Digital Clock and Data

Recovery for High-Speed

Serial Data Links

Ken Evans
[email protected]

Digital Clock and Data Recovery for High-Speed Serial Data Links 1
 Typical 6 Gbps SATA / SAS PMA Layer
•  Adaptive RX equalization
at 6.0 Gbps.
•  Digital clock and SS-
modulated data recovery.
•  1:10 ser/des to
accomodate 8b/10b
coding in PCS layer.
•  SSCG MPLL serving
shared RX/TX frequency
synthesizer.
•  1b transmitter de-
emphasis, programmable
to -6 dB, -3.5 dB
•  Link scaling for legacy
modes (1.5, 3.0 Gbps).

Digital Clock and Data Recovery for High-Speed Serial Data Links 2
 Typical 5 Gbps PCI-E or USB PMA Layer
•  Adaptive RX equalization
for 5.0 Gbps USB, none
for PCI-E.
•  Digital clock and SS-
modulated data recovery.
•  1:10 ser/des to
accomodate 8b/10b
coding in PCS layer.
•  SSCG MPLL serving
shared RX/TX frequency
synthesizer.
•  1b transmitter de-
emphasis, programmable
to -6 dB, -3.5 dB
•  Link scaling for legacy
PCI-E (2.5 Gbps).

Digital Clock and Data Recovery for High-Speed Serial Data Links 3
 Combining Macros into Multilane SERDES
ASIC

Digital Clock and Data Recovery for High-Speed Serial Data Links 4
 Notes on SERDES IP Macros
–  Multiple applications / standards can be satisfied
with universal macros – PMAs are almost identical.
–  Signaling rates can be dialed in by scalable link
design.
–  PCS layer functions tend to be universal, i.e.
scrambling, 8b/10b coding, BIST, loopback modes.
–  RX / TX equalization optional depending on
channel medium.
•  SATA 3.0, USB 3.0 require RX EQ, TX de-emphasis.
•  PCI-E 2.0 does not require RX EQ up to 34” FR4.
–  Mostly digital implementations mandatory at this
point.
•  Multi-lane solutions demand minimal thermal / area footprint.
•  Quick process migrations require maximum portability.
•  Macros must be digitally programmable to accommodate
specific application specs and usage customizability.

Digital Clock and Data Recovery for High-Speed Serial Data Links 5
 Relevant CDR Blocks
•  Digital PLL (DPLL)
recovers digital phase
error to drive loop into
lock.
•  Varying degrees of data
deserialization
embedded within CDR.
•  Digital-to-phase
conversion bridges digital
and analog domains.
•  Frequency synthesizer
required to produce clock
which is ideally matched
in frequency with RX
data.
•  Loop dynamics are
governed by both linear
and non-linear behavior.

Digital Clock and Data Recovery for High-Speed Serial Data Links 6
 Basic DPLL Structure
Block-level Arrangement
•  DPLL phase adjustment
accomplished by rotating
the phase of clock
produced by VCO.
•  DPLL frequency
adjustment accomplished
by adjusting rate of
phase rotation.
•  DPC is the DPLL’s
“VCO”, but does not
System Representation integrate.
•  Simplistically, phase
error sampler is viewed
as a slicer. All remaining
elements are linear.
•  Non-linear loop dynamics
difficult to analyze in
closed form….

Digital Clock and Data Recovery for High-Speed Serial Data Links 7
 Basic !!PD Mechanics
•  !!PD 2x oversamples
data by 00(I) and 900(Q)
clock phases.
•  In locked state, I clock
samples data at the eye
center and Q clock
samples zero-crossings.
•  Early and late phase
errors produced by
XORing sliced data and
edge samples.
•  No error produced when
no transition present –
important feature!
•  Early, late, and no-
transition states are all
mutually exclusive.
•  !!PD never produces
stable output – causes
“dither” jitter at lock point.

Digital Clock and Data Recovery for High-Speed Serial Data Links 8
 !!PD Behavior in Presence of Jitter
•  Clock and data jitter
causes !!PD to have a
“stochastically” linear
range.
•  At any given phase error,
there are odds “P” that
the edge clock will
sample Dk, and “1-P” that
it will sample Dk-1.
•  “Linear” range of !!PD is
the probability
distribution of edge
samples over phase
error.
•  K!!pd is the slope of the
distribution at zero phase
error.
•  Jitter is a combination of
rj and dj components.

Digital Clock and Data Recovery for High-Speed Serial Data Links 9
 !!PD Behavior in Presence of Jitter (cont’d)

•  Random jitter (rj) used

for simulation – gaussian
probability distribution.
•  Extremes of phase error
saturate at +/- DT, the RX
data transition density,
because !!PD produces
no error when no
transition is present.
•  !!PD gain at lock, K!!pd, is
given as 1 / [σj(2π)1/2].
•  “Linear” range is
contained within +/- 2σj.

