analysis, there is
basically no way today to provide a model of these standard
microprocessors to the Verilog and VHDL simulators. These
standard products were also hand laid out in many cases. So no
“auto place& route” topocell was built.

The ASSPs of today followed a different path. Many of the ASSPs

1-12  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

were done as ASICs and then made available as stand alone

functions that could be “packaged” as individual functions. And all
it took to make stand alone parts out of these Cores was to put some
I/Os on the boundary pins, and then package the device.

CHAPTER 1 - SECTION 6
Embedded Arrays
We have covered the fundamental features and benefits of Array-
Based technology and Cell-Based technology, and we have
introduced a new type of integrated product called the Application
Specific Standard Product or ASSP. There is one other design
concept that has existed in the ASIC world - it is called the
Embedded Array approach. Although this type of product is
different than either Array-Based or Cell-Based, it is not totally
different than either one of these two types of ASIC. As a matter of
fact, it exists as a marriage of the two technologies.

The basic concept of an Embedded Array is that a designer

“embeds” either large memories or cores or both into their design
and then fills in the area around the large cells with standard gate
array transistors. Thus the chip is customized because the designer
chooses the large cells that will be used on the design and where
they will be placed, but the random logic area will be designed and
laid out using standard gate array “masterslice” areas.

The big advantage of Embedded Arrays over either Array-Based or

Cell-Based designs is that you get BOTH the density of cores and
embedded memories from the Cell-Based products AND the
turnaround time of Array-Based products. The embedded cells are
floorplanned into the “custom masterslice” well in advance of final
place and route so that the diffusion masks can be defined and
wafers can be processed up to the metal layer steps. This
turnaround time advantage comes because you are not using Cell-
Based cells for the random logic portion of your design.

 LSI Logic Corporation 1988 - 2000 1-13

An ASIC Primer

Of course the other major advantage of the Embedded Array

concept is that if you need to make a change or if you have multiple
designs that can all use the same “customized masterslice”, then
you will save a significant amount of time turning the design around
to get prototypes AND the cost will be significantly lower for
multiple designs off the same basic masterslice.

An Embedded Array design requires the same cost as a Cell-Based

design for the first run through simply because all masks are
customized. But the turnaround time to prototypes is still much
better than for Cell-Based designs.

So the trade-offs are that you give up some of the random logic
performance vs. that available in a 100% Cell-Based design so that
you can gain turnaround time and reduce overall costs. These trade-
offs can offer an attractive alternative to designers that need the
capability of getting their designs into silicon fast, but who also
need the capacity offered by high density memories and core
technology. However, it is also important to note that as the volume
goes up and the performance requirements get tougher, the
Embedded Array will again fall short of the advantage offered by
Cell-Based ASIC technologies.

CHAPTER 1 - SECTION 7
Summary
Now that we have introduced the major categories of ASICs, we
need to cover some of the methodology required to design a circuit
in the ASIC world. The following Chapters cover various pieces of
ASIC methodology starting with “Basic Building Blocks” in
Chapter 2.

Figure 1.1

1-14  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

2
Design Trade-offs

Many trade-offs are going to have to be made during the course of a

design project. In this Chapter we will discuss some of the trade-offs
that must be considered. The trade-offs will involve giving up one
benefit to gain an even greater benefit. We will consider:

- NAND logic vs. NOR logic

- The proper use of Buffers

- On chip memory vs. off chip memory

- The use of a PLL

- Hard Coded logic vs. Soft Coded logic

- Cores and Megacells

- Some limitations of Synthesis tools

- I/O Cells

- ESD Cells

 LSI Logic Corporation 1988 - 2000 2-1

An ASIC Primer

Of course, these days most of the issues presented here are taken
care of by the design tools, but it’s still good for the designer to be
aware of what’s going on. Rather than suggest hard and fast rules
for every design, the discussions in this chapter are meant more to
make designers aware of the pros and cons of choosing a particular
course of action.

CHAPTER 2 - SECTION 1
Use NAND Gates Rather than NOR Gates
Because the mobility of the electrons in N-channel transistors is
approximately twice as fast as the mobility of the holes in P-
channel transistors, N-channel transistors exhibit about one-half of
the ON channel resistance of P-channel transistors of the same size.
Figure 2.1 shows that the P-channel transistors of the NAND gate
are in parallel. The P-channel transistors of the NOR gate are in
series.
B A B A

“NAND” “NOR”

Figure 2.1
Comparison of 2-input NAND and NOR Gates

2-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Therefore, the low-to-high and high-to-low propagation delays of

the NAND gate are more nearly symmetrical than those of the NOR
gate, with a resulting improvement in average speed. Again, even
though synthesis tools will normally take this discrepancy into
account, the engineer should be aware of it.

It is possible in a cell-based architecture to compensate for this

behavior by making the P-channel transistors larger, but then the
penalty is one of area rather than speed. For designs that are pad
limited, where the core area of the chip is very sparsely utilized,
using a larger transistor to get more balanced rise and fall times is
definitely an option.

As with any benefit however, there is usually a side effect. Larger

transistors have lower channel resistance and can therefore provide
more current to charge or discharge capacitance. However, when
input ramptimes become slow, it is possible for both the N and P
channel transistors to be on for some period of time. When this
happens, current can flow from the power supply to the VSS pin
until one of the two transistors shuts off. This current is referred to
as "crowbar" current and can cause an overall increase in chip
power dissipation. See Figure 2.2 for a diagram that shows the
effect of "crowbar" current.

 LSI Logic Corporation 1988 - 2000 2-3

An ASIC Primer

Vdd

Vthp Crowbar Current

Vthn VIN Out

Vin

time Vss

Both Transistors
ON!

Figure 2.2
Crowbar Current

So it is suggested that a designer be cautious in the use of large

output transistor cells.

CHAPTER 2 - SECTION 2
Proper use of Buffers
Buffers are required when a signal is routed either to a large number
of destinations, or to a destination far away, or both. In the case of
the far away destinations, the long metal run will add significant
capacitance to the path parasitic model. In the case of the high
fanout, the sum of the input capacitances will also cause a large
capacitive parasitic component. In either case, there are several
techniques which could be used to effectively "buffer" the
capacitive fanout, each with its own advantages and drawbacks.

The most straightforward technique would be simply to add a large

buffer at the source. But this begs the question of where the buffer
should be located; in the case where a signal drives a long wire, you
must ensure that the buffer is located near the source and not near
the destination. Then there is the question of how large a buffer to
use. Too small a buffer may do no good at all. Too large a buffer can
overload the driver, introduce a significant delay in its own right in

2-4  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

the case of a multi-stage buffer, or overdrive its output wire, causing

noise and reliability problems due to metal migration. And of
course we've already mentioned the problem of increased power
dissipation as the buffer gets larger.

Another possibility is to add several buffers, either serially in the

case of a long wire, or as a tree of buffers in the case of many
destinations. In this case, take care that the delay of the buffers
themselves isn't too large. Also, a tree of buffers may create skew
problems unless it is carefully planned.

Be aware that, in addition to buffers created from core transistors,

LSI Logic also offers the option of using I/O slot transistors as
internal buffers. The good news is that you can use these very large
buffers to drive many destinations from a single buffer to reduce
skew. The bad news is that, in doing so, you lose the use of an I/O
slot, and you must route the signal from the core logic to the edge
of the die and back as shown in Figure 2.3.

I/O slot buffer

Long wires

High fanout

Figure 2.3
I/O Slot Used as Internal Buffer

 LSI Logic Corporation 1988 - 2000 2-5

An ASIC Primer

In the LSI Logic G12 technology, the I/O slot buffer has been
replaced by a dedicated “megacell” having large geometry
transistors, but located within the core. You don’t lose an I/O slot,
but you still need to give the buffer extra power and ground routing,
which in many cases will limit its location to the periphery of the
die, so Figure 2.3 would still be more or less accurate.

Since most ASIC design toolsets used today rely on feedback from
the layout tools to some degree, the decision as to what type of
buffer scheme to use to resolve a delay or ramptime problem, may
in fact be made by the layout software given the correct user input.
For example, using a series of buffers along a wire (referred to as
repeater buffers) might be the overall best solution for driving a
fanout which is distributed all across the chip. The wire will be long
simply because the destination locations are all over the chip. But
driving a long wire from a single source, no matter how big the
driver is, will not be the most elegant solution. The concept of
repeater buffers will allow a signal to drive destinations across the
chip, and it will guarantee that no single piece of the wire in the
path is larger than a certain value. This means that none of the
repeater buffers in the path is likely to "see" a ramptime problem. It
also means that with a smaller capacitive load due to shorter wires
and smaller component fanout on any segment of the wire, the
repeater buffers can be smaller cells. This may mean that internal
cells are OK to use where a larger fanout may have even called for
the use of an I/O buffer slot. This tradeoff means that a designer
could save area and power dissipation by using repeater buffers.
The key here is that since the repeater buffers would be distributed
along the wire, they must have x-y coordinate information attached
to them. This means that cell placement information is necessary
for a tool or a human to be able to make this type of decision. See
Figure 2.4 for an example of the use of repeater buffers.

2-6  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Module A
A Die
Repeater
Wire Segment Buffers
lengths Do Not
exceed the
technology
maximum lengths Module B
which helps eliminate B
ramptime problems

Figure 2.4
Repeater Buffers

Many of the synthesis tools used today have a feature called "IPO”
or in-place optimization. The basic mechanism here is that the
software can choose a larger or smaller buffer based on fanout. The
difficulty is that the fanout is made up of both the destination
capacitance of the load plus the wire capacitance. But the wire
capacitance won't be known until the layout is complete.
Unfortunately, given the long run times associated with full layouts
for designs of today's size, engineers cannot always wait until the
layout is finished to make many of the decisions that will affect
design performance. The better ASIC toolsets provide designers
with an intermediate step to get layout type of feedback without
actually doing full layout. This step involves a chip floorplan. The
floorplan not only allows designers a means of refining wire delay
information, it can be tuned to a stage where cell placement
information is very close to that of the final layout. The only step
which is not performed is the actual wire routing. But given a good
cell placement, the details of the actual wire length can be
extrapolated with a fairly high degree of accuracy. This provides the
designer with the feedback in the form of a custom wireload model
which in many cases is close enough to the final routed wire model
that a designer can begin to make performance decisions without

 LSI Logic Corporation 1988 - 2000 2-7

An ASIC Primer

actually routing the design! Then, tools like the IPO algorithm can
decide whether or not to adjust the size of buffers needed to drive
specific high fanout lines.

One other technique that is possible in terms of "buffering" a signal

is to use the concept of split fanout. Figure 2.5 shows that instead of
driving the two high fanout modules from one driving source, the
driver could have been split into two buffers, where each buffer
drives a more limited fanout. By taking some care during layout,
any skew differences introduced by splitting the fanout can be tuned
to minimize overall skew between the two signals. It is important to
realize that by splitting the fanout, the cell driving the two buffers
might have to be enlarged to handle a larger fanout. It is also
important to realize that by splitting the fanout, the driver module
now has another port added to it. This may or may not be an issue
during a re-synthesis run if the original RTL code is not "told"
about the new port.

BUF1

A B
B
A
High Fanout of
Fanout BUF2 BUF1
BUF1
for BUF1

C Fanout C
of BUF2
Original Split Fanout
Die Die

Figure 2.5
Split Fanout

2-8  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

In summary, some of the buffering techniques that can be used are:

- increase the driver buffer size

- insert repeater buffers in the line

- use a buffer tree to limit individual buffer fanout

- split up the fanout so that each buffer drives a reduced

fanout

- use an IPO tool to adjust buffer size for load

- use an I/O buffer slot for very high fanout signals

Most designs will use some type of a buffering scheme. Many

designs will use a combination of some or all of the techniques
listed above.

CHAPTER 2 - SECTION 3
On Chip memory vs Off chip memory
ASIC technologies were originally designed to replace
combinatorial logic which was spread out over many small TTL
chips. In many of the very early systems, memory was not a big
consideration.

Today almost every single chip design includes some type and
amount of memory. So when does it make sense to put memory on
an ASIC chip? The basic decision is based on the criticality of the
design's performance. If it is crucial that the memory be right next
to the logic it interfaces with, then the ideal situation is to have both
pieces of the design on the same IC. But ASIC technologies do not
usually support the very high memory densities one might find in
DRAM chips. Even though the density of ASIC memories is
growing very fast, it will still never quite catch the density that can
be attained in the most dense memory technologies. One of the

 LSI Logic Corporation 1988 - 2000 2-9

An ASIC Primer

reasons is because ASICs still must provide for a way to build

random logic. Furthermore, today's designs include some analog
portions, some very high speed differential signals, very high
density cores that perform specific "non-memory" functions, and so
on.

So the ASIC chip has to do more than provide lots of memory.

Given this fact, ASICs can complement the high density memory
technologies.
It appears that many designs that are now being moved into ASIC
technologies do not require lots of high density memory. Many of
the memory needs are for cache memory, scratch pad memory, or
for small memories local to specific logic functions. So although
the amount of memory on board an ASIC chip may approach a
hefty size, much of it is spread out throughout the design so that no
one piece is considered "large".

In any case, memories are definitely special functions in that they

are always hard coded pieces of a design. This simply means that
the memory layouts are fixed once the configuration of the memory
is defined. Because of the organization of the bits, words, sense
amps and control logic associated with a memory it would not
make sense to have "soft-coded" memories where the layout was
not structured by definition.

The hard coding causes a couple of things to happen:

1. the memories usually have a fixed size and aspect ratio based
on configuration;

2. the I/O pins of the memories are usually in predetermined

locations around the periphery of the memory;

2-10  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

3. the density of the memories is usually such that no chip wires

may route over the memory in order to avoid bit errors induced
through crosstalk;

4. metal wires are typically much closer together inside the

memories which creates a balancing act between the wires
inside the memory and the wires on the rest of the chip;

5. due to the fact that memories tend to switch quite a bit of

capacitance when data changes (the bit lines are long),
memories can consume large amounts of power and therefore
usually require their own power bus connections to the I/O
power and ground pads;

Since hard coded entities like memories are densely populated

areas of a design, and since they are usually hard coded to
guarantee high performance or repeated performance, they tend to
switch quite a bit of current during normal operation. So these
entities carry their own power bus structure with them. It must be
handled carefully on the layout. Today most designers get involved
with the chip layout at least to the level of floorplanning. So even at
this level, the location of these large hard coded entities on the chip
must be carefully planned. The large power and ground buses that
usually accompany the hard coded blocks can be most effectively
dealt with if the large blocks are placed in a corner of the die or near
one edge of the die. If this cannot be done, as in the case where
there are too many of these blocks, then a special power bus matrix
will essentially have to be designed to handle the current flowing to
and from these blocks. These special power and ground matrices
can take up significant amounts of die area if not handled properly.
Figure 2.6 shows an example of potentially wasted area that results
when large cells like memories are placed in non ideal locations on
a die.

 LSI Logic Corporation 1988 - 2000 2-11

An ASIC Primer

Routing & Placement

Blockage
RAM 1

RAM Corner

Edge
RAM 2

I/O Regions
Bad Placement Acceptable RAM
of RAM Placement

Figure 2.6
A Bad Place for a Memory

Memory blocks present other challenges for chip designers

independent of their effect on chip layout. Issues such as test
introduce concepts such as memory BIST logic, built in self repair,
memory redundancy to allow for fixes without redoing masks, and
so on. A much more extensive discussion of memory testing can be
found in some of the Application Notes available in the ASKK
documentation on the WEB. Testing memories that are part of a
large ASIC design can be a non-trivial issue. The technique(s) used
to test the memories could either add quite a bit of additional logic
to the design or could complicate the test procedure required to
guarantee high defect coverage on the memories.

CHAPTER 2 - SECTION 4
The Use of a PLL
In the best of all worlds, all chip designs would be totally
synchronous and only have one clock signal. Alas, we live in the
real world. And unfortunately, even those designs that appear to be
synchronous can still suffer clock problems. Just about any scheme

2-12  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

used to distribute the incoming clock signal around a large design

will introduce serial delay into the clock signal before it actually
reaches the flip flops it will be controlling. The serial delay doesn't
necessarily create a big problem on chip since the skew between on
chip flip flops can be minimized if care is used in the layout of the
clock metal.

