Clock and Data Recovery for Serial Digital Communication

Rick Walker Hewlett-Packard Company Palo Alto, California

[email protected]

Basic Idea
Serial data transmission sends binary bits of track drifting txas a seriesprop delayor electrical identify bits information clock and of optical variation + pulses:
0111011000111110011010010000101011101100011111..

The transmission channel (coax, radio, ber) generally distorts the signal in various ways:

From this signal we must recover both clock and data 5

Eye diagram
Use a precise clock to chop the received data into periods Overlay each of the periods onto one plot symbol cell Y

jitter

Y amplitude distribution at Y-Y 7

Eye diagram construction

random data TX link synth trigger RX scope

Denition of Jitter
unit interval

-3T -2T -T 0 T 2T 3T 4T Impulses spaced equally in time (jitter free signal)

time

-3T -2T -T 0 T 2T 3T 4T Impulses spaced irregularly in time (jittered signal)

Errors treated as discrete samples of continuous time jitter

After Trischitta and Varma: Jitter in Digital Transmission Systems

20

Analytic Treatment of Jitter

Perfect Clock: x ( t ) = A cos c t

Jittered Clock: x ( t ) = A cos [ c t + ( t ) ]

( t ) is then treated as a continuous time signal

After Behzad Razavi: Monolithic Phase-Locked Loops, ISSCC96 Tutorial

29

Data Recovery with simple PLL

Jittered Data Signal D Q Retimed Data

Phase Detector PLL Low-pass Loop Filter Voltage Controlled Oscillator

28

Model of Loop
Phase Detector Loop Filter VCO

K 1 + -----------------( 1 + s )

Kv

1 -s

Warning: Extra Integration in loop makes for tricky design!

See Floyd M. Gardner, Phaselock Techniques, John Wiley and Sons, for good introduction to PLL theory

30

Loop Frequency Response a

(input data jitter)
80dB 40dB 0dB -40dB -80dB 1k 10k 100k 1M 10M 100M 1G 10G

1 + -----------------( 1 + s )

Kv -----s

c b

(vco phase noise)

open loop gain c/a c/b

31

Jitter Measurements

SONET has most complete set of jitter measurement standards, but the techniques are useful and relevant for datacom applications also.

Jitter Tolerance Jitter Transfer Jitter Generation

46

Jitter Tolerance Test Setup

laser transmitter data generator optical attenuator optical receiver xamp + limiter decision circuit retiming circuit bit error rate tester

FM modulated clock

sine wave generator

At each frequency, the sinewave modulation amplitude is increased until the BER penalty is equal to that caused by 1dB optical attentuation 47

After Trischitta and Varma: Jitter in Digital Transmission Systems

SONET Jitter Tolerance Mask

15 UI 1.5 UI 0.15 UI f0
Data Rate 155 Mb 622 Mb 2.488 Gb 10 Gb

acceptable range

f1
f0 [Hz] 10 10 10 ?

f2
f1 [Hz] 30 30 600 ?

f3
f2 [Hz] 300 300 6000 ?

ft
f3 [kHz] 6.5 25 100 400 ft [kHz] 65 250 1000 4000

from SONET SPEC: TA-NWT-000253 Issue 6, Sept. 1990, g 5-13

48

Jitter Transfer Measurement

data generator decision circuit clock retiming circuit D.U.T.

Phase detector

Signal Generator

Phase modulator

[TrV89] [RaO91] IN OUT network analyzer

49

Jitter Transfer Analysis

sideband clock amplitude

Assuming small angles (i.e.: only one dominant sideband on recovered clock):
ulta nt
V sideband Jitter pp ( rads ) = 2 2 atan ------------------------- V clock

res

Jitter transfer is dened as the jitter at the clock output divided by the jitter at the D.U.T input, plotted as a function of jitter frequency. ??

Jitter Transfer Specication

P[dB] acceptable range slope = -20 dB/decade

fc
Data Rate 155 Mb 622 Mb 2.488 Gb fc[kHz] 130 500 2000 P[dB] 0.1 0.1 0.1

This specication is intended to control jitter peaking in long repeater chains 50

Some Signal Degradation Mechanisms

Multiplex Jitter AC Coupling Optical Pulse Dispersion Skin Loss Random Noise E+O Crosstalk Intersymbol Interference

Multiplex Jitter
bit stufng events high speed data sub-rate data phase error [in UI] time Multiplex jitter is not a problem on the high rate channel itself - it only occurs on non-synchronous, lower speed tributaries that have been sent over the high-speed channel (e.g.: DS3 over SONET OC-48). 9

Voltage and Time aberrations caused by AC-coupling

t

V Vt t=0 t=t1

t1 V Vt AC coupled pulses droop as P --------------- 100 ------- 100 . V RC

Jitter is introduced by nite slope of pulse rise/fall time:

tr t1 t = --------------( 2RC )

10

Jitter Generation

decision circuit

computer

retiming circuit S.U.T.

spectrum analyzer

recovered clock

51

Jitter Generation (cont.)

