IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 42, NO.

9, SEPTEMBER 2007 1999

Equalization and Clock and Data Recovery

Techniques for 10-Gb/s CMOS Serial-Link Receivers
Srikanth Gondi and Behzad Razavi, Fellow, IEEE

Abstract—Two equalizer filter topologies and a merged equal- with other equalization methods [6] so as to mitigate both
izer/CDR circuit are described that operate at 10 Gb/s in 0.13- m effects.
CMOS technology. Using techniques such as reverse scaling, pas- Section II presents various gain peaking methods that are
sive peaking networks, and dual- and triple-loop adaptation, the
prototypes adapt to FR4 trace lengths up to 24 inches. The equal- used in Section III to develop two equalizer filter topologies.
izer/CDR circuit retimes the data with a bit error rate of 10 13 Section IV describes an adaptive equalizer architecture and
while consuming 133 mW from a 1.6-V supply. Section V the merged equalizer/CDR circuit. Section VI sum-
Index Terms—Adaptive equalization, analog equalization, marizes the experimental results for the two prototypes.
broadband receivers, DFE, FFE, high-speed links, lossy channel,
reverse scaling. II. GAIN PEAKING TECHNIQUES

A. General Considerations
I. INTRODUCTION
Copper traces designed as transmission lines on FR4 sub-
strates suffer from both skin effect and dielectric loss. For ex-

T HE frequency-dependent loss of traces on FR4 printed cir-

cuit boards (PCBs) poses increasingly more difficult de-
sign challenges as data rates approach 10 Gb/s and trace lengths
ample, a 30-in trace exhibits a loss of approximately 21 dB at
5 GHz and 34 dB at 10 GHz (Appendix I). The equalizer filter
must therefore provide adequate gain peaking around 5 GHz so
reach tens of inches. Preemphasis in the transmitter can partially as to equalize the signal spectrum.
compensate for the channel loss but at the cost of dynamic range. The design of gain-peaking circuits must satisfy many
For example, the 12 dB preemphasis in [1] requires a supply difficult requirements: 1) sufficient boost at high frequencies;
voltage of 2.5 V to accommodate the large amplified compo- 2) matching the inverse loss profile of the channel with reason-
nents without sacrificing the low-frequency swings. Thus, low able tolerance; 3) minimal low-frequency loss to minimize the
supply voltages and channel losses as high as 25 dB dictate that noise accumulation in cascaded stages and provide sufficient
most of the equalization be performed in the receiver. Examples swings for the CDR; 4) well-behaved phase response to achieve
include a bipolar implementation consuming 195 mW [2] and a a low jitter; 5) reasonable linearity so that the equalizer transfer
CMOS realization suffering from high intersymbol interference function indeed acts as the inverse of the channel loss profile;
(ISI) and lacking automatic adaptation [3]. 6) small input capacitance to satisfy the requirements; and
Recent work [1], [4], [5] has incorporated nonlinear (e.g., de- 7) tunability of the boost to allow adaptation.
cision-feedback) equalizers so as to accommodate sharp notches In addition, the challenges typically encountered in the design
in the channel due to impedance discontinuities and also avoid of low-voltage broadband amplifiers–such as limited bandwidth
the amplification of crosstalk due to high-frequency peaking. and drive capability–persist here as well.
The complexity and power dissipation of such realizations (e.g.,
210 mW for the 6.25-Gb/s 0.13- m CMOS receiver in [5]) are B. Peaking By Complex Poles or Real Zeros
justified for only demanding applications.
This paper describes adaptive equalization and clock and The availability of monolithic inductors may suggest the use
data recovery (CDR) techniques suited to 10-Gb/s binary of underdamped complex poles to provide the required boost
receivers. The concepts introduced here address the problem at high frequencies. For example, the shunt-peaking circuit of
of high losses. Loss compensation is achieved by using linear Fig. 1(a) yields a transfer function of the form
equalization techniques that tend to amplify crosstalk noise. If
crosstalk and impedance discontinuities due to connectors and (1)
vias also become critical, these techniques can be combined
where , ,
Manuscript received July 26, 2006; revised February 26, 2007. This work . While providing enough boost to match
was supported by Kawasaki Microelectronics America. Fabrication support was
provided by Taiwan Semiconductor Manufacturing Company.
the inverse of the FR4 frequency response, high- complex
S. Gondi was with the Electrical Engineering Department, University of Cal- poles introduce substantial phase distortion. As an example, a
ifornia, Los Angeles, CA 90095 USA. He is now with Kawasaki Microelec- cascade of two such stages is designed to equalize a 30-in trace
tronics, San Jose, CA 95131 USA. [Fig. 1(b)], yielding the equalized eye shown in Fig. 1(d).
The authors are with the Electrical Engineering Department, University of
California, Los Angeles, CA 90095 USA (e-mail: [email protected]). This issue restricts realizations to non-feedback structures
Digital Object Identifier 10.1109/JSSC.2007.903076 containing real zeros and poles or feedback structures having
0018-9200/$25.00 © 2007 IEEE
2000 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 42, NO. 9, SEPTEMBER 2007

Fig. 1. Complex pole peaking circuit. (a) Implementation. (b) Magnitude response. (c) Phase response. (d) transient response.

