L200 CDR I (2up)
L200 CDR I (2up)
Page 200-1
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
INTRODUCTION AND BASICS Why Clock and Data Recovery Circuits? In many systems, data is transmitted or retrieved without any additional timing reference. For example, in optical communications, a stream of data flows over a single fiber with no accompanying clock, but the receiver is required to process this data synchronously. Therefore, the clock or timing information must be recovered from the data at the receiver.
Data Recovered Clock
Fig. 4.2-01
Page 200-3
Properties of NRZ Data Most binary data is transmitted in a nonreturn-to-zero (NRZ) format. NRZ data is compared with return-to-zero (RZ) data below.
NRZ Data RZ Data
Fig. 4.2-02
Tb
The NRZ format has a duration of Tb for each bit period. The bit rate, rb = 1/Tb in bits/sec. The bandwidth of RZ data > bandwidth of NRZ data Maximum bandwidth of NRZ data is determined by a square wave of period 2Tb. In general, NRZ data is treated as a random waveform with certain known statistical properties.
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
The Challenge of Clock Recovery 1.) The data may exhibit long sequences of ONEs or ZEROs requiring the CRC to remember the bit rate during such an interval. The CRC must not only continue to produce the clock, but do so without drift or variation in the clock frequency. 2.) The spectrum of the NRZ data has nulls at frequencies which are integer multiples of the bit rate. For example, if rb = 1Gb/s, the spectrum has no energy at 1GHz. 500 MHz square wave with all even-order harmonics absent
1ns 2ns
1ns
Fig. 4.2-03
Page 200-5
NRZ Data Spectrum The autocorrelation function of a random binary sequence can be written as | | Rx() = 1 - Tb , || < Tb = 0, || = Tb From this, the power spectral density of a random binary sequence is written as, sin(Tb/2) 2 Px() = Tb T /2 b which is illustrated as,
Px()
2 Tb
4 Tb
6 Tb
8 Tb
Fig. 4.2-04
S.K.Shanmugam, Digital and Analog Communication Systems, New York: Wiley &Sons, 1979. ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
Edge Detection CRC circuits require the ability to detect both the positive and negative transitions of the incoming data as illustrated below,
NRZ Data Edge Detection
Fig. 4.2-05
Methods of edge detection: 1.) EXOR gate with a delay on one input.
Din Dout
Fig. 4.2-06
Dout
Fig. 4.2-07
Page 200-7
Edge Detection and Sampling of NRZ Data - Continued 3.) Use a flipflop that operates on both the rising and falling edges. This technique takes advantage of the fact that in a Din Edge phase-locked CRC, the edge-detected data is Detector multiplied by the output of a VCO as shown. In effect, the data transition impulses sample points XVCO on the VCO output. a.) Master-slave flipflop consisting of two D latches. D Latch D Latch
Din CLK D Q CLK D
Dout
Fig. 4.2-08
Dout Q
XVCO
Fig. 4.2-09
MUX
Fig. 4.2-10
P.E. Allen - 2003 Page 200-8
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
CLOCK RECOVERY ARCHITECTURES AND ISSUES Clock Recovery Architectures From the previous considerations, we see that clock recovery consists of two basic functions: 1.) Edge detection 2.) Generation of a periodic output that settles to the input data rate but has negligible drift when some data transitions are absent. Conceptual illustration of these functions:
Din
Edge Detector
High-Q Oscillator
Dout
Fig. 4.2-11
In essence, the high-Q oscillator is synchronized with the input transitions and oscillates freely in their absence. Synchronization is achieved by means of phase locking.
Page 200-9
Fig. 4.2-11
Operation: 1.) Assume the input data is periodic with a frequency of 1/Tb (Hz). 2.) The edge detector doubles the frequency causing the PLL to lock to 2/Tb (Hz). 3.) If a number of transitions are absent, the output of the multiplier is zero and the control voltage applied to the VCO begins to decay causing the oscillator to drift from 1/Tb (Hz). 4.) To minimize the drift due to the lack of transitions, LPF >> Maximum allowable interval between consecutive transitions. 5.) The result is a small loop bandwidth and a narrow capture range. Fortunately, most communication systems guarantee an upper bound of the allowable interval between consecutive transitions by encoding the data.