But the serial delay issue can cause a problem in a system where
multiple chips need to be synchronized. So in effect, we need a way
to synchronize the clock that reaches the flip flops on chip with the
incoming clock. A very effective way to do this is to use a Phase
Locked Loop module. This can basically guarantee that a designer
can adjust the phase of the clock on one chip so that the overall
system is synchronous. The PLL cell itself is relatively small, but it
introduces an annoying issue into the design: analog circuitry! This
requires that analog power and ground busses be added into the
design and brought out to chip pins properly so that they do not
interact with the digital power and ground busses on chip.

In general, the PLL will be able to adjust for just about any amount
of serial delay that could result from long wires and clock trees, but
the extra work required to handle the PLL cell correctly adds one
more item to a designers list of things to do.

CHAPTER 2 - SECTION 5
Hard Coded vs. Soft Coded Logic
Hard coded logic refers to logic that has been pre-layed out before
the actual place and route of the design itself begins. Some types of
logic almost require hard coded layouts just due to their nature.
Entities like memories are always hard coded simply because of
their regular structure which cries out for a fixed layout for all of the
bit cells due to their unique relationship to each other. Also, the
control signals for the memory are of such a regular nature that it
would be terribly inefficient to route them if the memory bit cells
were layed out in a random fashion.

 LSI Logic Corporation 1988 - 2000 2-13

An ASIC Primer

Generally speaking it is recommended that designers use soft coded

logic blocks whenever possible. This gives layout software a lot
more freedom to place cells and route metal than when hard coded
logic is used simply because hard coded cells tend to act as
blockages to place and route programs. This makes it more difficult
to complete a chip layout since alternate locations must be used to
find room to place and route the other cells and the wires in the
design.

All designs will have many signals that travel some distance across
the chip in order to reach all of their destination points. Given that
ASIC chips have regular orthogonal “tracks” for routing wires,
there are only so many routing tracks available on a given die size.
So it's important to make sure that routing "space" is used wisely.
When signals have to take detours around blockages, the bends or
corners actually use up more die area than a straight piece of metal
would, especially since changing direction usually means changing
routing layers. So if possible it would be desirable not to have to
bend any signal wires at all. Of course this is not possible, so the
goal is to try and minimize the overall number of bends in wires as
they route to and from their points of contact. The more hard coded
logic on the die, the more detours signal wires are likely to have to
take. This tends to cause longer wires, larger die sizes, and can
create congestion in areas where it would not have existed. So the
use of hard coded logic should be minimized.
Here are some cases where hard coded logic is actually required:

- memories

- need for repeatable performance

- special Cores

- macrocells

2-14  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

- critical paths

- datapaths

You may have a need for hard coded cells for reasons unique to
your design, but it is best if you can minimize the needs for cells
with fixed layouts.

CHAPTER 2 - SECTION 6
Cores and Megacells
LSI Logic has created a library of large system level building
blocks which implement many industry standard functions. For
some detailed information about LSI Logic’s CoreWare® modules,
visit the LSI Logic WEB page and specifically the section which
describes the CoreWare program. It is located at:

http://www.lsilogic.com/products/coreware/coreware.html#library

The types of cores that are generally available are functions that
have become very popular as building blocks for large systems. The
ATM core is very popular because many systems today deal with
asynchronous data which has to be moved from one place to
another. The JPEG and MPEG Cores allow designers to use
standard logic to handle still pictures and moving pictures without
having to design logic from scratch to implement these standards.
Many of the Cores handle bus interface issues for systems that have
to communicate with standard bus protocols. LSI provides Cores
for the PCI interface, the Fibre Channel interface, the USB
interface, the Ethernet interface, and others. In fact, as new and
more popular interfaces come along, LSI will introduce Cores to
allow designers to incorporate these standard protocols into their
system level chips.

 LSI Logic Corporation 1988 - 2000 2-15

An ASIC Primer

More and more Mixed Signal applications are in vogue today. And
LSI provides the most obvious Cores for the Mixed Signal
interfaces: the A/D and D/A converters. And many of the new I/O
cells that have been designed and are being designed, address the
area of Mixed Signal interface.

Much work has been done at LSI Logic in the area of DSP and
video compression. So in addition to Cores such as JPEG and
MPEG, you will find video compression chips, error detection and
correction chips, and new coding chips to allow for even denser
compression algorithms.

But it doesn't make sense to design these large Cores from scratch.
LSI provides these Cores as part of our growing Coreware library.
They are meant to be used in many cases much like designers use
basic low level cells - as is. In other words, the Core itself is already
built as a cell. In many cases it may include a hard coded topocell
which can be used by designers as a black box through the
synthesis process.

Coreware modules include deliverables such as complete

testbenches to exercise the Core(s), simulation shells for Verilog
and VHDL simulation, timing shells for PrimeTime Static Timing
Analysis, SPEF files for wire delay models, and full datasheets that
include information such as static and dynamic power dissipation
and switching performance parameters.

The LSI Logic FlexStream toolset includes programs which allow

designers to create specialized functions for their design work such
as datapath modules, which can be converted into hard coded
entities that can be dropped into designs to ensure repeatability and
optimized performance.

2-16  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

In addition to the CoreWare library, LSI has a team of Field

CoreWare engineers who specialize in certain core technologies
such as the Processor family: MIPS and SPARC RISC processors,
the Mini-RISC and Tiny-RISC processors, and some coprocessors.
Cores like the OAK Core, the ARM Core, and the HyperPHY Core
have dedicated engineering personnel who are available to work
with designers as applications support people to help optimize use
of these cores.

CHAPTER 2 - SECTION 7
Some Limitations of Synthesis Tools
Synthesis tools have improved dramatically over the last few years.
In fact, the improvements almost make one wonder if there is still a
need for engineers to make critical logic decisions! Well, as it
happens, even the best synthesis tools cannot know everything, and
some of the things that they don't know CAN HURT YOU!

Generally speaking, the two most well known synthesis tools used
by engineers are perhaps the Synopsys Design Compiler tool and
the Cadence Build Gates tool. And these vendors advertise that
these tools can synthesize up to about 200,000 gates in one run. No
doubt the tools actually can do this, but the question remains: "is it
wise to synthesize such a large piece of logic in one run"? LSI
Logic recommends that designers synthesize blocks to the 30K -
50K gate range in one run. When the blocks get much larger than
this, they may be creating logic which will occupy large sections of
the die. If these synthesized areas are too big, then guiding the
synthesis process with useful custom wireload information can be
difficult.

A few of the areas that need to be considered before synthesis are

the following:

• some foundry cells must be labelled as "don't use"

• memories cannot be synthesized by generic EDA tools

 LSI Logic Corporation 1988 - 2000 2-17

An ASIC Primer

• cores are foundry specific and must be black boxed

• I/O cells are not synthesized
• RTL code typically cannot specify timing constraints
accurately enough
• incomplete timing constraints can "mislead" synthesis tools
• initial wireload estimation techniques are usually inaccurate
• the way the RTL code is written can drive the type of logic
produced, so it needs to be checked for correctness before the
synthesis run.
Most often today, designers will go through a floorplanning cycle at
least once to get more accurate custom wireload data which is used
for a second synthesis run. And of course items such as the number
of layers of metal, what type of test logic is used, and the overall
clocking methodology will all help determine what the final
floorplan will look like. This information is extremely valuable for
the resynthesis run since it provides a physical reality check and
balance for the design.

Tools are beginning to emerge that will help designers examine

their RTL code, not only for synthesizability, but for design style
errors which can mislead the synthesis tools and create the wrong
type of logic. Within LSI, we have created a tool called lsivega
which can be used to do RTL code reporting and generate synthesis
scripts which follow the LSI Logic guidelines, and can also be used
to report on critical nets such as clocks and other control signals.
The tool generates design statistics and does some gate level DRC
checking, and provides a schematic browser capability. The scope
of this tool will grow over the next couple of years. Other tools such
as the SENTE Wattwatcher tool help designers do RTL power
estimation. Be prepared to see quite a bit of software that will be
written to help designers look at their RTL code BEFORE going

2-18  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

through possibly wasteful synthesis runs. We'll even be seeing RTL

floorplanning tools to help get better wire estimates without having
to synthesize to the gate level for a first pass.

CHAPTER 2 - SECTION 8
I/O Cells
Ten years ago, I/O cells consisted basically of input buffers, output
buffers of various current drive capabilities, and three-statable and
Bidirectional buffers. Today, the numbers of choices has grown
beyond our ability to memorize all of the types or names. We now
have mixed signal I/O with differential signal capability. We have
level translators to go from one voltage level to another on chip. We
have specialized GTL and NTL buffers, PCI buffers, the RAMBUS
interface I/O, very high speed cells, LVTTL cells, LVDS cells, and
on and on.

A designer used to have to decide which type of input threshold

they needed (CMOS or TTL) and what output current drive they
needed (4 ma, 8 ma, or 12 ma) and he/she was pretty much done!
Today, it's important to know how the I/O signals are being used in
the system soas to be able to select just the right combination of
specs. to meet the pin by pin needs of the system. In addition since
many systems today include JTAG logic (or boundary scan), it is
necessary to guarantee that each I/O cell has appropriate JTAG
logic available.

In addition to choosing the type of I/O cell to be used, it is

necessary that designers plan the number and location of power and
ground pads around the die so that issues such as crosstalk and
signal integrity are accounted for.

And in today's designs we must deal with "analog" power and

ground, cut power and ground buses, and ESD protection logic.

 LSI Logic Corporation 1988 - 2000 2-19

An ASIC Primer

CHAPTER 2 - SECTION 9
ESD Cells
It has always been true that CMOS ICs in particular have needed
special protection in the I/O area to prevent electrostatic damage to
the very thin gate oxide regions of the die. As chips have gotten
bigger, the types of I/O cells have gotten more varied, mixed
voltage buses now share regions over the I/O cells, and design
speeds have increased causing more simultaneous switching, and as
package materials and styles have become more diverse, the ESD
protection circuitry has had to evolve also to keep pace with a much
larger variety of options at the periphery of the die. LSI Logic
currently has two different types of ESD cells located in the I/O
region of the chip. These cells provide protection for both the I/O
and Core regions of the chip. Since ESD damage can occur when
the chip is powered down sitting on the storage shelf, it is even
necessary to protect against ESD damage between two ground pins
which are normally assumed to both be at 0 volts! One would not
intuitively know that protection circuitry was needed between two
different ground pins on the package (VSS and VSS2). But of
course when we consider that in a power down situation there may
be sneak paths through oxides when an ESD voltage builds up
between two ground pins, then we have to design circuitry to
protect against this case. And indeed, we do keep many signals such
as VSS and VSS2 separate on the die for good reasons. But when
we do this, it introduces the potential for damage that is non
obvious.

LSI Logic has designed two different types of ESD cells for the two
major regions of the die - the core and the I/O area. And of course
there are many variations of the I/O cell to accommodate the many
types of I/O cells which could all possibly be tied to their own cut
section of power bus.

2-20  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

The ESD cells occupy space and must be planned for in an ASIC
design because they add necessary reliability to the chip. In this
sense, they act as another type of I/O cell, since they occupy space
around the edge of the die, and one must obey certain rules about
spacing between ESD cells and to local power and ground pads.
There must be enough of them on the die to prevent parasitic
resistance from nulling out their effectiveness.

 LSI Logic Corporation 1988 - 2000 2-21

An ASIC Primer

2-22  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

3
Design for Speed

Most of what used to be considered “designing for speed” nowadays

means plugging tighter constraints into the synthesis tool. However, it
is important to know what kinds of techniques the synthesis tools do
automatically, since there are still occasions when an engineer must
do things the old fashioned way.

CHAPTER 3 - SECTION 1
Reduce the Effect of Wire Delays
In current deep-submicron technologies, delays resulting from
wirelengths have considerably greater relative effect on effective gate
delay than loading due to other cells. Thus it is imperative that you: 1)
properly floorplan your chip, 2) optimize drive strengths by using
high-drive cells when necessary, and 3) partition your logic in such a
way as to minimize the length of interconnect between cells.

The total loading capacitance is the sum of the output capacitance, the
wire capacitance, and the sum of input capacitances of the driven
functions. In both case (a) and case (b) in Figure 3.1, the sum of the
input loads is 12; but CW2 in case (b) represents the wire capacitance
to 12 places. (The term “input load” does not mean input capacitance,

 LSI Logic Corporation 1988 - 2000 3-1

An ASIC Primer

but it does relate to capacitance. Ultimately, LSI Logic uses the

total output loading number for delay calculations.)

IL=6

COUT CIN
IL=6

CW1
(a) CIN

IL=1

12 /
COUT CIN
IL=1

CW2
CIN
(b)

Figure 3.1
Capacitance Due to Cells vs. Capacitance Due to Wires

Interconnection trace resistance also affects delay; the LSI Logic

FlexStream™ software tools perform an RC tree analysis after
layout in order to guarantee very accurate post-layout delays.

CHAPTER 3 - SECTION 2
Use complex cells where appropriate
There is no such thing as a non-inverting CMOS logic gate. There
are only double inverting logic gates, and a few four or six stage
gates. There are, however, complex cells which perform multiple
levels of logic gating with only one inversion. With so much delay
in the wires, they can be faster than interconnected gates, but their
size, compactness, and high pin count can cause routing congestion.

3-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

CHAPTER 3 - SECTION 3
Feed Late-Occurring Signals into the Last Logic Levels
Place early-occurring signals in the first levels of your logic and
late-occurring signals in the last levels, as shown in Figure 3.2. This
allows the early-occurring signals time enough to set up on the last
stage. The later signal needs to propagate through only one low
fanin gate.

Early A ABC
occurring B
signal C ABC+D+E=ABCDE
D
E
ABCDEF
Late
occurring F
signal

Figure 3.2
Comparison of 2-input NAND and NOR Gates

This technique can also be used to speed up certain inputs on high

fan-in gates. The example in Figure 3.3 shows an 8-input NAND
gate used in the critical path. The propagation delay between all
inputs and the output is the same in (a). Signal H can be speeded up
by replacing the ND8 gate with the structure shown in (b).

A ABCD
B ND4
C ABCD+EFG=ABCDEFG
D
ND8 E ABCDEFGH
F ND3
EFG
G
H
Figure 3.3
Comparison of 2-input NAND and NOR Gates

 LSI Logic Corporation 1988 - 2000 3-3

An ASIC Primer

CHAPTER 3 - SECTION 4
Use Shift Counters in Place of Binary Counters
A Johnson counter is simply a shift register with the QN output of
the last stage connected to the data pin of the first stage. Johnson
counters are fast because there is no gating between the flip-flops in
the counter. However, a Johnson counter will use more flip-flops
than a binary counter. An n-stage Johnson counter gives a count
sequence of 2 x n counts. Thus, for a modulo eight counter, a three-
stage Johnson counter won’t be adequate and a fourth stage is
needed. Thus, the trade-off is speed versus gate count.

Qa Qb Qc

D Q D Q D Q
FD2 FD2 FD2

QN QN QN

CLEAR
CLK

Qa Qb Qc
0 0 0
1 0 0
1 1 0
1 1 1
0 1 1
0 0 1

Figure 3.4
3-stage Johnson Counter with Count Sequence

3-4  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

CHAPTER 3 - SECTION 5
Use Duplicate Logic to Reduce Fanout and Increase Speed
You can reduce fanout and increase the speed of your circuit by
allowing one of the outputs of the redundant logic to drive the
critical path while the other output drives the non-critical path, as
shown in Figure 3.5. Use this technique only if the timing skew
introduced by separating the two paths will not create race
conditions in the circuit.

This technique is useful only if the gain in speed from using

duplicate logic exceeds the loss of speed that results from
increasing the capacitive load of the previous macrocell.