concluding that the product of the gain, the boost factor, and the
bandwidth of the stage is limited by the of the technology.1
It is important to appreciate the impact of the limited
(about 75 GHz in 0.13- m CMOS technology) on equaliza-
tion. For a small-signal gain of 2, an undegenerated differen-
tial pair with an overdrive voltage of 300 mV and fanout of
unity yields a bandwidth of less than 12.5 GHz. The degener-
ated structure trades this gain-bandwidth product for the boost
factor, the low-frequency gain, and the linear range.
With the performance envelope imposed by (2), a cascade
of stages such as the circuit of Fig. 2(a) fails to provide the
required boost factor while accommodating a data rate of 10
Gb/s and a reasonable low-frequency gain (e.g., 6 dB–0 dB).
Fig. 2. RC-degenerated differential pair. (a) Circuit implementation. (b) Fre- As the number of stages in the cascade increases to achieve a
quency response: actual response (solid line); bode approximation (dashed line).
higher boost factor, the overall bandwidth tends to drop unless a
greater low-frequency loss is allowed in each stage. Simulations
indicate that, for a total boost factor of 22 dB at 5 GHz and an
low- complex poles. Of course, inductors can still act as shunt- overall bandwidth of 5.5 GHz,2 three degenerated stages with a
peaking elements to broaden the bandwidth of equalizer stages. low-frequency loss of 6 dB per stage are required. Such a high
An efficient method of boosting by means of real loss results in a sensitivity degradation of about 3 dB.
zeros is capacitive degeneration. Fig. 2(a) shows a de-
generated differential pair that yields a zero at C. Peaking By Passive Networks
and poles at and We propose the use of passive high-pass networks to pro-
, with a low-frequency gain vide boost at the front end of an equalizer, thus relaxing the
of . Fig. 2(b) depicts linearity and gain peaking of the subsequent active stages and
the frequency response. Improving the linearity of the stage, saving power consumption. Fig. 3(a) depicts an example where
degeneration nonetheless creates a trade-off between the the zero and pole frequencies are respectively given by
low-frequency gain and the boost factor, . Interestingly, and and the boost factor by
one can write if is neglected. To avoid degrading the input
1Here, C includes the input capacitance of the next stage.
2The overall bandwidth refers to that of the FR4 trace along with the equalizer
(2) circuit.
GONDI AND RAZAVI: EQUALIZATION AND CLOCK AND DATA RECOVERY TECHNIQUES FOR 10-GB/S CMOS SERIAL-LINK RECEIVERS 2001

Fig. 3. (a) Passive network. (b) Proposed passive network with series-inductive peaking.

Fig. 4. Illustration of the reverse scaling technique. (a) Block diagram. (b) Single-ended circuit realization.

return loss, , , and must be chosen high enough

so that . Also, since the input capacitance, ,
lowers the pole frequency to , we re-
quire that . For example, if ,
(for a boost of 4.7 dB), and fF, then must remain
below 20 fF, restricting the size of the transistors in the first ac-
tive stage.
The above restriction can be eased through the use of series
inductive peaking [Fig. 3(b)]. Here provides additional
peaking at frequencies above 5 GHz even for large values of Fig. 5. Transistor-level simulations of a five-stage equalizer with (a) no scaling,
. For example, if , , fF, and (b) reverse scaling.
and fF, the network yields a boost factor of 8 dB and
an of dB at 10 GHz.
boost characteristics are maintained but the time constant at the
D. Reverse Scaling output node of each stage is reduced. The total capacitance seen
at the interface between stages and can be expressed
With the trade-offs described above, the peaking circuits pre-
as , where includes the junc-
sented thus far provide a boost of approximately 8 dB at 5 GHz,
tion and overlap capacitances at the output of the th stage and
dictating the use of at least three peaking stages in the equalizer.
the input capacitance of the th stage. In a re-
Moreover, the cumulative low-frequency loss requires two ad-
verse-scaled design [Fig. 4(b)], the corresponding time constant
ditional gain stages to restore the signal level.
is given by
If identical stages are cascaded, the overall small-signal
bandwidth is given by [7]
(4)
(3)
revealing a bandwidth improvement resulting from the second
where denotes the bandwidth of a single stage and is term in the parentheses. For example, with fF,
equal to 2 for first-order stages and 4 for second-order stages. fF, and , the bandwidth increases by about
Thus, if , then the small-signal bandwidth drops by a 33%.
factor of 2.6 for five stages. For example, with three peaking As evident from the relationship , the value of
and two gain stages, the bandwidth of the equalizer falls below is dictated primarily by three factors: 1) the maximum toler-
4 GHz if no bandwidth enhancement techniques are employed. able value of ; 2) the minimum acceptable value of (the
Reverse scaling [8] provides bandwidth improvement in ap- input capacitance of the CDR circuit); and 3) the minimum ac-
plications where the input impedance need not be very high. In ceptable number of stages. The first component is determined
this work, we propose the concept of reverse scaling for equal- by the required input bandwidth and/or return loss. If the equal-
izer design [9]. As illustrated in Fig. 4, successive stages are izer path directly drives the CDR circuit, the minimum accept-
scaled down in size by a factor of such that the gain and able value of is given by the total input capacitance of the
2002 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 42, NO. 9, SEPTEMBER 2007

Fig. 6. Equalizer filter I with reverse scaling.