Digital Clock and Data Recovery for High-Speed Serial Data Links 10
 Three PRBS-31 RX Test Subjects

•  PRBS-31 patterns generated with 9b oversampling (512 samples per UI).

•  Sequence processed by raised cosine filter to mimic dispersive channel
loss.
•  White gaussian random jitter applied to waveforms: σrj = 1%, 3%, 5% UI.
•  Primary sequence length for simulation is 216 symbols.

Digital Clock and Data Recovery for High-Speed Serial Data Links 11
 Simulated !!PD Transfer Curves

•  Higher jitter lowers CDR

loop gain – has
consequences for
sinusoidal jitter tolerance
(JT) compliance.
•  Higher jitter broadens the
range over the data eye
that the CDR will behave
as a linear feedback
loop.
•  Lower jitter causes CDR
to behave as a non-linear
slewing feedback loop.

Digital Clock and Data Recovery for High-Speed Serial Data Links 12
 Effect of Sampler Offsets

•  Small random and systematic sampler offsets have a manageable effect on PD

behavior (left plot).
•  Larger offsets can cause skews in in the lock point as well as make the CDR’s loop
gain erratic and unpredictable (right plot).
•  Use circuit design techniques to minimize random offsets, and robust layout
techniques to minimize systematic offsets. Digital offset calibration is also
recommended.

Digital Clock and Data Recovery for High-Speed Serial Data Links 13
 Typical Error Generator Implementation

•  CDR loop cannot process error symbols at baud rate – digital filter cannot move that
fast.
•  Half-rate I/Q clocks from DPCs perform 1:2 data and edge deserialization.
•  Symbols are aligned and deserialized 2:10, and half-rate clock is divided by 5.
•  Array of unit !!PDs are used to generate 10b error sequence, which is eventually
decimated into a signed, 2b error signal moving at 1/10th the baud rate.

Digital Clock and Data Recovery for High-Speed Serial Data Links 14
 Sampling and Deserialization

Digital Clock and Data Recovery for High-Speed Serial Data Links 15
 2T Sampling and Alignment Timing

Digital Clock and Data Recovery for High-Speed Serial Data Links 16
 Deserialization Timing (Data)

Digital Clock and Data Recovery for High-Speed Serial Data Links 17
 Deserialization Timing (Edge)

Digital Clock and Data Recovery for High-Speed Serial Data Links 18
 10b !!PD Implementation

•  Unit !!PD cells generate

10 sets of 3b hot-coded
errors.
•  The 10b !!PD needs to
“look ahead” into the next
data frame to compute
the 10th 3b error code. A
redundant “wrap” bit,
shared between two
adjacent data frames, is
made available for this
purpose.
•  10 x 3b error signal is
retimed prior to
decimation.

Digital Clock and Data Recovery for High-Speed Serial Data Links 19
 10b !!PD Timing

Digital Clock and Data Recovery for High-Speed Serial Data Links 20
 Error Decimation by Majority Voting
•  Majority voting used to
summarize 5b error
signal.
•  Democratic process is
inherently non-linear,
must be simulated to
determine effect on
DPLL loop gain.
•  MUX matrix maintains a
“tally” of early, late, and
no-transition bits within
the 5b error word.
•  Might require inter-stage
retiming depending on
MUX implementation.

Digital Clock and Data Recovery for High-Speed Serial Data Links 21
 Majority Voting Simulation

•  Transfer curve simulation of !!PD in combination with MV yields KMV = 3.0.

•  Good numerical agreement over three jitter values, although higher jitter simulations
produce a noisier curve.

Digital Clock and Data Recovery for High-Speed Serial Data Links 22
 DPLL Loop Filter Design

Digital Clock and Data Recovery for High-Speed Serial Data Links 23
 DPLL Loop Filter Design (cont’d 1)
•  2nd order filter provides both phase and frequency integration.
•  Phase integration path accumulates phase differences between
sampling clocks and the RX data stream.
•  Frequency integration path accumulates frequency differences
between analog VCO and RX data stream. Allows DPLL to track
static errors due to host and receiver crystal frequency mismatches
(+/-300 ppm), and time-varying errors due to spread-spectrum clock
modulation (0 to -5000 ppm).
•  Gain implemented by bit shifting, i.e.
•  (+1) x 22 is ’01’  ‘0100’ (+4)
•  (-1) x 22 is ’11’  ’1100’ (-4)
•  Attenuation implemented by truncating dither bits, i.e.
•  (+4) x 2-2 is ’0100’  ’01’ (+1)
•  (-4) x 2-2 is ‘1100’  ’11’ (-1)

Digital Clock and Data Recovery for High-Speed Serial Data Links 24
 DPLL Loop Filter Design (cont’d 2)
•  Phase path has gain phug = 2{3,4,5,6} x 2-6.
•  Frequency path has gain frug = 2{0,1,2,3} x 2-12.
•  Frequency integrator requirements:
•  Must accumulate frequency error until integrator saturates: adder must never
overflow or underflow.
•  Frequency error can be either positive or negative polarity: adder must perform
signed (2’s complement) arithmetic.
•  Phase integrator requirements:
•  DPC has full-scale rotation range of 3600: adder must overflow and/or underflow
to rotate DPC beyond this range.
•  DPC maps unsigned binary code to phase value, adder must be unsigned as
well.