D Q Fanout=10
speed critical
ND3 Fanout=10 FD2
speed critical

D Q Fanout=9
non-speed
ND3 Fanout=9 critical
non-speed FD2
critical

ND3 Fanout=1
speed critical Fanout=1
D Q
speed critical
FD2
Gates

Flip-Flops

Figure 3.5
Duplicate Logic to Increase Speed

(One output signal drives a critical path; another drives a non-

critical path)

 LSI Logic Corporation 1988 - 2000 3-5

An ASIC Primer

CHAPTER 3 - SECTION 6
Use Input of Floorplanning
As newer technologies emerge, component speeds get faster and
faster. This is good. However, die sizes are getting larger and larger
and the average length of any given wire in a design can potentially
be longer. Smaller geometries carry with them increased channel
resistance which tends to reduce charge current. In addition to this,
power supply values are dropping quickly. After almost 30 years,
the standard 5.0v volt supply is now giving way to 2.5v, 1.8v, 1.2v,
and yes even 1.0v! Once again this is good in that the power
dissipation can be reduced. However this reduced supply voltage
also has the side effect of reducing available charge current.

The above changes demand that designers take more care in the
layout of their designs. The best way to manage a design layout at
the design level is to be “hands on” involved with the design
floorplan.

Placing design modules in the correct location on a die can

significantly input performance. And in fact, floorplanning takes
designers through the placement process, not only for modules and
blocks, but also for components down to the cell level. Figure 3.6
shows how all placement data can help designers close timing
before final routing. Notice that in Figure 3.6a, although the design
has been floorplanned, since no cells have been placed, wirelength
is estimated based on Block sizes and locations. However if cells
“x” and “y” as shown in Figure 3.6 had been located in positions 1
and 2 or 3 and 4 as in Figure 3.6a, the “generic” wire estimate
(shown by the dotted line in Figure 3.6a) could be quite far off. This
before could be either overly optimistic or pessimistic, and you, the
designer, don’t know which is the actual value of the wire length. In
figure 3.6b, all cells have been placed and the estimated wirelength
between “x” and “y” will be much closer to the final value. Using

3-6  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Wire Estimated A Cells Placed

3A 1 Center of Block X Wire Estimated
A to Center from Cell X to
of Block B Cell Y
B 2 B
Y

4 A
Floorplan with no Floorplan with Cells
Cells Placed Placed
a b

Figure 3.6
Better Wire Models After Cell Placement

the placement information, a designer can choose from a number of

techniques to solve timing problems BEFORE going to final layout
and detailed routing.

So Floorplanning not only helps to located the system logic on the

die (Blocks A and B in Figure 3.6) but allows a designer to place all
cells prior to Backend Layout.
Analyzing the results of the Floorplanning and cell placement
information can allow designers to select strategies to solve timing
and ramptime problems before final layout. This tends to generate
fewer surprises after layout is finished.

The Customer Education Department will be producing on-line

Application Notes beginning this year (2000). One of these
addresses the issue of floorplanning in more detail. Suffice it to say,

 LSI Logic Corporation 1988 - 2000 3-7

An ASIC Primer

floorplanning is now a crucial step in the design process.

Many other considerations need to be made at the Floorplanning

level in order to close timing after final layout. The locations of
memories and cores, special I/O cells, the location of the clock and
the flip flops it drives, the location of boundary scan logic, the
location of power and ground pads and ESD cells, all contribute to
Timing Closure.

A designer MUST be willing to invest time in a good Floorplan and

cell placement in order for the router to complete its work and not
destroy the delicate tradeoff between ever smaller geometries and
bigger die sizes.

3-8  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

4
Designing for Reliability

This chapter discusses a number of issues which, if not properly

handled, can cause designs to fail. Alternatives are presented to show
comparisons between marginal logic design techniques and rigorous
techniques that lead to more reliable silicon.

CHAPTER 4 - SECTION 1
Use Synchronous Design
Asynchronous circuits may cause problems, especially when (spike-
prone) combinational logic is used to clock or reset storage elements.
LSI Logic strongly recommends that your designs by synchronous.

Do not use asynchronous (ring) oscillators. Process variation renders

their frequency unpredictable. As a result, parts cannot be tested using
standard automatic test equipment. Moreover, ring oscillators
introduce noise, which also reduces testing reliability.

To avoid glitches, use synchronous pulse generators. Figure 4.1

illustrates the incorrect use of an asynchronous pulse generator whose
function is to provide a reset pulse whenever a positive edge occurs
on its input.

 LSI Logic Corporation 1988 - 2000 4-1

An ASIC Primer

A Z
A

Z
Waveforms
(a) from Glitch
Pulse Generator
(b) Width in (a)

Figure 4.1
Classic Prop-Delay Based Pulse (glitch) Generator

Note that the circuit in Figure 4.1 should not be used in LSI Logic
ASIC designs. The pulse width in the above circuit would appear to
be equal to the cumulative propagation delay of the string of
inverters. However, a distributive-capacitance model of the circuit
reveals that the pulse width on the output of the NAND gate is
equal to the difference in propagation delay between the inverter
string and the wire beneath that inverter string. During pre-layout
simulation, this circuit may perform satisfactorily. However, if the
wire should be longer than expected after layout, the resulting
increase in capacitance will cause the delay in that wire to increase
proportionately. Increased voltage, decreased temperature, and
process variation will decrease the delay of the inverter string. If the
wire delay equals that of the inverter string, no output pulse will be
generated; asynchronous pulse generators such as this can cause
glitches. Use a synchronous pulse generator as shown in Figure 4.2
in place of an asynchronous one.

4-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Z
A
D Q D Q
FD1 FD1
QN
CP

Figure 4.2
Synchronous Pulse Generator

In this synchronous pulse generator, the Z output is the decoded

output or the flip-flops. This pulse will always be exactly one clock
cycle long. The decoded output (Z) is used as the new clock signal
(see Figure 4.3).

CP

Figure 4.3
Synchronous Pulse Generator Timing Diagram

 LSI Logic Corporation 1988 - 2000 4-3

An ASIC Primer

CHAPTER 4 - SECTION 2
Avoid “Glitch” Generators
An inconsistency in signal timing produces what is conventionally
called a “glitch”. Figure 4.4 shows how a glitch might look
compared to a signal that is behaving predictably.

Expected Behavior

Glitch Behavior

Figure 4.4
Glitch Behavior

Some circuits, such as the 2-to-1 multiplexer in Figure 4.5,

invariably cause glitches.

D0

G1 Z
SEL
G2

D1

Figure 4.5
Glitch Generating Mux

4-4  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Figure 4.6 is the timing diagram for the circuit in the next example.
To see how the glitch is generated, assume that inputs D0, SEL, and
D1 are initially high. The output therefore is low. When set makes a
high-to-low transition, neither D0 nor D1 is selected for a period of
time. During this time, the output Z will glitch high regardless of
the states of D0 and D1. The glitch width is equal to the time it
takes G1 in Figure 4.5 to propagate a 1 to its output.

Note that the low-to-high transition of SEL in Figure 4.6 will not
cause a glitch on G2’s output.

HIGH

D0
HIGH

D1

SEL

SEL

Z
GLITCH

Figure 4.6
Timing Diagram of Mux Output Glitch

 LSI Logic Corporation 1988 - 2000 4-5

An ASIC Primer

It is safe to use the output of a potential glitch generator to drive the

data input of a synchronous network provided the output has
reached its steady state before the active edge of the clock is
applied. However, it is a bad practice to use such a circuit to drive a
clock line, a set line, or a clear line because the potential glitch may
cause a false clock signal or reinitialize internal registers. Figure 4.7
shows the 2-to-1 multiplexer driving a data input.

D0

SEL D Q

FD1
D1
CP QN
CLK

Figure 4.7
The Glitch Doesn’t Matter Here

CHAPTER 4 - SECTION 3
Use Care with Gated Clocks
Designers must use great care in the use of gated clocks since
improper timing of the gated signal can generate a false clock,
causing the flip-flop to clock in the wrong data. The circuit shown
in Figure 4.8 has a gated clock signal.

4-6  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

DATA
D Q

FD1
GATE

CLK CP

Figure 4.8
A Gated Clock

Again, gated clocks must be used with extreme caution, and require
very restrictive timing tolerances between the clock and the gate
signal. If you are unable to guarantee such tolerances, do not use
gated clocks. Moreover, because of skew between different local
clocks, gated clocks may cause hold-time violations. Interfaces
between local clock signals should be carefully monitored.

The alternative design shown in Figure 4.9 is perhaps a safer

implementation of the same logic function since it eliminates the
gated clock. When GATE is low, the data is allowed to circulate
continuously with each clock pulse. As soon as GATED goes high,
new data enters the flip-flop on the next active clock edge.
(Remember to allow proper setup time.)

Of course, today, one of the main advantages of gated clocks is that

they can reduce power dissipation overall - sometimes
considerably! So we cannot discard this design technique carte
blanche, but it is a more esoteric technique that requires some
design expertise.

 LSI Logic Corporation 1988 - 2000 4-7

An ASIC Primer

GATE

D Q

FD1
DATA
CLK
CP QN

Figure 4.9
Controlling Data Instead of Clock is Safer

CHAPTER 4 - SECTION 4
Avoid Race Conditions
Race conditions occur when the same signal follows two or more
paths having different delays to a common circuit element, or when
two signals caused by the same event (say a clock) follow different
paths to the same circuit element. (Race conditions sometimes
create a condition called “reconvergent fanout.”) Avoid designs
which function efficiently only if there is a predictable interval
between signals traversing different delay paths. This interval is
seldom predictable; pulse widths and delays vary widely due to
variations in layout, processing, power supply, and temperature.
Even where multiple paths have the same number of elements,
variations in fanout and wire delay will cause timing differences in
the different paths. As a result, a circuit that is subject to race
conditions is bound to be unreliable. Examples of race conditions to
be avoided are shown in Figures 4.10 through 4.13.

4-8  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

SD
DATA
D Q

FD4

CLK CP

Figure 4.10
A Race Condition

DATA

CLK

Q Unpredictable

Figure 4.11
Waveforms for Figure 4.10

In Figure 4.10, the Q output sets the flip-flop when a zero is clocked
into the flip-flop, creating the narrow negative pulse on the Q output
as shown in Figure 4.11. The output pulse width depends on the Q
output loading. Under heavy loading, the pulse may set the flip-
flop, which is close to its source, but then be so attenuated by a long
wire that it is invisible to its other fanouts.

 LSI Logic Corporation 1988 - 2000 4-9

An ASIC Primer

Under ideal conditions, the circuit in Figure 4.12 will generate a

narrow, positive pulse on each active clock edge. However, if the
clock line is part of a high-fanout network with a slow ramp rate,
then the rest signal may reach the flip-flop before the active clock
edge, with the result that no pulse will be generated. To see how this
can happen, look at the timing diagram in Figure 4.13. The inverter
threshold voltage may be slightly different from the flip-flop’s
threshold voltage (in this case, it is slightly lower). As a result, the
inverter goes low and resets the flip-flop before the CLK signal
reaches the flip-flop’s clock threshold.

Vdd
D Q

FD2

CLK

CD

IV
R

CLK

Figure 4.12
Another Race Condition

4-10  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Vth(IV) Vth(FF)

CLK

tPD(IV)

Q LOW

Figure 4.13
Waveforms for Figure 4.12

CHAPTER 4 - SECTION 5
Avoid Floating Nodes on Internal Three-State Buses
A floating node occurs when all of the buffers driving a bus are
disabled, as shown in Figure 4.14. In general, floating nodes should
not be used in LSI Logic ASICs. (Note, however, that a node
floating for a period of time equal to the delay of a 2-input NAND
gate with a fanout of four IV gates and two millimeters of wire is
acceptable.) You can avoid floating nodes on internal three-state
buses by adding a buffer to drive the bus when all of the other
buffers are off. (See Figure 4.15.)

 LSI Logic Corporation 1988 - 2000 4-11

An ASIC Primer

S0=1

S1=1

S2=1 “Floating”

Figure 4.14
Avoid Floating Nodes

S0

S1

S2

S0
S1
S2

Figure 4.15
Floating Node Avoided w/ Extra Buffer/Decode

4-12  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

A floating node will cause static current to flow on a driven gate

(such as the inverter in Figure 4.14), since the P- and N-channel
devices of the inverter can be at least partially ON at the same time.
A floating node may also cause the inverter gate to oscillate,
generating excessive noise.

Again, to correct this situation, you can add another buffer to drive
the bus, as shown in Figure 4.15.

The enable signal for the extra buffer in Figure 4.15 is equal to the
NAND function of the three original enable signals. Decoding these
multiple enable signals ensures that at least one buffer drives the
bus at all times. Take care to control the enable signals so as to
prevent bus contention. In some logic schemes, a bus may float for
a brief time, but this should be minimized as much as possible. LSI
Logic does not allow bus contention on our test hardware.

CHAPTER 4 - SECTION 6
Use Bidirectional Output Buffers with Pull-Up or Pull-Down Resistors to
Increase the Noise Margin
Figure 4.16 illustrates the use of bidirectional output buffers with
pull-up or pull-down resistors.

Pull-up and pull-down resistors tend to pull the signal away from
the threshold of the pad in question making noise less likely to
propagate back into a design, and thereby tending to reduce the
possibility of data corruption in storage elements.

 LSI Logic Corporation 1988 - 2000 4-13

An ASIC Primer

Figure 4.16
Bidirects with Pullup and Pulldown

Do not use the input half of a bidirectional buffer to drive a clock

line. Doing so will result in clock problems in your design. In most
LSI Logic ASICs, the typical pull-up or pull-down resistor is in the
range of 80K ohms. It takes tens of microseconds for an 80K pull-
up resistor to pull a pad high. During this time, any noise that
occurs a the pad may propagate to the clock input of internal flip-
flops, clocking in bad data. An on-chip, pull-up resistor should not
be relied upon to produce a correct input state for a floating signal.
That is not the purpose of these resistors. Use off-chip, high-
precision resistors to set bus voltages.

4-14  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

CHAPTER 4 - SECTION 7
Avoid Excessive Fanout
A logic element driving many elements is said to have a high fanout
and drive a heavy load. In an HCMOS circuit there is no theoretical
limit to how heavy a load a gate can drive. If a gate fans out even to
1000 places, the output will eventually reach its final state, albeit
after a very long delay. For practical purposes, the rise or fall or an
internal node ought not to exceed about 1.0 ns under nominal
conditions.

IV
Fanout=FO
VIN
VOUT

FO=50
FO=20

FO=10

VOUT TL TL TL

Time

Figure 4.17
Excessive Power Consumed Due to Excessive Fanout

The graph in Figure 4.17 shows that the ramp rate of a gate
becomes longer as the fanout increases. As the slope becomes
longer, the driven macrocells spend more time (TL) in the active
region, where both N- and P- channel transistors are ON, and

 LSI Logic Corporation 1988 - 2000 4-15

An ASIC Primer

consume excessive power as a result. This situation is undesirable

and should be avoided. Do not exceed recommended maximum
fanouts for library elements.

CHAPTER 4 - SECTION 8
Avoid “Dangerous” Decoding

D Q

Qa Qb Qc Qd

CLK Four-bit Binary Counter

CLK

Qa

Qb

Qc

Qd

Valid Clock Edge False Clock Edge

Figure 4.18
Dangerous Decoding of Terminal Count

4-16  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Use either Johnson Counters or separate flip-flops to decode

terminal counts. Figure 4.18 shows a procedure which, though
convenient, can cause false clock signals. In Figure 4.18, the
terminal count of a binary counter is detected with a NAND gate.
The decoded signal is then used to drive a clock line. To avoid false
clock signals, use a Johnson Counter or a Gray Code Counter
instead of a binary counter; or else use separate flip-flops for the
decoded signal.

In Figure 4.18 the internal loading on each of the Q outputs of the

counter varies, as do their delays. As a result, the inputs to the
NAND gate stabilize at different times. The output of the NAND
gate can switch between a one and a zero state, creating a false
clock signal, until it finally stabilizes to its steady state.

The same logic can be implemented with a Johnson counter rather

than a binary counter. Figure 4.19 shows the logic diagram of a
Johnson counter. Any terminal count can be decoded with a 2-input
NAND gate. Since only one flip-flop changes state during a single
cycle, no glitch is generated.

 LSI Logic Corporation 1988 - 2000 4-17

An ASIC Primer

D Q

FD2

QN

D Q D Q D Q D Q

FD2 FD2 FD2 FD2

QN QN QN QN

CLEAR

CLK

Figure 4.19
Decoding a Johnson Counter

The outputs of a binary counter can be decoded by decoding one

count before the terminal count and running the decoded signal into
a separate flip-flop, as illustrated in Figure 4.20. The new signal
coming from the output of the single flip-flop will have a clean
waveform.