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
Frequency Aided Acquisition Frequency acquisition can be accomplished with and without an external reference. If an external reference clock is available, frequency acquisition can be done with a secondary PLL loop having a PFD. Frequency acquisition with Frequency acquisition with a an external reference: frequency detector:
data Phase Detector Charge Pump1 Phase Detector Charge Pump1
Loop Filter
data
Loop Filter
ref clk
Freq. Detector
Charge Pump2
If no reference clock is available, a frequency detector has to be used which requires I and Q clocks and for typical implementations, the VCO frequency cannot be off more that about 25% of the data rate.
Page 200-11
PHASE DETECTORS FOR RANDOM DATA Linear Phase Detectors up down This type of detector is represented by the Hogge detector.
data
clk
A B
A B
up down
up down
C. R. Hogge, IEEE J. Lightwave Technology, pp. 1312-1314, 1985. ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
Linear Phase Detector Continued Transfer characteristics of the Hogge phase detector:
Normalized average up/down output
+1.0 +0.5
-
11001100 data density
0 -0.5
Phase error
-1.0
Linear gain characteristics Phase detector gain is 0.5 for 11001100 data transition density Small jitter generation due to PD Suffers from bandwidth limitations Have static phase offset due to mismatch
Page 200-13
Binary Phase Detectors This type of phase detector is represented by the Alexander type of phase detector. Clock is ahead Clock is behind data A C
D Q D Q up clk
A B C A B C
down D Q B D Q
Binary Phase Detector Truth-Table ABC Decision Tri-state Clock is ahead Error Clock is behind Clock is behind Error Clock is ahead Tri-state Output ------Down ------Up Up ------Down ------000
+1
Phase error
-1
High phase detector gain Causes higher output jitter compared to linear phase detectors Static phase offset set by sampling aperture errors Widely used in digital PLL and DLLs
P.E. Allen - 2003 Page 200-14
J.D.H. Alexander, IEE Electronics Letters, pp. 541-542, 1975. ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
+1
clk
0
A up/down D Q
Phase error
-1
M. Meghelli, et. al. ISSCC2000, pp. 56-57. ECE 6440 - Frequency Synthesizers
Page 200-15
FREQUENCY DETECTORS FOR RANDOM DATA Rotational Frequency Detectors Block diagram (Richman)
I-clk DE-FF D Q A DE-FF C D Q up
+1
fdata 2 fdata 4 3fdata 8 fdata 8
data
frequency error
0
Q-clk D Q DE-FF B D Q DE-FF D down
f data 8
f data 4
3fdata 8
f data 2
-1
Timing diagram example: (VCO clock is faster than the data rate)
data I_clk Q_clk A B up down
10
11
01
00 10
10
11
00
10
11
01
01
10
01
00
10
11 01
01
10 11 11
01
00
10
Pull in range: 25% of data rate Prone to false locking in presence of jitter and/or short data pattern AB changing from 00 to 01 DOWN pulse, AB changing from 10 to 00 UP pulse
D. Richman, Proc. of IRE, pp. 106-133, Jan. 1954. ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
Phasor DiagramExamples of Rotational Frequency Detectors Phasor diagrams for a 0101 data pattern:
Q Q Q Q Q Q Q Q Q
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
01 00
11 10
If the VCO frequency is off more than 50%, the frequency is in the wrong direction.
fc
3fc 4
fc 2
fc 4
frequency
fc 4
fc 2
3f c 4
fc error
-1
Page 200-17
Rotational Frequency Detectors Continued A simpler implementation of the Richman frequency detector (Pottbacker). The data samples the clock.
DE-FF D Q
I-clk
Q1
up / down
data
Q-clk
D Q DE-FF
Q2
D FD
Q3
Very similar characteristics to that of Richmans frequency detector, however, the implementation is simpler. Pull-in range: 25% of data rate Prone to false locking in the presence of jitter and/or short data patterns
A. Pottbacker, et. al., IEEE JSSC, pp. 1747-1751, Dec. 1992. ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
CDR ARCHITECTURES Clock Recovery Spectral Line, Early-Late Enam, Abidi 1992 Interleaved decision circuit
Din Interleaved Decision Circuit Dout D Q CLK Half-rate Clock F(s) VCO
VDD
Output+ Data
D2 Phase Detector f = BT
Data
Comments: Good example of CMOS solution to practical clock recovery circuits Circuit can be analyzed as spectral line or as early-late.