phase detector—a rather large value. Thus, a simple gain stage

preferably follows the equalizer. In our design, fF. As
mentioned above, the minimum number of stages is dictated by
the trade-off between the required peaking and the bandwidth.
For example, a five-stage cascade necessitates a bandwidth of
14 GHz at each node, limiting the total input capacitance to
450 fF. On the other hand, an of 10 dB at 10 GHz translates
to a total input capacitance of 213 fF. Allowing for ESD and
pad capacitance, we assume fF.3 Also, fF.
Thus, .
In this work, a scaling factor of 1.3 is chosen to achieve an
improvement of 22% in speed. Fig. 5 shows the simulated out- Fig. 7. Tuning behavior of five-stage cascade.
puts of the equalizer for scaled and unscaled designs.4 The eye
opening increases by approximately the same factor.5
Note that the use of monolithic T-coils [10] can greatly in- technique described in Section II but with slight variation in the
crease the tolerable input capacitance and hence provide addi- value of from one stage to the next to allow optimization for
tional bandwidth improvement with reverse scaling. high-frequency peaking and low-frequency loss.
As mentioned in Section II, simple resistively-loaded differ-
III. EQUALIZER FILTER DESIGN ential pairs cannot yield the required bandwidth. Thus, induc-
In this section, the peaking techniques presented in Section III tive peaking and negative Miller capacitances [11] have been
are used to design equalizer filters. These designs also incorpo- added to improve the speed without sacrificing the voltage head-
rate adaptive boosting so as to allow different trace lengths. room. To save area, only three of the stages incorporate induc-
tive peaking.
A. Equalizer Filter I The peaking stages in the equalizer path employ a variable
Shown in Fig. 6, the first equalizer [9] intersperses three degeneration resistance along with MOS varactors and
peaking stages with two gain stages to provide a boost factor to provide a wide boost range. As the control voltage rises, the
of about 22 dB at 5 GHz while exhibiting a low-frequency on-resistance of falls and so does the capacitance of
loss of less than 3 dB.6 The design exploits the reverse scaling and , raising the magnitude of the zero. Note that the simul-
3For the sake of simplicity, these calculations do not include the use of passive
taneous change of the resistance and capacitance greatly simpli-
peaking at the front end. fies the adaptation loop (Section IV). Fig. 7 illustrates the sim-
4The details of this circuit are shown in Fig. 6. ulated tuning behavior of the cascade as the control voltage is
5These designs incorporate other broadband techniques as well (Section III). swept from 0.1 V to 1.1 V.
6The two gain stages partially compensate the loss of the peaking stages, pro- As indicated in Fig. 6, the cascade employs capacitive
viding an output swing of approximately 1 V . coupling between some stages to isolate common-mode
GONDI AND RAZAVI: EQUALIZATION AND CLOCK AND DATA RECOVERY TECHNIQUES FOR 10-GB/S CMOS SERIAL-LINK RECEIVERS 2003

Fig. 8. Equalizer filter II architecture.

(CM) levels. This mitigates the voltage headroom issue and, range of 1–4 GHz. The pole provides additional adjustment of
more importantly, avoids variability in the CM level seen by the overall phase response in the all-pass path.7
- due to the preceding stage, thus maintaining a constant The use of series inductive peaking (with 3-nH inductors)
tuning range. The capacitors 0.25 pF are realized using in the passive boost stage allows wide input transistors (
multi-finger fringe structures having a parasitic component of m) in the first differential pair and hence reverse scaling
about 3%. The CM level is generated using a resistive divider. through the cascade. The low-frequency loss of 7 dB in the
The corner frequency associated with this capacitive coupling passive network degrades the sensitivity to some extent. With
is around 3 MHz, resulting in negligible droop with encoded inductive peaking and negative Miller capacitances, the band-
data. width of each active stage reaches 18 GHz. The summing stage
incorporates active feedback [10] to improve the speed while
avoiding inductors. Since active feedback trades low-frequency
B. Equalizer Filter II
gain for bandwidth, it has been applied to only one stage.
In Section II, the high-pass passive network was introduced to The choice between the two types of equalizers described
provide a peaking profile. The use of this network in the equal- above somewhat depends on the application. The former does
izer filter can save power consumption by relaxing the linearity not incorporate a passive network at the input, providing greater
and gain peaking requirements of the active stages. Fig. 8 shows sensitivity but consuming a higher power. The latter achieves a
the second equalizer filter architecture [12], which incorporates higher linearity and wider tuning range.
both passive peaking and reverse scaling. In contrast to the first
IV. ADAPTIVE EQUALIZER
equalizer, this design performs boost tuning by interpolation be-
tween a peaking path and an all-pass path (set by coefficients Fig. 9 shows the first adaptive equalizer architecture [9],
and ), thus achieving a wider tuning range than that ob- where the equalizer filter is followed by a slicer, and two loops
tained by means of MOS varactors and variable resistors. Also, a control the boost in the filter and the swing in the slicer. The
constant (linear) degeneration resistance in the differential pairs equalized data, , is sensed at node A rather than B because
yields higher linearity and a more accurately-defined low-fre- the slicer incorporates some peaking to improve convergence
quency gain. of the loops, thereby leading to larger jitter at B than at A when
The use of interpolation nonetheless presents a difficulty for the loops reach steady state. The need for the swing control
intermediate line lengths, i.e., if . The disparate de- loop is justified as follows.
lays through the two paths result in substantial ISI after their High-speed adaptive equalizers set the peaking in the filter
corresponding outputs are summed. Realized as a degenerated stages so as to compensate for the high-frequency loss of the
differential pair, the phase shift block in the all-pass path par- channel. To this end, the equalizer output is sharpened by a slicer
tially corrects this error. The zero of this differential pair is po- and the adaptation loop adjusts the peaking according to the
sitioned such that the phase response approximates the effective 7Simulations indicate that a 20% delay mismatch produces less than 5 ps of
response of the three zeros in the peaking path for the frequency ISI jitter at the equalizer output.
2004 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 42, NO. 9, SEPTEMBER 2007

Fig. 9. Proposed adaptive equalizer architecture.