4-18  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

D Q

FD1
Qan Qb Qc Qd
QN
Binary Counter

CLK

Figure 4.20
Alternate Glitch-Free Terminal Count Decode

CHAPTER 4 - SECTION 9
Do Not Use Unbuffered Transmission Gates
Transmission gates have the advantage of higher speeds into light
loads, but they must be used judiciously. Input signals of the
transmission gate are briefly shorted together, and when at least one
of the input signals to a transmission gate multiplexer is the output
of an unbuffered storage element (see Figure 4.21), this may result
in corruption of data in the storage element. All LSI Logic
proprietary flip-flops and latches have buffered inputs and outputs
to prevent such a malfunction. Customer-designed storage devices,
however, may not have this buffering protection. Do not build this
reliability problem into your design. So, if you must use unbuffered
storage elements, do not use transmission gate multiplexers.
Conversely, if you use the transmission gate MUX’s, then use
buffered outputs on storage devices that feed into these MUX’s.

 LSI Logic Corporation 1988 - 2000 4-19

An ASIC Primer

LSI Logic storage devices include this buffer.

Figure 4.21
Faulty Use of Transmission Gate

CHAPTER 4 - SECTION 10
Avoid Bus Contention
Bus contention occurs when a bus is driven from different sources
at the same time. There are two kinds of bus contention - internal
and external. Internal bus contention can occur on internal three-
state buses. External bus contention occurs when an output buffer is
enabled, and is outputting a logical value that differs from that of
the external driving source (see Figure 4.22, Case I).

4-20  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

VDD VDD
OE1 OE2
Q1 Q3
F

Q2 Q4
A
B
Gate 1 Gate 2
VDD

Gate 3

Case 1: Gate 1 driving low; Gate 2 driving high

- DC current path thru Q3 and Q2 to VSS (ground)
- Both transistors handle an “abnormal” current load.
- Voltage at “F” is not well-defined
- Output state of gate 3 is unknown

Case II: Gate 1 driving low; Gate 2 driving low

- Point F at “Logical 0”
- Useless for test purposes

Figure 4.22
Avoid Bus Contention

 LSI Logic Corporation 1988 - 2000 4-21

An ASIC Primer

In Case I of Figure 4.22, with Gate 1 driving low and Gate 2 driving
high, a DC current path exists through the upper transistor of Gate 2
and the lower transistor of Gate 1 to VSS (ground). Both transistors
handle abnormal current loads, and the voltage at point F is not
well-defined. The output state of Gate 3 is questionable. External
bus contention can be avoided by monitoring three-state control
signals to determine when the external source is allowed to drive
the pin.

In Case II, with Gates 1 and 2 both driving low, point F is at a

logical 0, but the situation is useless for test purposes. If one of the
drivers is faulty, the fault will not be detected, and weak output
current drive would be asked by the other driver. Bus contention
can be avoided by three-stating one buffer whenever the bus is
being driven by the other buffer or external signal. Bus contention
can be avoided by allowing a given bus to be driven by only one
driver at any one time. Bus contention will not be allowed during
testing. Therefore, bus contention must be eliminated during
simulation in order to avoid fatal errors when converting simulation
vectors to tester format.

CHAPTER 4 - SECTION 11
Buried Buses: Sometimes Yes, Sometimes No
“Buried” buses that route to many places tend to increase parasitic
capacitance, and result in poor floorplans, inefficient chip
utilization, routing problems for surrounding logic, and faulty die
size estimates. Note that bit-slicing datapaths will help minimize
such buses. Keep buses that route to many places at the top levels of
your design hierarchy so they can be easily seen and planned for.

4-22  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

B C
A D

E 128
E1 E2

F G

A top-level bus that 128-bit wide “buried” bus

routes to many places

Figure 4.23
Two Types of Buses

Sometimes buses can be legitimately buried if they are local buses

that basically connect one or two tightly coupled blocks. Under
these circumstances it is better to bury the bus so as to maintain
close placement of the tightly coupled blocks.

Figure 4.23 shows an example of each of the two types of buses.

 LSI Logic Corporation 1988 - 2000 4-23

An ASIC Primer

CHAPTER 4 - SECTION 12
Follow I/O Guidelines Correctly
A generic LSI Logic rule is that every VSS pad can support only a
certain number of standard output buffers (4 mA equivalents),
depending on what package you are in. But this rule should not be
misinterpreted. The unacceptable design in Figure 4.24 below
shows 2xN 4 mA equivalents between the two VSS pads. In the
G12 technology, it may not be unrealistic to have one VDD and VSS
pad pair for every six to eight 4mA buffers.

VSS

VSS

Figure 4.24
Too Many Outputs Between Power/Ground Pads

4-24  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Figure 4.25 shows the proper rule interpretation - N 4 mA

equivalents per VSS pad (either per side), and no more than N 4 mA
equivalents between VSS pads.

N/2

VSS

N/2

N/2 VSS N/2

Figure 4.25
Enough Ground for the Outputs

 LSI Logic Corporation 1988 - 2000 4-25

An ASIC Primer

Do not place clock signal pins, reset pins, preset pins, or other
major control signals between high-drive output buffers and VSS
(see Figure 4.26).

4 mA equivalents

VSS

Clock
Figure 4.26
Poor Placement of a Clock

4-26  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

In Figure 4.27 the clock pin is properly placed, with a VSS pin
between the clock pin and the high-drive output buffers, or with the
clock pin between two VSS pads. Note that inputs do not cause
nearly as much noise as outputs because they do not drive huge
loads, so having other inputs near the clock is not a problem.

4 mA equiv. VSS VSS 4mA equiv.

Clock

Eight 4 mA equiv. VSS Inputs

Clock

Figure 4.27
Proper Placement of a Clock

CHAPTER 4 - SECTION 13
Power Considerations
In some ASIC designs, minimizing power consumption is a high
priority. In others, it is not. In any case, it is always necessary to
know in advance how much power a chip will consume, since this
affects the operating temperature, which in turn affects:

 LSI Logic Corporation 1988 - 2000 4-27

An ASIC Primer

❑ Speed
❑ Package selection
❑ FIT rate
Speed

CMOS silicon chips slow down as they get hotter. Proper delay
prediction must therefore take junction temperature into account.
And, make no mistake: any temperature specified to any delay
prediction tool is a junction temperature, not an ambient
temperature. Junction and ambient temperature are related by the
equation Tj = (PD)(θ) + Ta, where PD is power in Watts and θ
(Theta) is the thermal resistance of the package in degrees C per
Watt. Since speed is a function of Tj rather than Ta, and the
program does not know Power or Theta, the temperature cannot be
ambient.

Package Selection

From the equation in the previous paragraph it is obvious that, to

reduce junction temperature, you must reduce Ta, PD, or θ. Perhaps
the simplest of these, from a design perspective, is to select a
package with a lower thermal resistance, so that heat generated by
the chip flows more easily to the ambient environment. Usually, the
disadvantage of a more efficient package is higher cost.

FIT rate

Many designers operate under the mistaken assumption that it is

necessary only to cool a chip sufficiently for it to run at the desired
speed. Unfortunately, this ignores the fact that hotter chips burn out
sooner. FIT stands for Failures In Time, and is a measure of how
many chips will fail for each ten thousand hours of operation. In

4-28  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

general, a ten degree C increase in junction temperature will double

the FIT rate. For a low volume design with a finite life expectancy,
this may not be a problem. But for any design requiring high
reliability, or any high volume design requiring moderate reliability,
it may be necessary to reduce the operating junction temperature
well below the rated maximum. A designer who fails to realize this
will end up with a design having very high cost of maintenance and
lots of angry customers. A designer who realizes this too late will
end up with a more expensive system (either because of increased
packaging cost [lower θ] or increased system cooling requirements
[lower Ta]) and, a design that is over-specified, since it will be
capable of higher frequency due to the lower operating temperature.
It is therefore crucial for successful design that the appropriate FIT
rate be specified early in the design cycle, and the desired operating
temperature be calculated with the FIT rate in mind.

Contributing factors

CMOS circuitry consumes power in two ways: by capacitance and

by resistance. The former is very frequency dependent; the latter is
less so. When a CMOS circuit element changes state, it is charging
and discharging capacitors. The more often the capacitors are
charged and discharged (that is, the higher the operating
frequency), the higher the power consumption.

There are simple equations for estimating power based on

frequency, but applying them to a large system will result in a very
high estimate. This is because a typical digital circuit element does
not change state on every cycle. In any power estimation program,
there are factors for both operating frequency and a rather nebulous
“activity factor”, that is, the average number of circuit elements
switching state in any given clock cycle.

 LSI Logic Corporation 1988 - 2000 4-29

An ASIC Primer

VDD VDD VDD

OFF 1/2 OFF/ON ON

‘0’ Switching ‘1’

‘1’ ‘0’
ON OFF
1/2 OFF/ON
Switching C

VSS VSS VSS

No Current No Current
Crowbar fcv2
Figure 4.28
CMOS Power Consumption

There are two types of resistive power consumption in CMOS. The

first is the so-called “crowbar” current, named after the habit of old-
time power generation workers of deliberately causing a short-
circuit with a crowbar to trip the breaker when de-powering a
transmission line, to avoid dangerous power arcs in unpredictable
places. When a CMOS circuit element, say an inverter, changes
state, it is momentarily in the linear region, and both N-channel and
P-channel devices are partially on. This creates a resistive path
directly from power to ground through both transistors. The longer
the ramptime on the input, the more time the device spends in the
linear region. This means that one way of reducing power
consumption is to ensure that there are no excessive ramptimes.

4-30  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

There are many reasons why all excessive ramptimes should be

fixed. A long ramptime not only causes the circuit elements it
drives to consume more power, it may also cause electromigration
which can lead to premature and catastrophic circuit failure. In
addition, a long ramptime will have a large effect on delay
prediction. A long input ramptime will cause a long output
ramptime on any given circuit element, so fixing one long ramptime
will likely fix others as well. Also, a long ramptime may well fall
outside of the characterization range of either the driving element or
the driven element, which means that the delay predictor is
basically making a guess with respect to these delay values.

The second type of resistive power consumption in CMOS is static

power from analog (that is, non-digital) circuit elements. There may
be a large number of analog circuit elements in a modern digital
CMOS chip, from PLLs (phase- locked loops), to high-speed, low
voltage-swing I/Os, to sense-amplifiers in high-density memories.
All of these elements consume static current independent of
operating frequency.

Conserving power

As mentioned earlier, reducing power consumption will make your

chip run faster and last longer, may reduce the required size and
complexity of both your power supply and your cooling system,
and perhaps enable you to use a less expensive package. And, of
course, reducing power consumption is absolutely critical in battery
powered applications.

One of the easiest ways to reduce power consumption is to gate

clocks; unfortunately, this can very easily cause testability
problems, especially when it comes to scan and the associated
ATPG (Automatic Test Pattern Generation). There are now
programs available which will automatically insert circuitry that not

 LSI Logic Corporation 1988 - 2000 4-31

An ASIC Primer

only gates clocks, but also ensures that scan testability issues are
addressed.

Another way of reducing power consumption is to enable power-

hungry elements such as analog circuits and high-density RAMS
only when necessary.

Power distribution
When asked to name the most important interconnection on a chip,
most logic designers will immediately answer, “clock.” Actually,
the answer is power and ground. Logic designers tend to take power
and ground for granted; it doesn't even show up on most
schematics. However, when routing an ASIC, power and ground
come first.

Current LSI Logic technologies make use of a four-bus system in

which power and ground for large geometry I/O transistors (Vdd
and Vss) are separate from power and ground for the core (Vdd2
and Vss2). These run around the periphery of the die, with Vss and
Vdd on the outside and Vdd2 and Vss2 on the inside.

The goal of routing power and ground is to ensure their even

distribution to all points on the die. Unfortunately, today’s
technologies don't have enough routing resources to create
dedicated power and ground planes on-chip, so a certain amount of
“droop” must be anticipated. Since power and ground buses are
located at the periphery of the die, the greatest power droop will be
concentrated in the center of the die. This is one reason why
memories, which cannot tolerate as much power droop as other
logic elements, should be placed near the edge of the die.

Another result of not having dedicated power planes is that power

flowing into and out of the core region must all flow to the allotted
Vdd2 and Vss2 pins. If care is not taken to provide a sufficient

4-32  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

number and proper location of these pins, a great many electrons

can all funnel through an insufficient cross-section of metal, which
causes metal migration over time. Eventually, these metal traces
will blow like a fuse. This is another example of electromigration.

Power isolation

Even when CMOS ASICs were 100% digital, it was sometimes

necessary to prevent noise from high current switching areas of the
die from interfering with more sensitive portions of the circuit.
Until the G12 technology, all LSI Logic chips had clock drivers and
output buffers sharing the same power and ground. Since this could
introduce unacceptable noise into the clock line, it was often
necessary to take drastic measures to isolate power and ground for
the clock from power and ground for the output buffers. These
drastic measures involve cutting the power and ground rings on
either side of the selected circuitry and providing the isolated
segment with its own power and ground pins. This becomes even
more important when what we need to isolate is not merely a clock
driver, but an analog circuit such as a PLL or A/D converter.

A power ring segment isolated by power cuts can also be supplied

with a voltage level different from that on the rest of the die. The
affected area might be supplied with a higher voltage level to
increase its performance, or with a lower voltage level to conserve
power, or simply to make an interface which is compatible with a
more exotic system.

Power and Testability

All LSI Logic ASICs are tested for leakage current, both for
reliability reasons (a chip with elevated leakage current will fail
after fewer hours of operation compared to a chip with normal
leakage), and as a means of defect detection. A typical leakage

 LSI Logic Corporation 1988 - 2000 4-33

An ASIC Primer

current measurement is on the order of 100µA. A typical high

density memory can draw almost 100mA, a difference of three
orders of magnitude. To ensure that structures which can draw
static current can be switched off, all designs in current LSI Logic
technologies must include a test pin connected to a special buffer
called IIDDTN. This cell’s primary purpose is to disable pullup and
pulldown “resistors” (actually transistors), but it can and should
also be used to provide a means of putting all circuit elements
capable of drawing static current into their power-down modes.

4-34  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

5
Designing for Testability

CHAPTER 5 - SECTION 1
Testing Redundant Logic
Redundant logic, such as that shown in Figure 5.1, may be used to
enhance reliability. However, unless the redundant logic can be tested
uniquely, the intended enhancement may prove illusory. Figure 5.2
shows a testable version of this circuit.

Logic A

IN OUT

Logic B

Figure 5.1
Untestable Redundant Logic

 LSI Logic Corporation 1988 - 2000 5-1

An ASIC Primer

TEST A
Logic A

IN OUT

TEST B Logic B

Figure 5.2
Testable Redundant Logic

CHAPTER 5 - SECTION 2
Use Combinational Logic
Combinational logic is easier to test than sequential logic. in
combinational logic, the outputs are always a function of the
current inputs. In sequential logic, the outputs depend on both the
current inputs and the outputs of storage elements driven by inputs
preceding the current ones. For example, in the multiplexer in
Figure 5.3, all possible states can be exercised by presenting
permutations of 1’s and 0’s on the input pins. The test vectors will
not exceed 29 power (512), and will cover all possible detectable
stuck-at faults.

5-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

AO2
D0
D1

AO2
D2
D3
AO2
Z
AO2
D4
D5

A
B
C

Figure 5.3
Combinational Logic is Easy to Test

CHAPTER 5 - SECTION 3
Add Test Lines to Sequential Logic
Sequential logic is more difficult to test than combinational logic
because it contains storage elements. You can make sequential logic
more testable by adding test lines to logically break feedback loops
between flip-flops.

Figure 5.4 shows a segment of untestable sequential logic. The

circuit’s initial output condition is unknown because of the
feedback loop from the outputs to the inputs of the flip-flops.
Therefore, input test vectors may not result in a unique set of output
vectors. The four possible output sequences for the logic shown in
Figure 5.4 are shown in the table below the diagram.