1) Measure Jitter Sidebands around Clock clock amplitude
V sideband Jitter pp ( rads ) = 2 2 atan ------------------------- V clock

sideband

2) Multiply Jitter components by Filter Mask 3) RMS sum total noise voltages over band 4) Convert RMS noise voltage to RMS jitter

OC-48 (2.488 Gb/s SONET) species 12 kHz hipass lter, and maximum 0.01 UI RMS integrated jitter. 52

res

ulta

nt

Decision Circuit

Quantizes amplitude at precise sample instant Typically uses positive feedback to resolve small input signals A master/slave D-ip-op carefully optimized for input sensitivity and clock phase margin is a common choice Latches input data on rising edge of clock signal

simplied schematic symbol:

D Q clock 39

Code Disparity
Disparity is dened as Nhigh - Nlow in past transmitted signal

+5 0 -5

encoded data signal signal disparity

In an unbalanced code the disparity can grow without limit. e.g.: 4B5B code of FDDI In a balanced code, the disparity is limited to a nite worst case value. e.g.: 8B10B of FibreChannel 16

After Tom Hornak: Interface Electronics for Fiber Optic Computer Links, (see bibliography for full citation)

Filter Method Examples

d dt X
2

bandpass lter e.g.: SAW lter bandpass lter LC tank Recovered Clock Output

NRZ Data Input

non-linear element delay

(this last circuit can be thought of as an NRZ-RZ converter) 25

Spectrum of NRZ data

variations due to DC balance strategy

power in dB

sin ( 2fT ) -----------------------2 fT

missing clock frequency

f = 0

1 2T

1T

3 2T

2T

22

NRZ and RZ signalling

NRZ = non return to zero data

+ +

+ +

neither clock nor data frequency in spectrum

data frequency clock frequency

RZ = return to zero data

+ +

+ +

clock, but no data frequency in spectrum

T NRZ signalling is almost universally used. 21

A detailed look at the spectrum of differentiated NRZ

missing clock frequency power in dB
k

f = 0

fc 2

fc fc A k sin 2 ----- k k 2 k

23

Reconstructing the Clock

Start with symmetric sideband pairs about fc/2:

fc 2 ----- A k sin 2k k
k

Mix data signal with itself (e.g.: square law):

1 sin sin = -- cos ( ) cos ( + ) 2

All the symmetric sidebands mix pairwise to coherently create a carrier frequency component:

direct implementation of this principle is Filter Method 24

Example Bipolar Decision Circuit

master latch slave latch gnd data in data out clock in Vbias -5V

many clever optimizations are possible 40

Summary of Filter Method

Jittered NRZ Data Signal D Q bandpass lter/limiter Con: Retimed Data

d dt

Pro:

Very simple to implement Temperature and frequency variation of lter group delay makes sampling Can be built with time difcult to control microwave tinkertoys using coax to very high frequencies Narrow pulses imply high fT Hi-Q lter difcult to integrate 26

Example MOS Decision Circuit

master latch

data in clock

slave latch

clock

decision circuits can be implemented in NMOS technology up to Gb/s speeds 41

Decision Threshold Generation

To minimize bit-error rate, the decision threshold X-X must centered in the signal swing. Two common ways of automatically generating threshold voltage are:

Peak detection of signal extremes, limited run-length required positive peak detector negative peak detector Decision Threshold

Signal

Decision threshold = signal average, balanced signal required Low-Pass Filter Decision Threshold 15

Signal

After Tom Hornak: Interface Electronics for Fiber Optic Computer Links, (see bibliography for full citation)

Quantized Feedback
AC-Coupled Transmission Link

TX

H ()

RX

Feedback voltage models missing DC information

1 H ()

D Q clock

Output Data (models ideal TX waveform)

11

Phasor Diagram

Graph of relative phase between clock and data Each complete rotation is 1 unit interval of phase slip Rotations/second = frequency error (in Hz)
0

Plot of data transitions versus VCO clock phase.

90

270

180

Data at 1/2, or VCO at 2x, the proper frequency look locked. This puts a limit on VCO tolerance to prevent false locking.