Fig. 10. Spectrum before and after slicer with swing mismatch.

a function of the transmitted signal level and other parameters

in the signal path, whereas the slicer output is not; and 2) the
equalizer circuitry is fundamentally different from that in the
slicer. For example, the differential pairs in the equalizer may
not experience complete switching whereas those in the slicer
must.
Thus, as illustrated in Fig. 10, the adaptation may settle such
that the signals at A and B exhibit equal high frequency ener-
gies (between and ) while A is very poorly equalized. The
above difficulty can be alleviated by adding a loop for low-fre-
quency adaptation to maintain equal swings at A and B [15]. In
[15], however, the swings in the equalizer path are adjusted, po-
Fig. 11. Simulated dynamics of the swing and boost control loops. tentially limiting the tuning range and interfering with the main
adaptation loop. The approach introduced in this work is shown
in Fig. 9, where the input and output swings of the slicer are
difference between high-frequency contents of the data before
compared after rectification, and the resulting error is used to
and after slicing [2], [13], [14]. Note that the slicer generates a
adjust the slicer output swing rather than the equalizer output
random binary sequence having a spectrum. If operating
swing. The swing is controlled by adjusting the tail current of a
optimally, the equalizer must do the same.
limiting differential pair in the slicer. The slicer therefore gen-
The adaptation requires that the equalizer not slice (hard-
erates swings that match the equalizer output swings, allowing
limit) the data. Otherwise, the input and output of the final slicer
nearly complete overlap of spectra at A and B after equalization.
carry similar spectra, and the error in the adaptation loop ap-
Unlike implementations that precede the rectifiers with high-
proaches zero even with incomplete equalization. This issue
pass filters (HPFs) [13]–[15], this design employs bandpass fil-
in turn necessitates adequate linearity in the equalizer path,8
ters (BPFs) to measure the energy in a band [ ] around
leading to two important effects: 1) the equalizer output swing is
5 GHz. This is because the gain peaking occurs in the vicinity
8A variable-gain amplifier can precede the equalizer to ensure linearity. of this frequency and must be controlled accurately.
GONDI AND RAZAVI: EQUALIZATION AND CLOCK AND DATA RECOVERY TECHNIQUES FOR 10-GB/S CMOS SERIAL-LINK RECEIVERS 2005

Fig. 14. Simulated dynamics of (a) boost and swing control signals, and (b) the
Fig. 12. Response of the adaptive equalizer to a 25% mismatch in the line. control voltage of the VCO in the CDR loop.
(a) Frequency response. (b) Eye diagram.

Fig. 15. CDR architecture.

Fig. 13. Proposed merged adaptive equalizer/CDR architecture.

The two dynamic feedback loops in the architecture of Fig. 9

can potentially “fight” each other, thus failing to converge to ap-
propriate settings. Such a conflict is avoided by choosing sub-
stantially different time constants for the two loops, namely,
65 ns for swing control and 105 ns for boost control. Fig. 11
Fig. 16. VCO and buffer circuit.
depicts the simulated dynamics of the two loops as they settle
to their final values for worst-case initial conditions. (This sim-
ulation is performed on the transistor-level implementation of A. Architecture
the circuit.)
A possible point of concern is the sensitivity of the proposed Shown in Fig. 13 is the architecture of the merged equal-
adaptive equalizer to mismatches in the transmission line. izer/CDR circuit [12]. The retimed data, , provides both
Fig. 12 shows the response of the adaptive equalizer for a 25% low-frequency and high-frequency information for adjusting the
mismatch in the transmission line at 3.4 GHz. As expected, boost and swing controls. The swing adjust circuit is realized as
there is only a gradual degradation in the eye compared to a single differential pair while the filter and the energy compar-
the transmission line with minimal mismatch [Fig. 5(b)]. This ison mechanisms are similar to those in Figs. 8 and 9, respec-
design may be followed by a multi-tap decision-feedback tively.
equalizer to remove the resulting ISI. The architecture in Fig. 13 entails a start-up issue. With
high intersymbol interference (e.g., for a 24-in trace), if the
equalizer begins with minimum boost, then each data edge
V. MERGED ADAPTIVE EQUALIZER/CDR CIRCUIT applied to the CDR spans several bit periods, thus prohibiting
proper phase-lock. The architecture contains three feedback
In broadband receivers, the equalizer is typically followed loops whose time constants must be chosen carefully to ensure
by a CDR circuit. The two functions can simply be cascaded convergence. Fig. 14 shows the dynamics of the three loops as
but we recognize that the retimed data produced by the CDR they reach steady state. Only with proper phase lock can the
circuit exhibits the properties of a sliced waveform, obviating CDR provide the retimed data to the feedback loop; hence, the
the need for an explicit slicer. Since slicers typically consume CDR loop must settle before the boost control loop. Note that
several inductors and considerable power to achieve the required the boost control loop is initially reset, providing maximum
bandwidth and gain, the consolidation of equalizer and CDR peaking and enabling the CDR to lock. Also, even without
circuits can yield savings in area and power dissipation. phase-lock, the retimed data from the CDR is sufficient for
2006 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 42, NO. 9, SEPTEMBER 2007

Fig. 17. XOR and V /I circuit.