Digital Clock and Data Recovery for High-Speed Serial Data Links 25
 DPLL Frequency Tracking
•  RX data stream is generally SS-
modulated by host-side TX.
•  RX clock is asynchronously SS-
modulated by MPLL unit.
•  Worst-case frequency difference
occurs when modulations are 1800
apart – 5000 ppm.
•  Including static frequency
differences, DPLL must elastically
track up to +300 ppm to -5300 ppm
over time.
•  Frequency integrator must have
sufficient bits to accommodate, with
integrator loss (2-6), decimation
factor (10), and DPC resolution (2-9).
•  Frequency resolution is the ppm
value that the DPLL can change
recovered clock frequency per
update time.

Digital Clock and Data Recovery for High-Speed Serial Data Links 26
 DPLL Frequency Tracking (cont’d)
•  DPLL step response
simulations performed with
and without static frequency
offset.
•  9b DPC resolution yields
512 samples per UI.
•  514 samples are allocated
between each retiming
interval.
•  Toffset = 106 x (514/512 – 1)
= 3900 ppm.
•  Slope of accumulation
proportional to amount of
offset tracked by DPLL.
•  Slope of accumulation will
time-vary according to RX
data spread-spectrum
modulation.

Digital Clock and Data Recovery for High-Speed Serial Data Links 27
 CDR Small-Signal Model (z-1)

•  Model is valid when DPLL is in lock:

•  Variations in ϕdata are contained within [ -2σj, +2σj ].
•  Variations in ϕdata, fdata are moving at a rate within the CDR’s tracking bandwidth
•  !!PD noise is always present, and can be input referred with value σ!!pd / K!!pd.

Digital Clock and Data Recovery for High-Speed Serial Data Links 28
 CDR Small-Signal Model (s)

•  For operating frequencies much lower than the update rate of the DPLL, Euler
difference equation can be used to transform z-domain to s-domain.
•  Loop gain L(s) has the same form as that of a typical charge-pump APLL, but with an
added exponential term that models the digital latency of the DPLL.
•  In the simplistic case where latency is omitted, the DPLL tracking response has
identical form with that of the APLL counterpart.
•  Loop bandwidth and transfer peaking can both be designed through this
transformation.

Digital Clock and Data Recovery for High-Speed Serial Data Links 29
 CDR Frequency Response

•  DPLL bandwith and peaking response is sensitive to CDR update rate and jitter.
•  For fixed σj, Tupd, bandwidth can be extended and peaking suppressed by increasing
phase update gain (phug).
•  Serial data link jitter should be empirically observed, and phug and frug values should
then be chosen to yield desired bandwidth and peaking.

Digital Clock and Data Recovery for High-Speed Serial Data Links 30
 CDR Frequency Response (cont’d)
σj = 0.01UI σj = 0.03UI σj = 0.05UI
Tupd = 2ns Tupd = 2ns Tupd = 2ns
phug frug BW peaking phug frug BW peaking phug frug BW peaking
(MHz) (dB) (MHz) (dB) (MHz) (dB)

1 / 8 1 / 4096 3.012 0.4267 1 / 8 1 / 4096 1.115 1.0568 1 / 8 1 / 4096 0.736 1.5644

1 / 8 1 / 2048 3.195 0.7674 1 / 8 1 / 2048 1.281 1.7968 1 / 8 1 / 2048 0.886 2.5744
1 / 8 1 / 1024 3.373 1.0672 1 / 8 1 / 1024 1.432 2.4037 1 / 8 1 / 1024 1.017 3.3691
1 / 8 1 / 512 3.545 1.3393 1 / 8 1 / 512 1.569 2.9265 1 / 8 1 / 512 1.133 4.0333

1 / 4 1 / 4096 5.832 0.1234 1 / 4 1 / 4096 1.968 0.3312 1 / 4 1 / 4096 1.213 0.5144

1 / 4 1 / 2048 5.925 0.2325 1 / 4 1 / 2048 2.06 0.6016 1 / 4 1 / 2048 1.304 0.9142
1 / 4 1 / 1024 6.018 0.3339 1 / 4 1 / 1024 2.151 0.8421 1 / 4 1 / 1024 1.391 0.4029
1 / 4 1 / 512 6.111 0.4298 1 / 4 1 / 512 2.239 1.062 1 / 4 1 / 512 1.474 0.5161