= missing transitions = actual transitions 43

Example Lock Detector

0

[FP15.5]

ideal data eye

270

90

noisy data eye

180

clock a,b D data

Q D Q Raw out-of lock indication 44

Aided Acquistion

Tricky task due to Nyquist sampling constraints caused by stuttering data transitions PD Input Data VCO FD loop lter 2 loop lter 1

Still subject to false lock if VCO range is too wide 42

After Behzad Razavi: Monolithic Phase-Locked Loops, ISSCC96 Tutorial

Phase Detectors

Phase detectors generate a DC component proportional to deviation of the sampling point from center of bit-cell Phase detectors are:

Continuous
90 180 90 0 180

Binary Quantized

Binary quantized phase detectors are also called Bangbang phase detectors 32

After Tom Hornak: Interface Electronics for Fiber Optic Computer Links, (see bibliography for full citation)

Training Loops
retimed data bang-bang drive Input PDET charge pump Clock VCO Data

select

dlock dtrans ock Reference Clock 2.488GHz/256 FDET Clock/256

State Machine

LOS

1/256 divider

An increasingly common technique is to provide a reference clock to the CDR circuit. This allows the VCO process-variation to be dynamically trimmed out, avoiding false locking problems. (Figure from paper FP15.5, 1997 ISSCC) 45

Coding for Desirable Properties

DC balance, low disparity Bounded run length High Coding Efciency Spectral Properties (decrease HF and/or DC component)

Many Variations are Possible!

Manchester [San82] mB/nB [Gri69][Rou76][WiF83] [YKI84] [Pet88] Scrambling [CCI90] CIMT [WHY91] 17

Simple 3B/4B code example

4B Output Data 3B Input Data Even Words 000 001 010 011 100 101 110 111 SyncA SyncB 0100 0010 0111 1000 Odd Words 0011 0101 0110 1001 1010 1100 1011 1101 1110 0001

Maximum Runlength is 6 Coding Efciency is 4/3 Sending Sync Sequence: SyncA(even), SyncA(odd), SyncB(even), SyncB(odd) allows the unambiguous alignment of 4-bit frame

18

Scrambling

Uses a feedback shift register to randomize data reversing process at receiver restores original data Scrambled Data Data Output

Data Input

XOR Shift Register n j 2 1 Clk PRBS Generators Caveat: Only guarantees balance and runlength under very specic data conditions!
After Tom Hornak: Interface Electronics for Fiber Optic Computer Links, (see bibliography for full citation)

1 2

19

VCO alternatives
LC Oscillator Speed Phase Noise Integration Multivibrator Ring Oscillator

Technology Dependent 1-10s of GHz, CMOS 1-2 GHz Good Poor (L, Varactor) Narrow/Slow Good Poor Excellent

Tunability Stability

Wide/Fast Poor (needs acquisition aid) Multi-Phase Clocks

Other

[Cor79, Ena87, Wal89, DeV91,Lam93, WKG94]

After Todd Weigandt, B. Kim, P.Gray, Timing Jitter Analysis for High-Frequency CMOS Ring Oscillators, March 10, 1994

36

Multivibrator VCO
Capacitor is alternately charged and discharged by constant current Tuned by varying Itune in current source Diode clamps keep output voltage constant independent of frequency Relies on non-linear switching for oscillation behavior, and so is limited to moderate frequencies. Itune
I tune Frequency = ---------------4CV be

After Alan B. Grebene, Analog Integrated Circuit Design, Van Nostrand Reinhold, 1972, pp 313-315

37

Example Ring Oscillator VCO

Input 1 Output Tune Input 2 Input 1 Output [SyA86] [EnA87] [Wal89]

Inpu
Input 1 Tune Input 2 Output

t2

38

Loop Filters

may be analog (integrator) or digital (up-down counter) should have provision for holding value constant (tristating) under long run-length conditions [Den88] [Dev91] [LaW91] [WuW92] VOUT
UP 0 0 DOWN 0 1 0 1 VOUT tristate ramp DOWN ramp UP tristate

UP

DOWN

1 1

35

Skin Loss

Nearly all cables can be modeled by the Skin Loss Equation with various k factors: T ( f ) = 10 1.0 linear amplitude k=.001 k=.0001 k=.00001 R3 R2
( k ) f

Ring 3 Ring 2 Ring 1 L3 L2 R1

0.0

1k freq (log scale) 10G

[YFW82]

Three-element equivalent circuit of a conductor with skin loss 12

Skin Loss Equalization at Transmitter

boost the rst pulse after every transition Error at sampling pt.