voltage to approximately , maximizing the capacitance

range of the MOS varactors and hence the tuning range of the os-
cillator. The buffer isolates the core from the large capacitances
associated with the PD devices and interconnects while sup-
pressing the data transitions that would otherwise couple from
the PD flip-flops to the VCO core.
The VCO phase noise is dominated by the modulation of the
varactor capacitances due to the noise current of . In this de-
sign, the simulated phase noise of the free-running VCO as pre-
dicted by SpectreRF is approximately equal to 106 dBc/Hz at
1-MHz offset.
The XOR gates used in the PD of Fig. 15 must exhibit sym-
metry with respect to their two inputs and operate with a low
Fig. 18. Die photograph of adaptive equalizer. supply voltage. Shown in Fig. 17 along with the / converter,
the XOR gate is a modified version of that in [17]. Here –
form the XOR core and copies the average output current
the swing control loop to converge. Thus, in this design, the into . To allow low-voltage operation, the drain voltage of
swing control loop is the fastest, and the boost control loop, the is raised by above its gate voltage, thus
slowest. saving one threshold in the headroom. The reference voltage
The merging of the equalizer and the CDR saves about is approximately equal to the common-mode level of A and B.
19 mW in power dissipation. Furthermore, it obviates several The / converter copies the average output current of the
inductors that would otherwise be necessary in the slicer. XOR, providing nearly rail-to-rail swings for the oscillator con-
The overall performance of the equalizer/CDR cascade de- trol line. Sensing the average voltage produced by the XOR, the
pends on the residual ISI, additive noise, and clock jitter. It is / converter remains free from a dead zone.
therefore necessary to quantify the trade-offs among these pa-
rameters so that a reasonable jitter budget can be allocated to VI. EXPERIMENTAL RESULTS
each stage. A framework for the analysis of these effects is pro- This section presents experimental results for the two circuits
posed in Appendix II. described in Sections IV and V. Both are fabricated in 0.13- m
CMOS technology.
B. CDR Circuit
This work employs a full-rate CDR circuit consisting of an A. Adaptive Equalizer I
Alexander phase detector (PD) [16] and an voltage-con- Fig. 18 shows a photograph of the die, which measures
trolled oscillator (VCO) (Fig. 15). While operating as a bang- 0.45 mm 0.36 mm. The circuit has been tested on a high-speed
bang circuit and exhibiting a high gain, the Alexander PD pro- probe station while sensing 10-Gb/s data that has traveled on
duces no output in the absence of data transitions, thus leaving an FR4 board. The pseudorandom bit sequence follows a 2 1
undisturbed. The high gain of the PD obviates a charge pattern. Fig. 19 shows the equalizer input and output waveforms
pump and permits the use of a simple voltage-to-current ( / ) at 10 Gb/s for 30-in and 6-in differential traces on the FR4
converter to drive the loop filter. Note that nodes X and Y need board without any external changes in the bias or other circuit
not provide a high bandwidth as only their average voltages are conditions. In this setup, the 30-in FR4 trace has a loss of 21 dB
sensed by the / converter—an important advantage of this re- at 5 GHz. The results show that the high-frequency adaptation
alization over those using charge pumps. loop accommodates varying loss conditions. To check the
The speed, jitter, and driving capability required of the oscil- operation of the low-frequency adaptation loop, the pattern
lator point to the use of an implementation. Fig. 16 depicts generator output swing was varied from 520–700 mV and
the VCO and its buffer. Resistor sets the core common-mode similar results were obtained. Note that the pattern generator
GONDI AND RAZAVI: EQUALIZATION AND CLOCK AND DATA RECOVERY TECHNIQUES FOR 10-GB/S CMOS SERIAL-LINK RECEIVERS 2007

Fig. 19. Measured results before and after equalization at 10 Gb/s for FR4 traces, (horizontal scale: 20 ps/div.): (a) before equalization for 30-in FR4 (vertical
scale: 100 mV/div.), (b) after equalization for 30-in FR4 (vertical scale: 100 mV/div.), (c) after equalization for 6-in FR4 (vertical scale: 100 mV/div.).

the equalizer. The circuit consumes 133 mW, of which 41 mW

is dissipated in the equalizer and 92 mW in the CDR.
Fig. 23(a) shows the recovered clock spectrum for a 24-in
FR4 at 10 Gb/s, displaying a phase noise of 109 dBc/Hz at
1-MHz offset. The jitter histogram in Fig. 23(b) suggests an rms
jitter of 2.22 ps. The VCO provides a tuning range from 8.9 GHz
to 11.6 GHz.

VII. CONCLUSION
The high loss of long traces on FR4 boards can be compen-
sated through the use of equalization techniques such as pas-
sive peaking networks, reverse scaling, and capacitive degener-
ation. To adapt to the line length, equalizers must incorporate
both boost and swing control while guaranteeing smooth, con-
flict-free convergence. Finally, the equalization and CDR func-
tions can be merged to eliminate slicers. This work has demon-
strated these concepts for a data rate of 10 Gb/s and trace lengths
Fig. 20. Die photograph of merged adaptive equalizer/CDR.
of 24 inches in 0.13- m CMOS technology.