 LSI Logic Corporation 1988 - 2000 5-3

An ASIC Primer

State diagram for A & B

00

10 01
A B
D Q D Q
11
FD1 FD1

QN QN
CLOCK
DATA

OUTPUT

DATA 0 0 1 1 1 1 1 1 1 1 1 1 1 cont. 1’s

OUTPUT when
initial A, B=00 1 1 0 1 0 0 1 0 0 1 0 0 1 repeat 001

OUTPUT when
initial A, B=01 0 1 0 1 0 0 1 0 0 1 0 0 1 repeat 001

OUTPUT when
initial A, B=10 1 0 1 0 0 1 0 0 1 0 0 1 0 repeat 010

OUTPUT when
initial A, B=11 0 0 0 0 1 0 0 1 0 0 1 0 0 repeat 100

Figure 5.4
Untestable Sequential Logic

5-4  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Figure 5.5 shows how adding a test line makes the same circuit
testable. The test line logically breaks the feedback loop. Holding
the test line low can produce a known output after a maximum of
two clock cycles. Then the test line must be brought high. An input
test vector then produces a unique output vector. The test line need
not be a separate external pin on the package; you can derive it from
an unused state of some other combinational logic on the chip. (If
this approach is used, however, take care that your test signal is
stable before you begin your test.) Note that if, in using this
approach, your logic should ever pass through the unused state, the
chip may be put into the test state. Design your chip in such a way
that it cannot be put into test state by accident.

ND3 D Q D Q
TEST
FD1 FD1

QN QN
CLOCK IV ND2
DATA
ND2
OUTPUT
ND2

DATA 0 0 1 1 1 1 1 1 1 1 1 1 1 continue 1’s

TEST 0 0 1 1 1 1 1 1 1 1 1 1 1 continue 1’s

OUTPUT x x 0 1 0 0 1 0 0 1 0 0 1 repeat 001

Figure 5.5
Testability Added by “Test” Input (Breaking the Loop)

 LSI Logic Corporation 1988 - 2000 5-5

An ASIC Primer

CHAPTER 5 - SECTION 4
Use Reset Signals to Initialize Storage Devices
Adding reset signals is one of the simplest and most effective ways
to improve the testability of circuits that contain storage elements.
The penalty is the additional wiring the buffering needed to drive
the fanout. Storage elements require a testing strategy that puts
them into known states. You can initialize storage elements by
clocking in known input data, but by adding a reset signal you can
reduce the number of initialization test vectors.

Since many subsystems (like the one shown in Figure 5.6) use reset
signals in their normal operation, a reset signal usually involves
little or no additional logic. The reset signal can be used to reset all
elements whose state cannot be readily initialized by other means.
For example, the divide-by-three circuit from Figure 5.4 (clock/3)
at the bottom of Figure 5.6 uses the reset signal from the shift
register above it.

5-6  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

DATA OUTPUT

Logic D Q D Q D Q D Q Logic

FD1 FD1 FD1 FD1

RESET
CLOCK

D Q D Q

FD1 FD1

Divide by three

Figure 5.6
Use Reset Lines to Initialize Storage Elements

A master reset is not always required. The technique used in Figure

5.5 illustrates how the use of a 3-input NAND gate enables you to
add a test line. This test line provides an alternative method of
initializing the flip-flops, so that a master reset signal need not be
used.

 LSI Logic Corporation 1988 - 2000 5-7

An ASIC Primer

CHAPTER 5 - SECTION 5
Make Logic Observable by Multiplexing Output Pins
In Figure 5.7a, the state of the three FD1’s can be observed by
placing a 3-to-1 multiplexer in front of the output buffer, as shown
in Figure 5.7b.

11 11
D Q D Q

FD1 FD1

12 12
3-to-1
D Q D Q MUX
FD1 FD1

13 13
D Q D Q

FD1 FD1
S0 S1
(a) (b)

Figure 5.7
Multiplex Output Pins

CHAPTER 5 - SECTION 6
Decode Input Pins
You can make deeply buried states (e.g., internal registers)
controllable for testing by decoding input pins, as illustrated in
Figure 5.8. You can monitor internal registers by connecting them
to package-pin test points. To increase the number of effective

5-8  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

package pins for the purpose of monitoring internal registers

(improving observability), make unidirectional pins bidirectional. If
pins cannot be made available in this way, enable on-chip test logic
with a combination of inputs that cannot occur in actual system
operations. Never, however, put a clock signal on a bidirectional
pin.

s0 s0 11
D Q D Q
11
IBUF FD1 S0/D3 FD1

s1 s1 12
D Q D Q
12
IBUF FD1 S1/D4 FD1

clk
IBUF
Internal 2:1 13
D Q D Q
Internal 13 signal MUX
Signal FD1 FD1

Internal 14
D Q 2:1 D Q
Internal 14 signal MUX
Signal FD1 FD1

TEST

CLOCK

Figure 5.8
Decode Input Pins

 LSI Logic Corporation 1988 - 2000 5-9

An ASIC Primer

CHAPTER 5 - SECTION 7
Break Up Long Sequential Chains
Long counters, although they may operate satisfactorily in circuit
systems, can add significantly to the number of test vectors needed
for proper testing. For example, the 16-state counter shown in
Figure 5.9 would require 216 power (65,536) test vectors merely to
apply a single clock to the control circuitry. Assuming the control
circuitry has 212 possible states for thorough testing, about 2.7 x
108 test vectors would be needed to test the circuit completely.

System inputs

100MHz
input Low
clock speed
16-Stage controller
prescaler -212
states

System outputs

Figure 5.9
Long Counter is Difficult to Test

A practical solution is to break the 16-stage counter into two 8-

stage counters, as shown in Figure 5.10. You may then test two
portions of the counter and the control circuitry in parallel; 4096
test patterns are sufficient for this task.

5-10  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Test point Test point

A02 A02
Input 8-Stage 8-Stage
clock prescaler prescaler

System
inputs

Normal operation Low

speed
Test controller
mode -212
states

System
outputs

Figure 5.10
Two Short Counters are Easier to Test

CHAPTER 5 - SECTION 8
Use Level-Sensitive Design
In large designs, unavoidable skew along a clock signal may make
edge-triggered flip-flops unreliable due to hold-time violations. One
way to avoid this problem is to use a level-sensitive design
technique. In Figure 5.11, ENA and ENB are non-overlapping,
active-high enable signals. Signals are transferred into one level of
the design using the ENA signal and into alternative levels using the
ENB signal. In this way, the flow of data through the design can be
staged and very well controlled.

 LSI Logic Corporation 1988 - 2000 5-11

An ASIC Primer

Combinational logic

LD1 LD1 LD1

D Q D Q D Q

G G G

(Level 1) (Level 2) (Level 3)

ENA
ENB

Figure 5.11
Level Sensitive Design

It is even possible to gate enable lines with complete safety

(although with some loss of speed) as long as all gating signals
originate with latches using the opposite enable phase (see Figure
5.12).

LD1
LD1
D Q
D Q AND2
A
Z G
ENA G B

ENB

Figure 5.12
Safely Gated Enable Line

5-12  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

CHAPTER 5 - SECTION 9
Partition Logic into Blocks
A large design may contain deeply buried functional blocks which
are both difficult to control and to observe. In such a case, it may be
desirable to divide the design into easily testable blocks. These
blocks may be made functionally transparent while the design is in
normal operational mode. With only a slight increase in delays
between blocks. Figure 5.13 and 5.14 show how this can be done.

From Prvious latch

LS2
D1 Q1

C1

D2

C2

Block C3 Q2 Block
A B

LS2
D1 Q1

C1

D2

C2
C1
C2 C3 Q2
C3
To next latch

Figure 5.13
Block Level Boundary Scan

 LSI Logic Corporation 1988 - 2000 5-13

An ASIC Primer

During normal operation, C1 is always high, and C2 is always low.

Thus the latches are merely transparent elements. In test mode, if
Block A is to be observed, C1 is held until the proper time and then
brought low. Then the data is scanned out using C2 and C3 as non-
overlapping enable lines. To control Block B, C1 is left low and
data is scanned in as before.

LSI
Block A D1 Q Block B

D2
MUX21H
A
Z
Normal B
function Normal
S
output
Normal Normal
input function
TEST

Figure 5.14
Block Partitioning Using Muxes

CHAPTER 5 - SECTION 10
Do Not Design Your Own Storage Devices
Timing Considerations

In contrast to LSI Logic defined storage devices, customer-designed

storage devices do not have known layouts. Timing analysis of
customer-designed storage elements is therefore unreliable. If you
need a storage device not available in an LSI Logic library, ask that

5-14  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

it be coded by LSI Logic as a hard-coded macrocell. The layouts of

hard-coded macrocells are known and can therefore be accurately
modelled and simulated by LSI Logic simulation software.

Figure 5.14 depicts a customer-designed storage device that has

been flattened for analysis. LSI Logic layout software treats this
combination of gates as a random collection of gates - i.e., not as a
functional block (in this case, a D flip-flop). In other words, our
software does not identify (or recognize) the gates comprising a
user-defined storage device as: data inputs, clock inputs, reset, etc.
It does not, therefore, analyze signal-timing, relationships among
the signals in a user-defined storage device. The concept of set-up
time, for example, which is essential to the analysis of a storage
device, is irrelevant to gates not identified as those of a storage
device. (Note that Figure 5.15 is the logic schematic of a D flip-
flop.) See if you can identify the pins. The answers are at the end of
this chapter.

(2)
C
(1)
E

B
F

A
D
(3)
(4)

Figure 5.15
Customer-Designed Storage Device

 LSI Logic Corporation 1988 - 2000 5-15

An ASIC Primer

Logic Considerations

Very often the inputs to a transmission-gate multiplexer are briefly

shorted to one another. As a result, a logical 0 on one input could
change the state of latched data going into another input. LSI Logic
solves this problem by buffering the outputs of its hard-coded
storage devices in order to make feedback paths inaccessible at the
output pins of the storage element. Figure 5.16 depicts a potentially
dangerous use of a transmission-gate multiplexer, on of whose
inputs is an unbuffered storage device. Note that the signal on B can
follow the dotted line path into the storage device, toggling the state
of the latch.

U1

U2
A

Combo data B

Figure 5.16
Dangerous Use of a Transmission Gate Mux

There are other compelling reasons why designers should abstain

from creating their own storage devices, such as non-identical
layout of logically identical storage cells, differing AC performance

5-16  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

of logically identical storage cells, and so on.

CHAPTER 5 - SECTION 11
Use a Serial Scan Approach
Standard scan design uses a one-phase clock. Place a 2-to-1
multiplexer in front of each register, as shown in Figure 5.17.

In this case, each flip-flop receives its data input either from its
normal path or from the Q output of the previous flip-flop in the
scan chain, depending on the state of the test pin. The top flip-flop
is an exception; its input comes from a test input pin. The bottom
flip-flop’s Q output goes both to its normal destination and also to a
test output pin.

The circuit is placed in test mode by setting TEST to HIGH. In this

mode, all the flip-flops are connected or “chained” together as a
single serial shift register. This allows a serial data stream to be pre-
loaded into the storage elements. When TEST is set to LOW, the
circuit reverts to its normal operating mode and is clocked for one
or more cycles to exercise the combinational logic. As a final step,
the circuit should be placed in test mode again to shift out the
contents of the scan chain for examination. Because scan chains can
be very long for large circuits, LSI Logic uses a simulation
technique that allows scan chains to be parallel loaded and
unloaded during simulation. LSI Logic’s lsitest and lsimts programs
later convert these parallel patterns into serial patterns for actual use
on the testers.

Entire PC boards may be put into test mode if properly designed.

All of the individual circuits can then be read and analyzed. Serial
scan design enables you to divide a large circuit into two or more
independent sub-circuits for testing. You can shift out and examine
the state of every storage element in your design.

 LSI Logic Corporation 1988 - 2000 5-17

An ASIC Primer

TESTIN IBUF

IBUF

MUX
D Q
TEST
FD1

MUX
D Q

Combinational FD1
Inputs IBUF logic

B4 TSTOUT
MUX

D Q

FD1

CLOCK IBUF

Figure 5.17
Scan Design

5-18  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

CHAPTER 5 - SECTION 12
Multiplex Output Pins to Make Scan Chain Observable
Pin requirements need not be increased for circuits using scan
design since scan outputs can be multiplexed with regular output
signals, as in Figure 5.18. Do not, however, multiplex critical path
outputs.

IBUF Shift
data in FD1S

MUX
D Q
FD1
CP
MUX

Large D Q
IBUF combo
FD1
logic
CP

IBUF
MUX

Logic D Q
MUX

FD1

TEST CP
IBUF

IBUF B4

Figure 5.18
Sharing the Scan Output

 LSI Logic Corporation 1988 - 2000 5-19

An ASIC Primer

Note that several scan flip-flops are available in LSI Logic libraries,
(The circuitry enclosed within a dotted line in Figure 5.18 and
identified as an FD1S is such a flip-flop.) Each scan flip-flop has a
test input (TI) and test enable (TE) input in addition to its normal
inputs. Scan test flip-flops have an “S” suffix on their macrocell
names.

CHAPTER 5 - SECTION 13
Implementing the Vil/Vih (NAND) Tree

What is the NAND Tree?

The LSI Logic NAND tree is a dedicated circuit for testing each
input buffer's Vil/Vih. It is formed from a special NAND gate
embedded within each input or bidirect buffer. One input of the
NAND gate is connected to the output of the buffer, and the other is
given the port name PI. The output of the NAND gate is referred to
as PO. This means that, for a non- inverting input buffer, the PI pin
is active when the input signal is at 1 and inactive when the input
signal is at 0. One unidirectional output buffer is chosen as the
NAND tree output. This buffer is driven, through a MUX if desired
soas to avoid having a dedicated NAND tree output pin, by the
nearest PO pin. The corresponding PI pin is then connected another
PO pin, and the process is repeated until all PO pins are connected.
This leaves one unconnected PI pin, which is tied to logic 1. The
result is a daisy-chain of NAND gates, all leading to a single output
buffer. There is now a purely combinational path from each input
buffer to a single output buffer.

Why is it important to have a combinational path?

The Vil/Vih test is performed with each input at its own Vil-max or
Vih-min. This leaves no noise margin at all. Any storage element
within the circuit would be susceptible to the first noise spike that

5-20  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

came along.

Normal Output1
Normal Circuit

Input Buffer

Input Buffer

Input Buffer

Procmon
Input Buffer TESTOUT output pin

1. All outputs, except TESTOUT, are masked during the DC parametric tests.

Figure 5.19
NAND Tree Structure

How is the NAND tree sequenced?

To save routing resources, it makes sense to connect the NAND tree

in a sequence dictated by the bonding diagram, so that the wires
between NAND gates will be as short as possible and will be
confined to the edges of the die.

 LSI Logic Corporation 1988 - 2000 5-21

An ASIC Primer

A C

E
D
VDD

B TO

Figure 5.20
NAND Tree Sequence

How is the NAND tree tested?

In its simplest form, the Vil/Vih test is performed by setting all

inputs to logic 1, then setting each to 0 in order of connectivity,
starting with the buffer furthest from the output (that is, the one
with its PI pin tied to 1). If there is an even number of buffers, the
output is now at 1; if there is an odd number, the output is at 0. As
each input goes from one to zero, the output toggles. The overall
result is a function of each input properly responding to both its
own Vil-max and Vih-min. Notice also that the Vil/Vih test is a
static test. It is only the voltage levels that are important, not the
transitions themselves, and since the circuit itself is combinational,
the particular sequence of the patterns is also irrelevant. We could
re- arrange the sequence so that all the patterns that cause 1 on the
output came first, and all those that cause a 0 came second, and we

5-22  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

would still have a valid Vil/Vih test.

The test must also take into account special cases such as Schmitt
triggers (where direction and sequence are important; they have to
be toggled both ways), and other special buffer types such as PCI.

Figure 5.21
NAND Tree Test

Subtleties

The NAND tree has some surprising subtleties for something that
seems at first glance to be so simple. These subtleties arise not from
the NAND tree itself, but from the input buffers which drive it. The
two most obvious problems are the bidirects and the MUX. How
can we have a purely combinational circuit whose output is a

 LSI Logic Corporation 1988 - 2000 5-23

An ASIC Primer

function of all the inputs, and still guarantee that the bidirects will
not suddenly go from input mode to output mode? After all, if we're
testing the voltage response of the input half of the bidirect, we
can't very well have it slip into output mode. But one of the other
inputs that we need to toggle would have to be holding the bidirects
in input mode, and if we toggle that input to test it, we risk
switching the bidirects to output mode. What must happen is that
the global three-state enable pin must appear in the NAND tree
closer to the output than any bidirect buffer, and it's sense must be
such that the bidirects are forced to input mode when the PI pin is
active.