Iclock
Fig. 4.2-16
Page 200-19
Clock Recovery - Quadricorrelator Analog version has three loops sharing the same VCO.
NRZ Data sin 1t Edge Detector cos 2t
Loop II
Fig 4.2-13
LPF
cos (1-2)t
Edge detector plus three loopsLoops I and II perform frequency detection Loop III performs phase detection Operation: The signal at P is (1-2) cos2(1-2) VCO is driven by sin(1-2)t + (1-2) Loops I and II drive the VCO to lock when 12. As |1-2| approaches zero, Loop III begins to generate an asymmetrical signal at node M assisting the lock process. Finally, when 12, the dc feedback signal produced by Loops I and II approaches zero and Loop III dominates, locking the VCO output to the input data.
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03) P.E. Allen - 2003 Page 200-20
d dt
Quadricorrelator Continued The use of frequency detection in the quadricorrelator makes the capture range independent of the locked loop bandwidth, allowing a small cutoff frequency in the LPF of Loop III so as to minimize the VCO drift between data transitions. Because Loops I and II can respond to noise and spurious components, it is desirable to disable these loops once phase lock has been attained. Since the combination of an edge detector and a mixer can be replaced with a doubleedge triggered flipflop, the quadricorrelator can be implemented in a digital form as shown below.
0 90 D CLK
Page 200-21
JITTER IN CDR CIRCUITS Jitter Influence on Clock Recovery Types of jitter: Long term jitter
Ideal Reference Oscillator Output T
Fig. 4.2-17
- Diverges for a free-runnng oscillator - Meaningful only in a phase-locked system - Depends on PLL dynamics Cycle-to-cycle jitter
T+T1
T+T2
T+T3
T+T4
Fig. 4.2-18
- Of great interest in many timing applications - Mostly due to the oscillator - Usually too fast for the PLL to correct
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03) P.E. Allen - 2003 Page 200-22
Jitter Due to Device Noise 1/f noise: 1/f noise is inversely proportional to frequency and causes the frequency to change very slowly. Easily suppressed by a wide PLL bandwidth. 2BKIi 2 B 2 (V2/Hz) and ini = fLi2 (A2/Hz) eni = fWiLi where KF B = 2CoxK Thermal noise: Thermal noise is assumed to be white and is modeled in MOSFETs as, 2 8kT 2 8kTgm (V2/Hz) and ini 3 (A2/Hz) eni 3gm The relationship between phase noise and cycle-to-cycle jitter is, 4 T 2(rms) 3 S() (-o)2 o
(Razavi: IEEE Trans. on Circuits and Systems, Part II, Jan. 99)
ECE 6440 - Frequency Synthesizers P.E. Allen - 2003
Page 200-23
Din
VCO fosc
Fig. 4.2-20
Din
D CLK Q
Dout
VCO
Fig. 4.2-22
P.E. Allen - 2003 Page 200-24
Substrate and Supply Noise How Do Carriers Get Injected into the Substrate? 1.) Hot carriers (substrate current) 2.) Electrostatic coupling (across depletion regions and other dielectrics) 3.) Electromagnetic coupling (parallel conductors) Why is this a Problem? With decreasing channel lengths, more circuitry is being integrated on the same substrate. The result is that noisy circuits (circuits with rapid transitions) are beginning to adversely influence sensitive circuits (such as analog circuits). Present Solution: Keep circuit separate by using multiple substrates and put the multiple substrates in the same package.