itself suffers from a peak-to-peak jitter of 15 ps. The circuit APPENDIX I

(excluding the output buffer) consumes 25 mW from a 1.2-V LINE MODELING
supply. Fig. 24(a) plots the loss profile of a differential 30-in
The double trace visible in Fig. 19(c) for a 6-in line results microstrip line on an FR4 board. Designed for a differential
from approximately 2 dB of ripple in the cascade frequency re- characteristic impedance of 100 , the two traces have a width
sponse of the FR4 board and the equalizer. Confirmed by sim- of 6 mil and a spacing of 12 mil. (This profile is obtained by
ulations, this ripple possibly arises from the imperfect match simulation of the structure in SONNET.) The skin effect and
between the peaking profile provided by the equalizer and the dielectric loss exhibit approximately and dependencies,
inverse of the channel frequency response. respectively, and the loss profile can be represented as
B. Merged Equalizer/CDR Circuit
(5)
Fig. 20 shows the photograph of the die, which measures
0.94 mm 0.65 mm. The merged equalizer/CDR circuit uti- where and depend on the trace and board properties, re-
lizes the equalizer filter architecture proposed in Section III.B. spectively, and denotes the trace length. Thus, skin effect is
The circuit has been tested in a chip-on-board assembly with a dominant at lower frequencies and dielectric loss, at higher fre-
10-Gb/s 2 1 bit sequence. quencies. In this example, skin effect loss becomes equal to di-
Fig. 21 depicts the measured input and output eye diagrams electric loss at approximately 2.5 GHz. The strong frequency
for an FR4 trace length of 24 inches having a loss of 18 dB at dependence of the losses in the microstrip makes it difficult to
5 GHz. For this setup, the equalizer/CDR circuit was tested with use the standard transmission line model [Fig. 24(b)], where the
a supply voltage of 1.6 V. The sensitivity of the merged circuit series resistance and the parallel conductance remain con-
is plotted in Fig. 22, indicating a bit error rate (BER) of less than stant with frequency.
for a differential transmitted voltage of 640 mV . This While broadband models have been developed for skin effect
sensitivity can be improved by adding a simple gain stage after [18], an accurate model is not available for dielectric loss. It is
2008 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 42, NO. 9, SEPTEMBER 2007

Fig. 21. Measured results before and after equalization/data recovery at 10 Gb/s for FR4 traces (horizontal scale: 20 ps/div.): (a) before equalization/data recovery
for 24-in FR4 (horizontal scale: 10 ps/div., vertical scale : 75 mV/div.), (b) after equalization/data recovery for 24-in FR4 (horizontal scale: 20 ps/div., vertical
scale : 20 mV/div.), (c) after equalization/data recovery for 6-in FR4 (horizontal scale: 20 ps/div., vertical scale: 20 mV/div.).

Fig. 22. BER sensitivity graph for equalization/data recovery for 24-in FR4 at
10 Gb/s. Fig. 24. (a) Loss profile of 30-in microstrip on FR4 board. (b) Narrowband loss
model.

Fig. 23. Measured results after equalization/clock recovery for 24-in FR4 at
10 Gb/s. (a) Recovered clock spectrum. (b) Recovered clock jitter histogram
(horizontal scale: 5 ps/div.).

possible to extract the -parameters of a given line across a fre-

quency band using a vector network analyzer or a field simulator
and provide it to a tool capable of generating a spice model (e.g.,
SONNET). However, this approach provides little intuition and,
more importantly, requires entirely different sets of data for dif-
ferent line lengths. It is therefore desirable to develop a phys- Fig. 25. (a) Model for variable skin effect. (b) Proposed broadband model for
ical, scalable circuit model that includes frequency-dependent a 1-in trace.
dielectric loss and easily lends itself to circuit simulations. Such
a model must also represent the phase profile of the line with
reasonable accuracy. thus raising the loss at higher frequencies. Additional sections
Modeling begins with a correction of the circuit in Fig. 24(b) can be added to refine the model, but simulations indicate that
to accommodate frequency-dependent skin effect. Illustrated in one branch is adequate for modeling up to 7 GHz.
Fig. 25(a), the section consisting of and is added [18], In order to include the dielectric loss, the constant parallel
with ( ) modeling the low-frequency resistance of conductance, , in Fig. 24(b) is replaced with a frequency-de-
the line and forcing the high-frequency current through , pendent impedance that incurs greater loss at higher frequencies
GONDI AND RAZAVI: EQUALIZATION AND CLOCK AND DATA RECOVERY TECHNIQUES FOR 10-GB/S CMOS SERIAL-LINK RECEIVERS 2009

Fig. 26. Profiles of actual microstrip (dashed lines) and scalable model (solid
lines) for a 30-in trace. (a) Magnitude profile. (b) Phase profile. Fig. 27. (a) Error due to additive noise based on the sampling points in the eye.
(b) BER as a function of clock skew.