The other obvious problem is the MUX. What determines whether

we're in test mode or normal mode, and how do we test that signal
itself? Since most designs have an asynchronous reset pin (which
is, by most conventions, active low), and most designers really don't
care what the chip does while it is in the reset state, the reset state
actually makes a very good Vil/Vih test mode. Just tie the reset
signal to the select pin of the MUX. Ah, but wait! One of the LSI
Logic design rules states that no PO pin can be unconnected. The
simple solution is to drive the select of the MUX not with the Z
output of the reset buffer, but with the PO output of the reset buffer.
The PI pin is tied high. Testing the reset buffer for Vil/Vih is not a
problem: simply toggle the reset signal when the two inputs to the
MUX are in opposite states. One input of the mux (from the
PROCMON “A” pin) comes from the NAND tree, the other (“N”)
from the “normal” (non-test mode) circuitry. If the normal circuitry
is not initialized, you'll be switching between the NAND tree and a
X (unknown) state, which is not a valid test. This is yet another
reason to use reset as the Vil/Vih test mode; presumably, reset will
initialize the N pin.

5-24  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

The PROCess MONitor

OK, so now we have a daisy-chain of NAND gates driven by each

input and bidirect, with a MUX at the output and all the bidirects
close to the far end. What else? It turns out that, since a string of
NAND gates this long has so much propagation delay, it actually
doesn't hurt anything to make it even longer, which is why another
test structure, the PROCess MONitor or PROCMON, is added to
the end of the NAND tree. The PROCMON is a small hard-coded
macro with four inputs (A, E, S, and N) and one output (Z). It
consists of a special asymmetric delay chain, an exclusive NOR
gate to configure the delay chain as a pulse generator, and some
select circuitry to enable the E pin to select between pulse mode
and inverted pass-thru mode. It even has the 2:1 MUX as its output
stage (where the PROCMON's S and N pins are the S and A pins of
the MUX, respectively). By using the global three-state enable
signal as the mode (or “E”) pin of the PROCMON, the bidirect
enable problem is handled. During the static Vil/Vih test, pulse
mode wouldn't do us much good, so we must be in inverted pass-
thru mode. The preferred global three-state input buffer is the
ICPTNU cell. The polarity of the signals on the PROCMON is
deliberately designed so that connecting the PO pin of the ICPTNU
to the E pin of the PROCMON will give valid results.

There is one other important restriction in connecting the NAND

tree. If proper connection of the PROCMON and the ICPTNU
buffer takes care of bidirect conflicts, do we still need to worry
about where the bidirects fall within the NAND tree? Well, actually,
yes. Testing the PROCMON involves toggling both the ICPTNU
buffer and the input buffer closest to the PROCMON. If either of
the two input buffers closest to the PROCMON are bidirects, an
unknown logic value may propagate to the output when the
ICPTNU buffer allows the bidirects to become outputs. Therefore,
the two buffers closest to the PROCMON must be unidirectional

 LSI Logic Corporation 1988 - 2000 5-25

An ASIC Primer

input buffers.

ICPTNU

D Q
QN
CD
INPUT_B PROC_DRV
N E
INPUT_A Z
A
PROCMON
S
RESET

Figure 5.22
A Complete NAND Tree Circuit

The PROC_DRV

A brief soon-to-be historical note here: In the G10 and G11

technologies, there is a special PROCMON output buffer called a
PROC_DRV. It has an EN pin which changes its drive strength from
4mA to 2mA. This EN pin must be (manually) connected to the
same source as the PROCMON S pin so that during the Vil/Vih test,
the 2mA mode is active and the noise due to the switching of the
output buffer is reduced. This was important at one time because of
TTL compatible input buffers. TTL buffers had a very low Vih-min,
and were very susceptible to noise during the Vil/Vih test. Now,

5-26  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

remember that the NAND tree is connected in sequence around the

die. Realize also that this may mean that the input buffer at the far
end of the NAND tree may be very close physically to the NAND
tree output. If this input or bidirect were TTL compatible, an
unfortunate thing would happen. During the first pattern of the Vil/
Vih test, the TTL compatible input buffer at the far end of the
NAND tree and physically close to the output buffer goes from 1 to
0. This would send a transition all the way around the periphery of
the die and cause the output to switch, which would cause a noise
spike on the ground bus right next to the noise-sensitive TTL input
buffer now sitting at logic 0. The noise spike causes the TTL input
buffer to glitch, sending a transition all the way around the
periphery of the die.... You get the idea! Since the voltage level in
the G12 technology is low enough so that TTL compatibility is no
longer possible, the rationale for the PROC_DRV is gone as well.

The IIDDTN

Today’s technologies also have another dedicated test input buffer,

called the IIDDTN. The purpose of this buffer is to turn off all of
the pullup and pulldown “resistors” (actually transistors) on input
buffers, and power down any other applicable structures in the I/O
region. It also has another (actually identical) output which the user
can connect to any other circuit elements (e.g. high density RAMs,
PLLs, cores) which need to be powered down for the Idd and Iddq
tests. What does this have to do with the Vil/Vih test? Well, it turns
out that some input buffers have their input thresholds altered when
the IIDDTN buffer is asserted. This has unfortunate repercussions
if it happens in the middle of the Vil/Vih test, so if you allow the
IIDDTN pin to be included at some arbitrary point in the NAND
tree, and if something “downstream” has its threshold altered, the
Vil/Vih test becomes much more complex, to say the least. Add to
this the fact that this input buffer is purely for test and nobody really
cares whether its input thresholds are a little off, and it becomes

 LSI Logic Corporation 1988 - 2000 5-27

An ASIC Primer

apparent that maybe the IIDDTN shouldn't even have PI and PO

pins, and a declaration has been made that it is methodologically
permissible to leave it completely out of the NAND tree.
Unfortunately, doing so will cause errors in several FlexStream
Programs. You can either leave the IIDDTN buffer out of the
NAND tree and ignore these errors, make sure (in the absence of
any clear documentation) that none of your input buffers are
affected by the IIDDTN buffer, or carefully position the IIDDTN
pin within the NAND tree such that it is “out of the way” so to
speak. The best place to put it is as the second-closest buffer to the
PROCMON. Then you need only ensure that the input buffer
closest to the PROCMON is sufficiently simple (e.g. an IBUF) soas
to be unaffected.

The PROCMON2A

Yet another wrinkle in the NAND tree saga is the fact that the G12
technology has two different kinds of transistors, High Performance
(HP) and Low Leakage (LL). Thus the need for two different kinds
of PROCMONs and, for designs that contain both kinds of
transistors, a third, called PROCMON2A, that contains both of the
other two, along with an extra input (SEL) to choose between them.
In non-pulse mode, the path from the A pin through the low leakage
transistors is non-inverting, to differentiate it from the high
performance path and ensure that the input buffer driving the SEL
pin can cause a transition at the output and thus be tested for Vil/
Vih. Since the state of the SEL pin is only important during test
mode (presumably reset, as suggested earlier), the choice of which
input to hook to the SEL pin is not important but, as before, it
should be the PO pin from the input buffer, with the PI pin tied to
logic one.

5-28  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

lsints

There is a FlexStream program which will automate almost all of

the busy-work associated with hooking up the NAND tree and
PROCMON. It is called lsints, for NAND Tree Synthesis. The only
input data required for this program, other than a few rather obvious
command line options, is the UCF (User Control File) file. (OK, it's
a stupid set of initials, and it's kind of like saying “ATM machine”
but what the heck, it's easy to remember.) The most important
information in the UCF file are the connections to each pin of the
PROCMON cell. These take the form of “procmon_xpin” where
“x” can be a, e, n, s, z, or sel. In most cases, the program wants the
name of the external signal, not the name of an internal net or cell.
Further, the program assumes that specifying, for example,
“procmon_spin reset” means that you want the PO pin of the
“reset” input buffer to drive the “S” pin of the PROCMON. By
default, the program creates a modified netlist file, Vil/Vih test
pattern simulation files, and an interface file for the lsitest program.
Future plans include the creation of a WGL file so that you won't
have to bother running the simulation.

Since there are always people who want to make extra work for
themselves, not to mention the occasional redesign, the program
has a “pattern_only” mode which will create test patterns but not a
netlist, which of course assumes that the NAND tree already exists
in the circuit. There are some discrepancies between the UCF file
for “full” mode and the UCF file for “pattern_only” mode. In “full”
mode, the arguments for procmon_npin and procmon_zpin are
identical. In “pattern_only” mode, the argument for procmon_npin
must be an internal net name. Also, in “full” mode the
“procmon_apin” command is ignored, but in “pattern_only” mode
it is required.

 LSI Logic Corporation 1988 - 2000 5-29

An ASIC Primer

Answers to Figure 5.15:

(2) = CLK

(1) = Q

(3) = DATA

(4) = Q

5-30  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

6
Automatic Tester Requirements

CHAPTER 6 - SECTION 1
Overview
Your device will be tested at two different stages during the
manufacturing process: Once at wafer probe, and once at final test
after the device has been packaged. During both testing sequences,
the automatic tester uses the test patterns which you created during
tester functional and parametric simulations. The tester verifies the
correct operation of each circuit by clocking in test patterns, and then
checks to see whether the output patterns are identical to the ones
predicted by simulation. The tester also verifies static and switching
parametric values on the pins of the device.

Information about how the testers used at LSI Logic operate can help
you to decide the best way to create test patterns for your circuit.

 LSI Logic Corporation 1988 - 2000 6-1

An ASIC Primer

CHAPTER 6 - SECTION 2
Test Flow
The types of tests performed by the tester and the sequence in
which they occur is shown in Figure 6.1. When LSI Logic tests a
device, the tests are executed in the order shown in Figure 6.1. This
method reduces test time by identifying the most common failure
modes early in the sequence. Most manufacturing defects will be
detected in the first three tests.

LSI Customer Customer Customer Customer

Tester Functional VIL / VIH 3-State IDD
Setup Patterns Patterns Patterns Patterns
Open / Short X

Gross IDD X

Gross Functional X

Three-State X

VIL/VIH X
PROCMON X

Functional X

Static IDD X

VOH X

VOL X

IIL/IIH X

IOZ X

Figure 6.1
Test Flow

6-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

CHAPTER 6 - SECTION 3
Automatic Tester Configuration
LSI Logic currently uses a number of different automatic testers:
These Production Testers can handle pin counts up to 784 pins, can
run at speeds up to 100MHz, and can handle up to 8 scan chains.
Some of the testers are also setup to provide mixed mode capability.
These testers physically apply voltage and/or current to the device
under test, measure the voltage and current levels on the inputs and
outputs, and measure propagation delays through representative
paths.

Described most simply, a tester operates in a periodic fashion. Input

signals change states at the beginning of the test period. Output
signals are strobed at the end of the period to determine whether the
measured values match the simulated values, as shown in Figure
6.2.

T0 T0 T0

Input

0 0
Output

Strobe

Figure 6.2
Fundamentals of Tester Operation

 LSI Logic Corporation 1988 - 2000 6-3

An ASIC Primer

Figure 6.3 depicts how the automatic tester applies input patterns
and measures outputs of the device under test. The shaded box at
the left labelled A represents the local memory in the tester as well
as the test patterns contained in the local memory. The patterns
from the input columns are taken out of the local memory and
applied to the device under test one pattern at a time. At the same
time, the patterns from the output columns are compared with the
output pins of the device under test.

A single binary bit in the local memory does not contain sufficient
information for the tester to apply the correct waveform to the
device under test. Therefore, a waveform generator is needed to
generate the proper waveform characteristics. All LSI testers have
at least six such generators, referred to as input timesets.

6-4  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

RZ DNRZ NRZ

1 1 0 1 1 CLK DATA CNT

DUT
OUT1OUT2
1 1 1 1 0

1 0 1 0 1

1900 1800 (nS)

TESTER LOCAL
MEMORY
AUTOMATIC TESTER

Figure 6.3
Automatic Tester Equivalent Diagram

A timeset is like the waveform generator that you’d normally use on

the lab bench. Like a waveform generator; each timeset can be
programmed to have a particular frequency, delay, pulse width had
duty cycle. However, unlike a waveform generator, once a timeset is
programmed to a particular set of timing characteristics such as
frequency, you cannot change these characteristics for the duration
of a given test block.

 LSI Logic Corporation 1988 - 2000 6-5

An ASIC Primer

By using the combination of the binary patterns in the tester’s local

memory and the timing information in the input timeset, you can
specify the precise input waveform to be used as input stimuli to the
device under test. The six input timesets are used to replicate
simulation input. Two output timesets are used to measure
functionality and propagation delays.

Figure 6.3 shows the equivalent automatic tester configuration.

Up to six timesets may be used. The number of timesets used is

governed by the number of groups of input pins that have identical
timing characteristics. Suppose, for example, that a device has
sixteen data pins, two clock pins, and three control pins. Also
suppose that this device has a timing diagram like the one in Figure
6.4. If you were to apply signals to all the data pins simultaneously,
then you could group all sixteen pins into a single timeset. Since the
two clock pins have different delays and duty cycles, they must
each use a separate timeset. Two of the control pins, CNTRL1 and
CNTL2, have identical delays and therefore are grouped into a
single timeset. CNTL3 has a different delay, and so it must use its
own timeset.

As a result, five timesets are required for this device as shown in

Figure 6.4.

6-6  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

tg1T
TS!

CLK1 tg2

CLK2 tg3

CNTL1

CNTL2 tg4

CNTL3

tg5

Figure 6.4
Example Showing Five Timesets

CHAPTER 6 - SECTION 4
Test Period Length
The minimum test period size must be greater than or equal to two
parameters:

• 10 ns minimum (the timing constraint on the

100 MHz tester.

• the maximum propagation delay in the circuit

 LSI Logic Corporation 1988 - 2000 6-7

An ASIC Primer

plus tester overhead, which will be explained

later

The test period must be greater than the maximum propagation

delay in the circuit because at the end of each test period, only the
steady-state values of the output pins are measured. The test period
must therefore be long enough so that an input change causes
outputs to change state within the same test period, and not in a
subsequent test period.

The maximum propagation delay is defined as the longest delay of

the circuit. This is typically a path from an input pin to an output
pin, as shown in Figure 6.5.

IBUF
B4

Many macrocells
Figure 6.5
Maximum Propagation Delay Path

If the test period is not long enough to accommodate the maximum

propagation delay, the circuit may be improperly tested. For
example, in Figure 6.6, we purposely shortened the test period to
one-half the length of that in Figure 6.2, creating twice the number
of test periods and output strobes. As you can see in Figure 6.6, the
state of the output will be different at the third strobe point using
the new cycle time than it was at the second strobe point using the
original cycle time. The circuit still operates correctly, but the
choice of too small a cycle time in Figure 6.6 make it appear as if

6-8  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

the outputs are in the wrong state. Differences between worst-case

and best-case simulations could cause discrepancies like these if the
cycle time was not set long enough to cover expected delays.

T0 T0 T0 T0 T0 T0

Input

Original strobe times

1 1 1 0 0 0

Output

Strobe

Figure 6.6
Defining a Shorter Test Period

T0 T0 T0

Input

Best case
Worst case
1 1 1

Output 0 0 0

Strobe
T0 ≥ Tester minimum
Tpd MAX + Tester Guardband

Figure 6.7
Minimum Cycle Time

 LSI Logic Corporation 1988 - 2000 6-9

An ASIC Primer

In most cases, actual (manufactured) parts work faster than their

simulated counterparts. The outputs will switch earlier, shifting the
output waveforms to the left as shown Figure 6.7.

Notice that if the test period T0 is shortened beyond a certain point,

the nominal- and best-case simulation results might be as expected,
but the worst-case simulation results could fail functionally because
the output could change too close to the cycle boundary or even
after the cycle boundary if the cycle time is too small. The 10 ns
minimum cycle time is due to the 100 MHz limitation of the tester.
Although the real silicon will always be faster than the worst-case
simulation prediction, reliable testing must be based on the
possibility that real parts could vary enough in performance to
reach the worst-case prediction. LSI Logic simulation requires that
parts be tested rigorously but not so rigorously as to fail good parts
that are simply closer to the worst-case limit than others.