1 / 2 1 / 4096 11.903 0.0334 1 / 2 1 / 4096 3.833 0.0943 1 / 2 1 / 4096 2.301 0.151

1 / 2 1 / 2048 11.95 0.0648 1 / 2 1 / 2048 3.88 0.1787 1 / 2 1 / 2048 2.347 0.2822
1 / 2 1 / 1024 11.996 0.095 1 / 2 1 / 1024 3.927 0.2576 1 / 2 1 / 1024 2.394 0.4029
1 / 2 1 / 512 12.043 0.1242 1 / 2 1 / 512 3.973 0.3326 1 / 2 1 / 512 2.44 0.5161

1 1 / 4096 25.539 0.0087 1 1 / 4096 7.756 0.0253 1 1 / 4096 4.586 0.0413

1 1 / 2048 25.562 0.0172 1 1 / 2048 7.78 0.0492 1 1 / 2048 4.609 0.0795
1 1 / 1024 25.585 0.0255 1 1 / 1024 7.803 0.0722 1 1 / 1024 4.633 0.1161
1 1 / 512 25.608 0.0337 1 1 / 512 7.826 0.0946 1 1 / 512 4.656 0.1514

Digital Clock and Data Recovery for High-Speed Serial Data Links 31
 Effect of Loop Latency
σrj = 0.03UI σrj = 0.03UI
phug = 2-3 phug = 1
frug = 2-12 frug = 2-12

•  Low loop bandwidth settings are relatively insensitive to increases in digital latency.
•  High loop bandwidth settings see a severe degradation in phase margin.
•  Bandwidth and peaking values shift – be careful!
•  Lower phase margin causes the loop to have higher quality factor – bad for jitter transfer.
•  Limit cycle jitter worsens due to lengthened phase correction time.

Digital Clock and Data Recovery for High-Speed Serial Data Links 32
 Jitterless CDR Behavior
•  A jitterless CDR does not
obey linear loop dynamics
at all.
•  A 5b step in ϕdata causes
the DPLL to slew towards
the lock point.
•  The lock point is
characterized by a 1 lsb
prbs dithering sequence,
which is known as “limit
cycle” behavior.
•  Dithering is caused by the
full-scale meta-stable
output of the !!PD, and
characterized by σ!!pd.
•  Dither magnitude will grow
beyond 1 lsb if loop latency
increases – another reason
to retime sparingly.

Digital Clock and Data Recovery for High-Speed Serial Data Links 33
 Jittered CDR Behavior

•  4b step applied to ϕdata at CDR input.

•  Low jitter condition shows response is dominated by non-linear slewing. Very close
to the lock point, the phase error falls within the linear range of the !!PD and the
response tracks the s-domain model.
•  High jitter condition shows response closely tracks the s-domain model because
majority of input step is contained within the linear range of the !!PD.
•  Important observation: phase / frequency tracking capability is band-limited.

Digital Clock and Data Recovery for High-Speed Serial Data Links 34
 CDR Jitter Tolerance
•  Jitter Tolerance (JT) testing
is a frequently-used
SERDES compliance test.
•  Measure of sinusoidal jitter
tracking capability.
•  Compliant CDRs do not
penetrate the mask
definition.
•  Mask typically has a corner
frequency of rb / 1667.
•  Low-jitter environments can
meet the mask specification
over a variety of loop
settings, as shown.

Digital Clock and Data Recovery for High-Speed Serial Data Links 35
 CDR Jitter Tolerance (cont’d)

•  As jitter increases, the lower bandwidth settings start approaching the mask
boundary. For high jitter conditions, some settings cause mask violations.
•  For a given level of RX jitter, the DPLL requires relatively low bandwidth settings to
maintain good phase margin and jitter transfer performance. But Jitter tolerance
requires higher bandwidth settings….
•  Conclusion: jitter transfer and jitter tolerance are conflicting design goals and must be
balanced.

Digital Clock and Data Recovery for High-Speed Serial Data Links 36
 References

1.  J. Lee et al., “Analysis and Modeling of Bang-Bang Clock and Data Recovery
Circuits”, IEEE Journal of Solid-State Circuits, Vol. 39, No. 9, September 2004, pp.
1571-1580.
2.  J. Sonntag et al., “A Digital Clock and Data Recovery Architecture for Multi-Gigabit/s
Binary Links”, IEEE Journal of Solid-State Circuits, Vol. 41, No. 8, August 2006, pp.
1867-1875.
3.  P. K. Hanumolu, et al., “A Wide-Tracking Range Clock and Data Recovery Circuit”,
IEEE Journal of Solid-State Circuits, Vol. 43, No. 2, February 2008, pp. 425-439.

Digital Clock and Data Recovery for High-Speed Serial Data Links 37