[Fig. 25(b)]. The conductance to the substrate, therefore, in- where denotes the rms value of noise. Carrying out the inte-
creases with frequency. Accurate modeling of the dielectric loss gration in (6) and neglecting higher-order terms, we have
necessitates at least two such sections: the loss in the –
branch becomes appreciable above 0.4 GHz and that in the
– branch above 20 GHz, thus providing good accuracy up
to 7 GHz. The above model can simply be cascaded to represent
longer traces. Fig. 26 compares the loss and phase profiles for
a 30-in trace as predicted by field simulations and the model,
demonstrating a reasonable fit.9
(10)
APPENDIX II
BER ESTIMATION For example, with mV, mV,
mV ,10 and sampling in the middle of the eye, we obtain
The BER of equalizer/CDR chains is a function of additive .
noise, ISI, and CDR skew and jitter. In this section, we propose In the presence of clock skew and jitter, the sampling point de-
a method of estimating the BER based on these parameters. The viates from the maximum eye opening, rapidly raising the BER.
objective is to develop an intuitive understanding of the tradeoffs (Since the ISI-induced jitter in the data waveform contains pre-
and hence arrive at a reasonable compromise. dominantly high-frequency components [19], the CDR circuit
Consider the eye diagram shown in Fig. 27(a), where fails to track the corresponding phase variations.) For
and denote the peak received swing and the peak eye to in Fig. 27(a), we approximate the eye opening as
opening, respectively. We first exclude clock jitter. The ISI and plot as a function of the
due to limited bandwidth or incomplete equalization leads to a clock skew [Fig. 27(b)]. Here, the above values of , ,
roughly uniform distribution of the amplitude between and are used and .
. Denoting this distribution by and that of additive We can now incorporate the effect of clock jitter by weighting
Gaussian noise by , we observe that an error occurs if (10) according to the jitter distribution, , where
noise causes a level between and to fall below zero denotes the random departure of the sampling point from the
(or a level between and to exceed zero). Thus, the center of the eye. As illustrated in Fig. 28(a), if places the
probability of error is given by sampling instant at , then is lower and the effect of
more pronounced. The error probability density function across
the bit period can therefore be expressed as
(6)
(11)
Also, Plotted in Fig. 28(b) for a Gaussian jitter distribution having
an rms value of ps, this result reveals that errors are
(7) most likely to occur in the vicinity of ps, where the
dramatic rise in still compensates for the fall in . Beyond
(8) this point the clock jitter becomes so improbable that errors are
less and less likely to occur.
and The framework developed above can be used to formulate
the trade-offs between required swings, additive noise, ISI, and
(9) tolerable clock jitter. For example, with mV and
, the plots in Fig. 29 can be generated, where
9Simulations indicate that if the proposed model is altered to incur one more
dB of error, the equalizer output suffers from 2.6 dB of additional vertical clo- 10The additive noise is obtained by integrating the noise at the output of the
sure and 0.03 UI of additional jitter. equalizer path up to 100 GHz.
2010 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 42, NO. 9, SEPTEMBER 2007

[4] T. Beukema, M. Sorna, K. Selander, S. Zier, B. L. Ji, P. Murfet, J.

Mason, W. Rhee, H. Ainspan, B. Parker, and M. Beakes, “A 6.4 Gb/s
CMOS SerDes core with feed-forward and decision-feedback equal-
ization,” IEEE J. Solid-State Circuits, pp. 2633–2645, Dec. 2005.
[5] R. Payne, P. Landman, B. Bhakta, S. Ramaswamy, W. Song, J.
D. Powers, M. U. Erdogan, A. -L. Yee, R. Gu, W. Lin, Y. Xie, B.
Parthasarathy, K. Brouse, W. Mohammed, K. Heragu, V. Gupta, L.
Dyson, and W. Lee, “A 6.4 Gb/s CMOS SerDes core with feed-forward
and decision-feedback equalization,” IEEE J. Solid-State Circuits, pp.
2646–2657, Dec. 2005.
[6] M. Meghelli, S. Rylov, J. Bulzacchelli, W. Rhee, A. Rylyakov, H.
Ainspan, B. Parker, M. Beakes, A. Chung, T. Beukema, P. Pepelju-
goski, L. Shan, Y. Kwark, S. Gowda, and D. Friedman, “A 10 Gb/s
5-Tap-DFE/4-Tap-FFE transceiver in 90 nm CMOS,” in IEEE Int.
Solid-State Circuits Conf. Dig. Tech. Papers, Feb. 2006, pp. 80–81.
[7] R. P. Jindal, “Gigahertz-band high-gain low-noise AGC amplifiers in
fine-line NMOS,” IEEE J. Solid-State Circuits, vol. 22, pp. 512–521,
Fig. 28. (a) Error due to additive noise and clock skew. (b) Error probability Aug. 1987.
density across the bit period. [8] E. Sackinger and W. C. Fischer, “A 3-GHz 32-dB CMOS limiting am-
plifier for SONET OC-48 receivers,” IEEE J. Solid-State Circuits, vol.
35, pp. 1884–1888, Dec. 2000.
[9] S. Gondi, J. Lee, and B. Razavi, “A 10 Gb/s CMOS adaptive equal-
izer for backplane applications,” in IEEE Int. Solid-State Circuits Conf.
Dig. Tech. Papers, Feb. 2005, pp. 328–329.
[10] S. Galal and B. Razavi, “10-Gb/s limiting amplifier and laser/modu-
lator driver in 0.18-m CMOS technology,” IEEE J. Solid-State Cir-
cuits, vol. 38, pp. 2138–2146, Dec. 2003.
[11] J. A. Mataya, G. W. Haines, and S. B. Marshall, “IF amplifier using
C compensated transistors,” IEEE J. Solid-State Circuits, vol. SC-3,
pp. 401–407, Dec. 1968.
[12] S. Gondi, “A 10-Gb/s CMOS merged adaptive equalizer/CDR circuit
for serial-link receivers,” in IEEE Symp. VLSI Circuits Dig. Tech. Pa-
pers., Jun. 2006, pp. 240–241.
[13] M. H. Shakiba, “A 2.5 Gb/s adaptive cable equalizer,” in IEEE Int.
Solid-State Circuits Conf. Dig. Tech. Papers, Feb. 1999, pp. 396–397.
[14] A. J. Baker, “An adaptive cable equalizer for serial digital video rates
Fig. 29. BER curves for different combination of received swings and clock to 400 Mb/s,” in IEEE Int. Solid-State Circuits Conf. Dig. Tech. Papers,
jitter. Feb. 1996, pp. 174–175.
[15] J.-S. Choi, M.-S. Hwang, and D.-K. Jeong, “A 0.18 m CMOS
3.5-Gb/s continuous-time adaptive cable equalizer using enhanced
low-frequency gain control method,” IEEE J. Solid-State Circuits, pp.
each curve represents the acceptable combination of and 419–425, Mar. 2004.
that yields a given BER. Worth noting here is the sharp rise [16] J. D. H. Alexander, “Clock recovery from random binary data,” Elec-
tron. Lett., vol. 11, pp. 541–542, Oct. 1975.
in the required swing as the clock jitter exceeds 3 ps . Also, [17] B. Razavi, Y. Ota, and R. G. Swartz, “Design techniques for low-
the experimental result in Fig. 23(b) indicates an rms jitter of voltage high-speed bipolar circuits,” IEEE J. Solid-State Circuits, vol.
2.2 ps, which from Fig. 29, would dictate a minimum swing of 29, pp. 332–339, Mar. 1994.
[18] C. S. Yen, Z. Fazarinc, and R. L. Wheeler, “Time-domain skin-effect
520 mV for BER . This value is fairly close to the model for transient analysis of lossy transmission lines,” Proc. of IEEE,
measured result of 640 mV in Fig. 22. vol. 70, pp. 750–757, Jul. 1982.
[19] V. Stojanovic and M. Horowitz, “Modeling and analysis of high-speed
links,” in IEEE Custom Integrated Circuits Conf. Dig. Tech. Papers,
Sep. 2003, pp. 589–594.
ACKNOWLEDGMENT