The minimum tester cycle time can be determined by following

these three steps:

1. Determine the number of nanoseconds it takes for

an input signal’s active edge to change from low
to high after the beginning of the tester period.

2. Determine the number of nanoseconds it takes

for an output to reach steady state after the
input signal’s active edge changes. Add this
time to the time determined in Step 1.

3. Add about 2 ns tolerance to the time calculated

above. Section 6.7, Output Timesets, explains
the origin of the 2X + Y ns tolerance.

For example, in Figure 6.8, CLK is an input signal whose active

edge goes from low to high 3.0 ns after the beginning of the tester

6-10  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

period. MAT is an output that in the worst case reaches steady state
10.0 ns after the active edge of CLK. Adding another 2X + 4 ns
tolerance demands a minimum test period size of 27 ns.

CLK 3.0 ns

10.0 ns 14 ns
MAT

Figure 6.8
Determining the Minimum Test Period

CHAPTER 6 - SECTION 5
Input Timesets
Using input waveforms, you may specify that input signals change
state either at the beginning of the test period, or after a fixed delay
measure from the beginning of the test period. Alternatively, you
may specify that input signals have a narrow duty cycle within the
period.

Together, the six input timesets replicate a set of simulation inputs.

Four types of waveforms can be used for each of the input timesets
as follows:

NRZ Non-Return-to-Zero. The input changes at

the beginning of the test period and stays
in the same state throughout the entire
test period. It is not necessary for the
input to change state at every test

 LSI Logic Corporation 1988 - 2000 6-11

An ASIC Primer

period, but if it does, it must change state

only at the beginning of each cycle. This is
the preferred input type. Bidirectional pins
must usually use NRZ signals.

DNRZ Delayed-Non_Return-to-Zero. The input

changes state only after some fixed delay
measured from the beginning of the test
period. The delay must be at least X ns to
satisfy our lowest cost tester hardware
specifications, but no more than the length
of the period minus X ns. It is not necessary
for the input to change state during each test
for all periods. DNRZ waveforms are now
allowed for bidirectional pins.

RZ Return-to-Zero after pulsing high. The

input has a duty cycle within the test
period. The first edge must be at least
X ns after the start of the test period.
The second edge must occur no later
than X ns before the end of the test
period. The minimum pulse width is
X ns. Note that in any period in which
there is no pulse, an RZ signal must be
held low. The position or delay of the
pulse must be the same in every test
period where it occurs. RZ waveforms
are not allowed for bidirectional pins.

RTO Return-to-One after pulsing low. The

input has a duty cycle within the test
period. The first edge must be at least
X ns after the start of the period.
The second edge must occur no later

6-12  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

than X ns before the end of the test

period. The minimum pulse width is
X ns. Note that in any period in which
there is no pulse, an RTO signal must
be held high. The position or delay of
the pulse must be the same in every test
period where it occurs. RTO waveforms
are now allowed for bidirectional pins. Again
all times are specified for our least expense
tester.

The four types of input waveforms are illustrated in Figure 6.9. The
numbers in the figure refer to the following waveform conventions:

1. Input waveforms cannot change within X ns from

the end of the test period.

2. Input waveforms cannot change within X ns from

the beginning of the test period.

3. Pulse width must be rater than or equal to 10 ns.

4. Dissimilar Inputs (such as clock and data pins)

must not change states within Y ns of each other.

 LSI Logic Corporation 1988 - 2000 6-13

An ASIC Primer

T0 T0 T0

Input

4
≥ 4 ns
D1 D1
1
≥ 10 ns
D2 D2

D3 3
2
≥ 10 ns ≥ 10 ns

Figure 6.9
Four Types of Input Waveforms

Since they are mechanical devices, the automatic testers are neither
as stable nor as precise as the simulation results. When two or more
signals change at the same time, the simulation will accurately
predict circuit operation, but it will not analyze the effect of any
tester head skew (time difference) between the input signals. Even
if you specify the input signals in the same timeset, the tester may
not be able to apply them at exactly the same time. The skew
between any two input pins may be up to Y ns. If the sequence of
certain input signals is critical, either apply them at different
periods, or separate them in different timesets with sufficient delay
between them (normally X ns).

For ideal testing conditions, all inputs should be NRZ type with
data pins changing on the non-active edge of the clock pins, as
illustrated in Figure 6.10.

6-14  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

T0
Data

CLK

Figure 6.10
Idealized Input Test Patterns

T0 T0 T0
Test period 1 Test period 2 Test period 3

For RZ

For RTO

For DNRZ

Figure 6.11
Legal Timeset Initialization

Once assigned to an input timeset, inputs cannot be changed to

another timeset within the same test block.

 LSI Logic Corporation 1988 - 2000 6-15

An ASIC Primer

The first test vector of each set of patterns must be an X for every
input pin and a Z for each bidirectional pin. An example of legal
timeset initialization patterns is shown in Figure 6.11. Undriven
non-bidirectional inputs should have a value of either 0 or 1. The X
or Z values are required only on the first pattern of the set (time 0).

Except for zero delay NRZ, input value changes are not legal at
cycle boundaries, and different input timesets must not have the
same delay value.
Bidirectional pin conflicts are not allowed during functional or
three-state test patterns at any time.

In most situations, you can generate all of the test patterns using
only NRZ waveforms for all input pins. However, if the number of
patterns required to test the part is larger than about 32K, use
DNRZ, RZ, or RTO types of waveforms for test pattern
compaction.

Figure 6.12 shows a set of ten test patterns using only NRZ
waveforms. In Figure 6.13, the same set of test patterns is
compacted by using an NRZ waveform for the DATA in, and an RZ
waveform for the CLK in. This arrangement uses only five test
patterns.

6-16  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

DATA T0 T0 T0 T0 T0 T0 T0 T0 T0 T0

CLK

Figure 6.12
Input Test Patterns Using NRZ Waveforms Only

T0

DATA

CLK

Figure 6.13
Adding RZ to NRZ Reduces the Number of Patterns

CHAPTER 6 - SECTION 6
Examples of Valid and Invalid Waveforms
Figure 6.14 shows how a set of simulation inputs is replicated by
input timesets. Look at the figure to observe that Inputs DATA 0 to
7 are NRZ waveforms, CLOCK 0 is an RZ waveform with 4.0 ns

 LSI Logic Corporation 1988 - 2000 6-17

An ASIC Primer

delay and 4.0 ns pulse width, ENABLE 1 in a DNRZ waveform

with 4.0 ns delay, and SELECT 1 and SELECT 2 are DNRZ types
of waveforms with 6.0 ns and 8.0 ns delay, respectively.

Cycle 1 Cycle 2 Cycle 3 Cycle 4

DATA 0 ~ 7

4ns 4ns
CLOCK 0 4ns 4ns

ENABLE 1
4ns

SELECT 1 6ns
6ns

SELECT 2
8ns 8ns

Figure 6.14
Simulation Inputs Replicated by Timing Generators

Figure 6.15 shows a set of simulation inputs that cannot be

replicated by input timing generators because of the following
errors:

6-18  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

Error 1
DATA 0 ~ 7

2ns
4ns
CLOCK 0 4ns
Error 3 Error 2

4ns
ENABLE 1
8ns
Error 4
4ns
CLOCK 1 4ns
4ns 6ns

Figure 6.15
Simulation Inputs that Can’t be Replicated by Timing Generators

• Error 1 - Inconsistent delay. Illegal use of an NRZ

or DNRZ timeset.

• Error 2 - Inconsistent pulse width. Illegal use of

an RZ timeset.

• Error 3 - Inconsistent delay. Illegal use of a DNRZ

timeset.

• Error 4 - Inconsistent delay. Illegal use of an RZ

timeset.

CHAPTER 6 - SECTION 7
Output Strobes
The output strobes are similar to a camera in that they capture or
strobe output states at specified points in time: when the outputs
reach steady state. The captured states are then compared with

 LSI Logic Corporation 1988 - 2000 6-19

An ASIC Primer

simulation results.

LSI Logic uses two output strobes to monitor the circuit’s output
pins in order to measure functionality of your design. They are used
to capture the steady-state results of the output pins at the end of
each test period.

Each set of test patterns should be strobed to detect a functional

state. The output strobe checks all valid output pins in every test
period to capture the steady-state signal whenever the output is in a
0 or 1 state. The X and Z states, however, are not checked during
functional testing. All Z states are checked during three-state
testing.

The strobes must be set at the end of the test periods. The minimum
test period size is equal to Tpdmax +TOL ns, where Tpdmax is the
maximum propagation delay within the circuit.

The TOL tolerance for the tester is illustrated in Figure 6.16. The
minimum pulse width of a timeset is X ns. The measured value is
compared with the simulated value during this X ns.

6-20  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

15
IN

OUT
10 14
STROBE

IN’
4
OUT’

STROBE’ 10 10

Figure 6.16
Tester Tolerance

Due to the tester’s tolerance, a timeset must not change state within
X ns from either the beginning or the end of the test period.
Therefore, the trailing edge of the strobe cannot be closer than X ns
from the end of the test period. The accounts for 2X ns out of TOL
allowed for tester tolerance. An additional Y ns is added to strobe
time to compensate for the physical limitations of the tester, as
explained in the next paragraph.
When simulating the chip in software, you can specify that an input
signal make a low-to-high transition 15 ns after the beginning of the
test period. Likewise in the tester, you can program the driver for
this input pin to make a low-to-high transition 15 ns after the
beginning of the test period. However, it takes Y ns for the voltage
level at the output of the driver to reach the input threshold voltage
level of the input buffer on the chip. This tolerance factor must also
be allowed for when determining the minimum cycle time.

 LSI Logic Corporation 1988 - 2000 6-21

An ASIC Primer

Wafer testing is performed at a much slower test speed due to the

large inductance and capacitance of the wafer sorting hardware.
Because of this, all output signals must be stable before they are
strobed. All output changes caused by input changes in the same
time block must be at the final steady state before they are strobed.

The above-mentioned testing requirements may mean testing your

design below operational frequency. However, switching
performance testing is not limited. Testing switching performance
is accomplished by comparing output delays with simulation
delays. Overall switching performance is based on the
measurement made using the critical path. To make this
comparison, use the performance strobe (TS8) to measure the delay
of a representative path. In this situation, we make the assumption
that the delays of all paths in the chip are within specification if the
propagation delay through this path matches the delay from
simulation. This assumption is valid since cells on the same silicon
die and in the same package are subject to the same external
voltage, temperature, and process variations. This delay times will
track each other.

CHAPTER 6 - SECTION 8
Rules for Creating Time Blocks
Simulation test patterns are sometimes broken up into time blocks
that run sequentially. The term time block is used in Chapter 7,
Prepare Test Patterns for Simulation, and in this chapter.
Because the terminology is used in a different context in this
chapter, we will refer to time blocks used for simulation as
simulation time blocks, and time blocks used for the tester as tester
time blocks. You may have one or more simulation time blocks for
every set of test vectors in a single tester time block. The following
rules apply to tester time blocks.

1. You may create multiple tester time blocks of tester

functional test patterns . A different number of test

6-22  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

blocks are allowed for different technologies. LSI

Logic will charge additional NRE charges.
Collectively, these are called tester functional
patterns.

2. Additionally, one of each of the following time

blocks must also be created for the tester:

• One tester time block for tester

VIL/VIH test patterns.

• One tester time block for three-

state test patterns.

• One tester time block for IDD

test patterns.

3. Each tester time block must be independent and

must contain its own initialization patterns.

At the beginning of the test program, two initial vectors are always
required. One defines the function of all pins at the outset of testing,
and one specifies which output pins will be strobed. Each time you
want to add or delete output pins from the output pins to be strobed,
a new test vector is required. If your circuit contains bidirectional
pins, allow two test vectors for each simulation cycle in which the
function (direction) of those pins will be switched in order to avoid
contention.

If a time block has more than the maximum number of test vectors
allowed for the tester you have selected, you have one of two
choices: either eliminate some test vectors or divide them into two
tester blocks.

The test period size must be the same for each set of test vectors in
a given time block. However different time blocks may have

 LSI Logic Corporation 1988 - 2000 6-23

An ASIC Primer

different test period sizes, input timing specifications, and output

strobe timing.

Within any set of functional test patterns, all output pins must be
fully stabilized Y ns prior to the beginning of the next test period.
For example, if the slowest output on the chip has a delay of 9.6 ns,
the test period for that particular set of patterns must be at least 9.6
ns + Y ns.

The tester time block for the VIL/VIH,IDD, and three-state patterns
must be 2 uS (2000 ns).

Remember that all test simulations using test conditions are

performed at 25°C, nominal VDD, and worst-case process
conditions.

6-24  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

7
Automatic Tester Requirements

CHAPTER 7 - SECTION 1
Overview
Recall from Chapter 6 that the LSI Logic automatic testers use test
patterns created during tester functional and parametric simulations.
The tester verifies the correct operation of each circuit by clocking in
the test patterns, monitoring the output pins, and checking them
against the output patterns predicted by simulation.

The following sets of simulations should be prepared:

• system simulation

• tester functional simulation

• VIL/VIH simulation
• three-state functional simulation

• tester IDD simulation

• operationally, patterns for LSI Logic proprietary

functions (i.e. cores, mixed signal, etc.)

 LSI Logic Corporation 1988 - 2000 7-1

An ASIC Primer

This chapter provides guidelines you should follow in preparing

simulation test patterns, and includes the following major sections:

Section 7.2 Prepare Simulation Test Patterns

Section 7.3 Determine the Output Strobe Timing

Section 7.4 Avoid Voltage Spikes and Bus Contention

CHAPTER 7 - SECTION 2
Prepare Simulation Test Patterns
This section contains information you should know in order to
prepare simulation test patterns, including system simulation
patterns, tester functional patterns, VIL/VIH patterns, IDD
measurement, and three-state patterns.

Prepare System Simulation Input Patterns

System simulation input patterns are used to exercise the network

under user-specified environmental conditions. The purpose is to
validate the design and its performance. By analyzing the results,
you will be able to determine the network’s operational limits. The
results of this system simulation should indicate to you whether
your network behaves as you predict, and what AC performance
limits exist. If the system does not perform as expected, you may
have to make design modifications and re-create system simulation
test patterns.
There are no waveform restrictions in system simulation, which
means that input patterns may be either periodic or asynchronous.
There are also no timing restrictions; the simulator can analyze a
circuit at any frequency, and at a resolution as fine as 10 psec.

Input patterns may be skewed and/or the frequency of clock signals

may be changed. The absence of waveform and timing restrictions

7-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

in system simulation enables the design to be simulated more

thoroughly than the automatic tester can test it.

System simulation should be performed in ascending hierarchical

order from the most primitive module in the design to the most
inclusive. This approach saves time when debugging the networks,
and ensures that the foundation of the circuit is sound.

Prepare Tester Functional Simulation Input Patterns

Once the desired results are obtained from system simulation, input
patterns should be created for use by the automatic tester. The tester
functional simulation patterns enable the tester to verify the
device’s functionality, measure the propagation delay through the
representative path, and measure some Static parametric values
such as output voltage levels (VOL, VOH) and three-state output
leakage current. Functional simulation is performed only on the
top-level module, and should thoroughly exercise the network - i.e.,
cause every node in it to change state, or toggle, both from a 0 to 1
and from a 1 to 0 at least once. Full scan testing is used today by
most designers to create high fault coverage test vectors. And the
IDDQ tool in the FlexStream toolset allows designers to check their
toggle coverage.

To perform tester functional simulation, it is necessary to extract a

subset of the system simulation patterns and scale them down to the
timing and waveform requirements of the automatic tester. The
recommended way of revising system simulation patterns for use as
tester functional patterns is to reduce the frequency at which
patterns are applied to the inputs of the network.

Occasionally, multiple test blocks (tester memory loads) may be

used to perform complete functional testing. Each block can have
different timing characteristics. For example, each block of patterns
can test the device with a different test period size, different input

 LSI Logic Corporation 1988 - 2000 7-3

An ASIC Primer

timing, and different output strobe times. Different output pins may
be monitored on each block of patterns.