usable signal See Paper: FP15.1, 1997 ISSCC before after 14

Skin Loss Equalization at Receiver

transmission (linear scale) 1.4 3dB boost 1.0 0.5 0.0
0 1

equalized pure skin loss 1k freq (log) 10G

2x improvement in maximum usable bit-rate

[WWS92] 13

Agenda

Overview of Serial Data Communications Signal Degradation Mechanisms Data Coding Techniques Clock Recovery Methods Components Used in Clock/Data Recovery Jitter Measurements

Diversity of CDR applications

Clock and Data Recovery (CDR) applications span the range from ultra-high-volume, low cost datacom applications to very high precision, long-haul telecom applications Many different trade-offs are made to tailor each circuit to the target application area

1.25Gb/s Gigabit Ethernet Transceiver <$10 in volume (datacom application)

1cm

2.488Gb/s SONET CDR ~$500 (telecom application) 3

Q-Factor in resonant circuits

Voltage envelope of ringing circuit falls to 1/sqrt(e) in Q radians. 1.0 1.0/sqrt(e)

Q also equals the center frequency of a lter divided by the full-width of the resonance measured at the half power points: Fcenter/

amplitude

Q/2*PI cycles

Fcenter

High-Q lter can be emulated by PLL with low loop B.W.

27

Bit Error Rate (BER) Testing

Pseudo-Random-Bit-Sequence (PRBS) is used to simulate random data for transmission across the link PRBS pattern 2N-1 Bits long contains all N-bit patterns Number of errored-bits divided by total bits = BER. Typical links are designed for BERs better than 10-12

PRBS data generator clock in

TX link synth

RX

PRBS data receiver clock in

Drawbacks of Simple PLL

1) timing pulses 2) transfer function (linear vs BB), 3) quadratic, BB 4) critical problem is the stuttering data

??

Self-Correcting Phase Detector

UP Data Clock
1 = Data................. 2= Clock (Early)..... 3 = 1 retimed.......... 4 = Clock................. 5 = 3 retimed.......... 6 = 1 xor 3 (UP)..... 7 = 3xor 5 (DOWN)

D Q

D Q

DOWN Data

See FA9.6, 1997 ISSCC [Hog85][Shi87]

33

Binary Quantized Phase Detector

NRZ data is sampled at each bit cell and near the transitions of each bit cell The sign of the transition sample is compared with the preceeding and following bit cell sample to deduce the phase error B A T B Output DQ DQ A
0 0 0 0 1 1 1 1 0 0 1 1 0 0 1 1 0 1 0 1 0 1 0 1 tristate vco fast ? vco slow vco slow ? vco fast tristate

Data

DQ Clock

DQ

T A T B

[Ale75][WHY91] [LaW91][ReG73]

34

CDR Design Checklist

1) Eye Margin

how much noise can be added to the input signal while maintaining target BER? (voltage margin) How far can clock phase alignment be varied while maintaining target BER? phase margin) how much does the static phase error vary versus frequency, temperature and process variation? Is input amplier gain, noise and offset sufcient?

CDR Design Checklist (cont)

2) Jitter Characteristics

what is the jitter generation? (VCO phase noise, etc) what is the jitter transfer function? (peaking and bandwidth)what is the jitter tracking tolerance versus frequency?

CDR Design Checklist (cont)

3) Pattern Dependency

how do long runlengths affect system performance? is bandwidth sufcient for individual isolated bit pulses? are there other problematic data patterns? (resonances)

does PLL bandwidth, jitter, and stability change versus transition density? 4) Acquisition Time what is the initial, power-on lock time? what is the phase-lock aquisition time when input source is changed?

CDR Design Checklist (cont)

5) How is precision achieved?

are laser-trimming or highly precise IC processes required? 6) Input/output impedance

are external capacitors, inductors needed? does the CDR need an external reference frequency?

Is S11/S22 (input/output impedance) maintained across the frequency band? are reection large enough to lead to eye closure and pattern dependency? is t >15 dB return loss maintained across the band?

CDR Design Checklist (cont)

7) Power Supply

does the CDR create power supply noise? how sensitive is the CDR to supply noise? Is the VCO self-modulated through its own supply noise? (can be deadly) what is the total static power dissipation? what is the die temperature under worse case conditions?

CDR Design Checklist (cont)

8) False lock susceptibility

can false lock occur? are false lock conditions be detected and eliminated? can the VCO run faster than the phase/frequency detector can operate? (another killer) have all latchup/deadly embrace conditions been considered and eliminated?