Page 200-25
Digital Ground
p+
Hot Carrier Put substrate connections as close to the noise source as possible
;;
vin vout vin vout
n+ n+ p+ p+
VGS
vin
;;; ;;;
n+ n+
iD
"AC ground"
ID
;
VDD(Analog)
p+
Substrate Noise
iD iD
VGS
vGS
Heavily Doped p
Lightly Doped p
Intrinsic Doping
Lightly Doped n
Fig. SI-01
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
Digital Ground
p+
Put substrate Hot connections Carrier as close to the noise source as possible
;;
vin vout vin vout
n+ n+ p+ p+
VGS
vin
vin VGS
;;; ;;;
n+ n+
"AC ground"
;
VDD(Analog)
p+
Substrate Noise
Heavily Doped p
Lightly Doped p
Intrinsic Doping
Lightly Doped n
Page 200-27
Computer Model for Substrate Interference Using SPICE Primitives Noise Injection Model:
VDD
Digital Ground
p+
;;
vin vout vin vout
n+ n+ p+ p+
VDD
VDD(Digital)
vin
Substrate
Coupling
Coupling
p- substrate
Cs1 = Capacitance between n-well and substrate Cs2,Cs3 and Cs4 = Capacitances between interconnect lines (including bond pads) and substrate Cs5 = All capacitance between the substrate and ac ground Rs1,Rs2 and Rs3 = Bulk resistances in n-well and substrate L1,L2 and L3 = Inductance of the bond wires and package leads Fig. SI-06
P.E. Allen - 2003 Page 200-28
Heavily Doped p
Lightly Doped p
Intrinsic Doping
Lightly Doped n
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
Computer Model for Substrate Interference Using SPICE Primitives Noise Detection Model:
VDD(Analog)
VGS
vin vin
VGS
;;; ;;;
n+ n+ Lightly Doped p Intrinsic Doping Lightly Doped n
;
RL vout RL vout
VDD
RL L4 L6 VGS Substrate Cs6 Rs4 Cs7 vout CL
Analog Ground
p+
Cs5
L5
p- substrate
Cs5,Cs6 and Cs7 = Capacitances between interconnect lines (including bond pads) and substrate Rs4 = Bulk resistance in the substrate L4,L5 and L6 = Inductance of the bond wires and package leads
Heavily Metal Doped n
Heavily Doped p
Fig. SI-07
P.E. Allen - 2003
Page 200-29
Other Sources of Substrate Injection (We do it to ourselves and cant blame the digital circuits.)
Inductor Substrate BJT Collector Base n+ p+ p- well Emitter n+ Collector n+
;;;;;;;;
Fig. SI-04 Intrinsic Doping Lightly Doped n Heavily Metal Doped n
P.E. Allen - 2003 Page 200-30
Heavily Doped p
Lightly Doped p
Also, there is coupling from power supplies and clock lines to other adjacent signal lines.
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03)
What is a Good Ground? On-chip, it is a region with very low bulk resistance. It is best accomplished by connecting metal to the region at as many points as possible. Off-chip, it is all determined by the connections or 20 bond wires. Settling Time to within 0.5mV (ns) The inductance of the bond wires is large enough 16 to create significant ground potential changes for 12 Peak-to-Peak Noise (mV) fast current transients. 8 di 4 v = L dt 0 Use multiple bonding wires to reduce the ground 0 1 2 3 4 5 6 7 8 Number of Substrate Contact Package Pins noise caused by inductance. Fig. SI-08 Fast changing signals have part of their path (circuit through ground and power supplies. Therefore bypass the off-chip power supplies to ground as close to the chip as possible.
C2 t=0
Vin
C1
Vout
VDD VSS
Fig. SI-05
P.E. Allen - 2003
Page 200-31
Summary of Substrate Interference Methods to reduce substrate noise 1.) Physical separation 2.) Guard rings placed close to the sensitive circuits with dedicated package pins. 3.) Reduce the inductance in power supply and ground leads (best method) 4.) Connect regions of constant potential (wells and substrate) to metal with as many contacts as possible. Noise Insensitive Circuit Design Techniques 1.) Design for a high power supply rejection ratio (PSRR) 2.) Use multiple devices spatially distinct and average the signal and noise. 3.) Use quiet digital logic (power supply current remains constant) 4.) Use differential signal processing techniques. Some references
1.) D.K. Su, M.J. Loinaz, S. Masui and B.A. Wooley, Experimental Results and Modeling Techniques for Substrate Noise in Mixed-Signal ICs, J. of Solid-State Circuits, vol. 28, No. 4, April 1993, pp. 420-430. 2.) K.M. Fukuda, T. Anbo, T. Tsukada, T. Matsuura and M. Hotta, Voltage-Comparator-Based Measurement of Equivalently Sampled Substrate Noise Waveforms in Mixed-Signal ICs, J. of Solid-State Circuits, vol. 31, No. 5, May 1996, pp. 726-731. 3.) X. Aragones, J. Gonzalez and A. Rubio, Analysis and Solutions for Switching Noise Coupling in MixedSignal ICs, Kluwer Acadmic Publishers, Boston, MA, 1999.
ECE 6440 - Frequency Synthesizers Lecture 200 Clock and Data Recovery Circuits - I (6/26/03) P.E. Allen - 2003 Page 200-32
(To be continued)