Tester functional input patterns must be periodic. They must also

conform to one of the following types:

• Non-Return-to-Zero (NRZ)

• Delay-Non-Return-to-Zero (DNRZ)

• Return-to-Zero (RZ)

• Return-to-One (RTO)

A more specific description of recommended tester functional input

waveforms may be found in Chapter 6, Automatic Tester
Requirements.

Prepare Tester VIL/VIH Simulation Input Patterns

Tester threshold simulation patterns are used to test the circuit for
input low voltage and input high voltage. These patterns enable the
tester to measure the circuit for input threshold voltages. The tester
VIL/VIH simulation must be performed under the following
conditions: VDD =<Nominal Value>, TEMP=25 C, worst-case
process, and a test period of 2000 ns. Only 0 ns delay Non-Return-
to-Zero (NRZ) signals are legal for the VIL/VIH test patterns.

The Required VIL/VIH NAND Tree Test Circuit

In order to measure input voltage accurately while ignoring current

spikes, every design must have built-in VIL/VIH check circuitry.
Figure 7.1 shows the recommended VIL/VIH test circuit. This
circuit includes a 2-input NAND gate at every input pin as well as a
TESTOUT output pin. The NAND tree provides a purely

7-4  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

combinational path from all the input pins to a single TESTOUT

pin. Most LSI Logic products already have these gates built into all
input and bidirectional buffers.

Inputs
mi VDD To circuit

m2 To circuit

m1 To circuit
procmon mpout
mode
VDD

Figure 7.1
The Vil/Vih NAND Tree

The VIL/VIH test is executed with the input driver levels set at
VIL, (max) and VIH (min). This leaves no noise margin for the
input buffers. While switching, output buffers can generate large
current spikes capable of causing the input buffers to misread
voltage (and therefore logical) levels.

Recall that a typical output buffer may sink or source 4 mA in the

static mode. Dynamically, these buffers generate instantaneous
currents of 50 mA or more. Several output buffers switching
simultaneously can generate a current spike of one or more amps.
As a result, internal storage elements may be latched into an
incorrect state, causing incorrect states on the output pins.

To alleviate the noise problem created when a large number of

output pins switching simultaneously, only the TESTOUT pin is
monitored when the VIL/VIH test us executed on the tester. All

 LSI Logic Corporation 1988 - 2000 7-5

An ASIC Primer

other output pins are temporarily masked. The noise generated by

the simultaneous switching of a large number of output pins may
cause some internal storage elements to go into the wrong state.
This does not affect the tester’s ability to perform the VIL/VIH test,
however, as the tester is monitoring only the TESTOUT pin, but it
is recommended that all other outputs accept the TESTOUT output
be either three-stated or in a non-switching mode if possible to
reduce noise into the ground bus during the sensitive input
threshold measurements.

The VIH/VIH test patterns are specific. The first pattern must
consist of 1’s on the input pins for non-inverting input buffers and
0’s on the inputs of inverting input buffers. Then 0’s or 1’s are
applied according to the procedure shown in Figure 7.2 (all inputs
are non-inverting in this example). Note that a circuit containing n
number of input pins requires the use of at least (n + 1) VIL/VIH
test patterns. To test threshold both low-to-high and high-to-low
may require twice this number of vectors, but the overall vector
length is short and is determined by the number of inputs and
bidirect pins in a design.

For networks that do not have an extra output pin for the VIL/VIH
test circuit, a regular output pin or pins may be used to multiplex a
regular output with the VIL/VIH test output. Figure 7.2 shows VIL/
VIH test patterns. Make sure not to use a critical path output,
however, if you do intend to multiplex a functional output with this
TESTOUT signal.

7-6  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

TESTOUT

M1 M2 M3 Mi i=even i=odd

1 1 1... 1 1 0
1 1 1... 0 0 1
1 1 0... 0 1 0
1 0 0... 0 0 1
0 0 0... 0 1 0

Figure 7.2
Example NAND Tree Input Patterns

If regular input pins were used to multiplex the output for the VIL/
VIH circuit, a set of supplementary test patterns would have to be
added to the VIL/VIH test patterns. The purpose of these
supplementary test programs would be to check those input pins
that were used to multiplex the VIL/VIH test output.

The example circuit in Figure 7.1 includes only unidirectional input

pins. On designs that have a bidirectional bus, running the normal
parametric NAND tree pattern may cause bidirectional drivers to
turn on, creating a bus conflict condition. This problem may be
resolved by splitting up the NAND tree into different sections.
Figure 7.3 is an example of such a circuit. In this example,
bidirectional pins have been disabled with the TE pin so as to place
them in the input mode.

 LSI Logic Corporation 1988 - 2000 7-7

An ASIC Primer

M1 5V

M2

M3 MUX
TN
A

B TOUT
S
Normal
system
M4 output
TN signal

TE

Figure 7.3
A NAND Tree with Bidirects

This allows the input section of these buffers to be tested as if it

were an input buffer only. The 2-input NAND gates on the three-
state control lines allow these bidirect pins to be forced into the
proper state for this test. A MUX is used to direct this NAND tree
output to an existing pin to eliminate the need for additional chip
pins. The select line of this MUX is controlled by the TE Pin.
Figure 7.4 shows the test patterns used for this example.

Only a 0 ns delay, Non-Return-to-Zero (NRZ) signal is allowed for

inputs during the VIL/VIH test.

The use of a test-enable input pin in a circuit containing

bidirectional pins permits writing of input test patterns as though
they were for a circuit containing only unidirectional input pins.

7-8  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

M1 M2 M3 Mi TE TOUT

1 1 1 1 1 0
1 1 1 0 1 1
1 1 0 0 1 1
1 0 0 0 1 0
0 0 0 0 1 1
Put “0” on input B of 2:1 mux 0 0

Figure 7.4
Test Patterns for NAND Tree with BiDirects

Prepare Test Patterns for the Chip IDD Measurement

Only a few patterns are needed to set up the circuit for Static
leakage current (the IDD test). Test conditions should be set up in
which all Static current is turned off to facilitate easy measurement
of Static leakage. The test condition must eliminate the following
sources of Static current:

• floating internal nodes

• internal bus contention

• a low state on an input or output pin with a pull-up
resistor

• a high state on an input or output pin with a

pull-down resistor

• any undriven input without a pull-up or pull-down

resistor

 LSI Logic Corporation 1988 - 2000 7-9

An ASIC Primer

• enabled memory sense amplifiers

• static current in Mixed Signal and Analog I/Os

The first two items mentioned above take priority over the others if
all cannot be attained.

The IDD test patterns must be executed under the following

conditions: VDD = <Nominal>, temperature = 25° C, process =
worst case, test period = 2000 ns.

Prepare Three-State Test Pattern

Three-state test patterns are used to test the three-state enable

circuitry. These test patterns are typically very short, consisting of
only a few vectors. Three-state functional test patterns must
exercise every possible way that the three-state outputs can be
three-stated.

When the automatic tester executes the tester three-state functional

patterns, the output pins are monitored for the high impedance (Z)
state. All output pins in the 1 or 0 state are masked.

You must execute the three-state simulation under the following

conditions: VDD = <Nominal>, temperature = 25° C, process =
worst case, test period = 2000 ns.

To summarize, there are at least four sets of tester patterns that you
must create:

1. Tester functional 3. Tester IDD

2. Tester VIL/VIH 4. Tester three-state

Additional sets of patterns may be required if you are using LSI

Logic proprietary functions. Predefined simulation files are
provided by LSI Logic if you choose to use them.

7-10  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

CHAPTER 7 - SECTION 3
Determine the Output Strobe Timing
Recall from Chapter 6 that two strobes are available. Both strobes
allow a direct comparison of expected (simulated) and actual
(measured) outputs. The first strobe is placed near the end of each
test period to test the functional operation of most of the output
pins. The second strobe can be used to get a feeling for the
performance of the chip. Place the second strobe earlier in time than
the first strobe within the cycle to measure a representative path
delay using the remaining output pins, as shown in Figure 7.5.

Input

Output
#1

Output
#2

Strobe
#1
(TS7)

Strobe
#2
(TS8)

Figure 7.5
Using Output Strobe #2

When choosing a representative path for the second strobe, you

should pick one of the longest paths in the design because it will
contain the most internal macrocells and will, therefore, be
representative of most of the circuit elements, as shown in Figure

 LSI Logic Corporation 1988 - 2000 7-11

An ASIC Primer

7.6. In a shorter path, the delays being measured are really those of
the input and output buffers and not the delay of the internal
macrocells. An example of this is shown in Figure 7.7.

IBUF

D Q B4
LD1

G QN Many macrocells

Figure 7.6
Long Path Delay, Dominated by Chip Circuitry

OUT
D Q
CLK
CP

Figure 7.7
Short Path Delay, Dominated by Tester Capacitance

7-12  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

CHAPTER 7 - SECTION 4
Avoid Voltage Spikes and Bus Contention
This section describes information about how to avoid voltage
spikes and bus contention. It is helpful to understand this
information before creating simulation input patterns.

Avoid Voltage Spikes

A voltage spike, also known as a “glitch” is the abrupt transition on

any node from one state to another and then back again. Voltage
spikes occur when the propagation delay (or tPD) of a given gate is
longer than the duration of the input signal pulse width. The reason
to avoid voltage spikes is that they prevent a gate from reaching
either a logical 1 or a logical 0 state. Figure 7.8 depicts a 2-input
NOR gate (NR2) with a low-to-high propagation delay of 0.6 ns
and a high-to-low propagation delay of 0.2 ns.

A B LOW

tPDLH=0.6 ns Z
tPDHL=0.2 ns

tPDLH=0.6 ns tPDHL= 0.2 ns

Figure 7.8
NOR Gate Output Edge Asymmetry

Figure 7.9 shows the conditions in which a voltage spike will occur
on the NR2 gate in Figure 7.8.

 LSI Logic Corporation 1988 - 2000 7-13

An ASIC Primer

Low
B

A
Voltage level questionable
.3 ns
Output was not given enough
time to stabilize after input
Z
tPDLH=.6 ns

t+.3 t+.6

Figure 7.9
NOR Output Asymmetry Generating a Spike/Glitch

In Figure 7.9, input B is low and input A is high. When input A

makes the high-to-low transition at time t, output Z begins to go
high. Input A returns to the high state after 0.3 ns; this pulls the
value of output Z back to a low, which it attains in 0.1 ns. The solid
line for output Z depicts its transition back to the low state; the
dotted line shows the transition output Z would have made to the
high state if output A had not returned to a logical 1 at t+0.3 ns.

To prevent the voltage spike, input A must hold steady at logical 0

for 0.6 nsec or more, affording output Z time enough to change
from logical 0 to logical 1, as illustrated in Figure 7.10.

7-14  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

B Low

OK to change after t+.6

A

High

Low
Z

t t+.6
If

Figure 7.10
Avoiding the Glitch

Figure 7.10 shows the conditions under which a voltage spike will
not occur on the same NR2 gate.

Avoid Bus Contention

Bus contention occurs when an output buffer is enabled, and is

outputting a logical value that differs from that of the external
driving source. External bus contention can be avoided by three-
stating the output buffer whenever the bus is being driven by an
external source. Conversely, when the bus is being driven by an
output buffer, the external source should not be driving the pin. Bus
contention is avoided by allowing only one driver to drive a given
bus at a time. Buses must not be allowed to contend during any test
pattern.

 LSI Logic Corporation 1988 - 2000 7-15

An ASIC Primer

Figure 7.11 illustrates how external bus contention can be avoided.

A B During transitions from

one source to another,
X X neither signal drives the pin.

0 X
0 0 Output half A B
of bidirect
1 X A B Z 0
1 1 External Z 1
Source C
X 0 Z Z
X 1 Input half 0 Z
of bidirect
0 1 1 Z
1 0 Z Z
Values for signals
A, B that will
cause external
bus contention
Figure 7.11
Avoiding Bus Contention

Signal A is created by an external source C. Signal B is generated

by the output buffer. The table to the left of the drawing shows
values for A and B that will result in external bus contention. Bus
contention can occur when A or B (or both) are in an unknown state
(X), and when A and B drive the pad at the same time, or when not
enough time is allowed when switching between driving sources.

The table to the right of the drawing by contrast, show values that
will not result in external bus contention. Bus contention will not
occur when either A is in a high impedance (Z) state, while B is in a
defined state (1 or 0), or vice-versa, or when both A and B are in the
high impedance (Z) state.

7-16  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

8
The Beginning

The material in this book should be digested before you begin your
ASIC design. Most of the techniques discussed here have been tried
or used in real designs at LSI Logic. Techniques that do not work or
have caused simulation problems in ASICs have been analyzed, and
this book has proposed alternatives to resolve difficulties caused by
the original logic configurations.

The purpose of this book is to save you valuable design time by

comparing design trade-offs early in your design cycle and pointing
out things that are likely to cause problems if used in your design.

Before you begin your design, assemble all of the information

necessary to complete your ASIC design. LSI Logic technology
design manuals, datasheets and databooks are absolutely required
items in order to begin a design. Information on packaging is critical
for some trade-offs that may need to be made relating to power and
ground, ESD protection, or pin count. If you will be needing
functions or cells for your design that are not in any existing LSI
Logic libraries, you must request these functions and fells as early as
possible, probably during the first technical review. Know who your

 LSI Logic Corporation 1988 - 2000 8-1

An ASIC Primer

LSI Logic ASIC Customer Engineer (ACE) is from the first day
you begin your design work. That person will be your technical
contact throughout the duration of your interface with LSI Logic.
Use them as a resource.

More on the Testing Issue

Testing is becoming a very complex issue as designs grow in size.

Various test techniques have been addressed in this book, but many
others not covered are available to you as a designer. You will need
to partition your logic in such a way as to decide what goes on each
ASIC you design and then build in logic that will allow the design
to be tested most efficiently. Sometimes, a few additional test points
or access points can reduce the number of tests by a significant
amount. If you wish to use Memory BIST, full scan ATPG, JTAG
boundary scan, or other techniques, you must discuss the ground
rules early in your design with your ACE engineer. LSI Logic has a
series of Application Notes in the ASKK documentation package
which describe the various test techniques in great detail.

Training Classes

Training classes are available for our FlexStream Design Tools on a

regular basis at our Milpitas California facility. All information
related to training classes and material is available on our Customer
Education WEB page located at: (Internet site)

http://www.lsilogic.com/techsupp/training/index.html

For LSI Logic personnel, you can reach a more expanded list of
training classes at: (Intranet site)

http://webvision.lsil.com

All of our tools training material is downloadable from these two

8-2  LSI Logic Corporation 1988 - 2000

 An ASIC Primer

WEB sites. LSI employees can access on-line classes through a PC

interface using a WEB Browser or via UNIX machines which have
been configured to receive video and audio media files.

Additional classes are available on-line to LSI employees at this

time. These cover topics such as: Test Methodology, Packaging,
ESD Protection, and so on. These will become available to
customers over time as we add Internet media capability.

In the future, we will be providing expanded WEB based classes on

our Internet WEB site and will also add “live” classes on the WEB
via “chat room” type of access to keep travel costs down and keep
providing more timely training. See our WEB pages for details.

Questions about this Book

If you should have questions about the material contained in this

book, please contact the LSI Logic Customer Education
Department at:

LSI Logic Corporation

1551 McCarthy Blvd., MS E-197

Milpitas, CA 95035

(408) 433-7687

One of the members of our Customer Education Department will

respond to your questions. We are here to serve you and would
appreciate any feedback related to this document or any other
Training issues that you would like to discuss with LSI Logic.

Summary

LSI Logic wants to make your ASIC design experience a successful

 LSI Logic Corporation 1988 - 2000 8-3

An ASIC Primer

one. It is our philosophy that each design must work right the first
time; this requires that we provide you the highest quality products,
software tools, and training in the industry. Our Training WEB sites
Training classes, and reference documents are available to provide
you with the information necessary to make your job as easy as
possible when you are using LSI Logic’s ASIC products and design
tools.

Our floorplan for your successful design includes:

• Attitude -
that together we will make your design a success.

• Service -
which we provide at every step of the design process.

• Innovative ASIC technology -

keeping you and your products ahead of the competition.

• Customer satisfaction -
an LSI Logic priority.

Thank you for choosing LSI Logic as your ASIC vendor.

8-4  LSI Logic Corporation 1988 - 2000