The authors wish to thank V. Pathak and Kawasaki Micro- Srikanth Gondi received the B.E. (Hons.) degree in
electronics America for support of this work, Dr. J. Lee for de- electrical and electronics engineering from Birla In-
sign assistance, and TSMC for fabrication support. stitute of Technology and Science, Pilani, India, in
1997, the M.Sc. degree in computer engineering from
Iowa State University, Ames, in 2002, and the Ph.D.
degree in electrical engineering from University of
REFERENCES California, Los Angeles, in 2006.
He was a Research Assistant at Iowa State Uni-
[1] K. Krishna, D. A. Yokoyama-Martin, S. Wolfer, C. Jones, M. versity until 1999, where he focused on mixed-signal
Loikkanen, J. Parker, R. Segelken, J. L. Sonntag, J. Stonick, S. Titus, circuits for the gigabit ethernet transceiver and the
and D. Weinlader, “A multigigabit backplane transceiver core in 1394 home-networking applications. From 2000 to
0.13 m CMOS for serial data transmission across high loss legacy 2002, he was a Mixed-Signal Design Engineer at Texas Instruments, Dallas, TX,
backplane channels,” IEEE J. Solid-State Circuits, pp. 2658–2666, working on high-speed interface architectures and circuits for wireline applica-
Dec. 2005. tions, realized in BiCMOS technologies. He was a Graduate Student Researcher
[2] G. Zhang and M. M. Green, “A BICMOS 10 Gb/s adaptive cable equal- at the University of California, Los Angeles, from 2002 to 2006, where he devel-
izer,” in IEEE Int. Solid-State Circuits Conf. Dig. Tech. Papers, Feb. oped circuits and architectures for serializer-deserializer (SerDes) applications.
2004, pp. 482–483. He is currently a Senior Design Engineer in the R&D Division of Kawasaki
[3] Y. Tomita, M. Kibune, J. Ogawa, W. W. Walker, H. Tamura, and Microelectronics America, working on analog front-ends for both wireless and
T. Kuroda, “A 10-Gb/s receiver with series equalizer and on-chip wireline applications. His current research interests include mixed-signal equal-
ISI monitor in 0.11 m CMOS,” IEEE J. Solid-State Circuits, pp. izers, data converters, phase-locking and clock recovery for data communica-
986–993, Apr. 2005. tions.
GONDI AND RAZAVI: EQUALIZATION AND CLOCK AND DATA RECOVERY TECHNIQUES FOR 10-GB/S CMOS SERIAL-LINK RECEIVERS 2011

Behzad Razavi (S’87–M’90–SM’00–F’03) re- includes wireless transceivers, frequency synthesizers, phase-locking and clock
ceived the B.S.E.E. degree from Sharif University of recovery for high-speed data communications, and data converters.
Technology, Tehran, Iran, in 1985 and the M.S.E.E. Prof. Razavi received the Beatrice Winner Award for Editorial Excellence at
and Ph.D.E.E. degrees from Stanford University, the 1994 ISSCC, the best paper award at the 1994 European Solid-State Circuits
Stanford, CA, in 1988 and 1992, respectively. Conference, the best panel award at the 1995 and 1997 ISSCC, the TRW Inno-
He was with ATT Bell Laboratories and Hewlett- vative Teaching Award in 1997, and the best paper award at the IEEE Custom
Packard Laboratories until 1996. Since 1996, he has Integrated Circuits Conference in 1998. He was the co-recipient of both the Jack
been Associate Professor and subsequently Professor Kilby Outstanding Student Paper Award and the Beatrice Winner Award for Ed-
of electrical engineering at the University of Cali- itorial Excellence at the 2001 ISSCC. He was recognized as one of the top 10
fornia, Los Angeles. He was an Adjunct Professor authors in the 50-year history of ISSCC. He received the Lockheed Martin Ex-
at Princeton University from 1992 to 1994, and at cellence in Teaching Award in 2006 and the UCLA Faculty Senate Teaching
Stanford University in 1995. He is the author of Principles of Data Conversion Award in 2007. He served on the Technical Program Committees of the Inter-
System Design (IEEE Press, 1995), RF Microelectronics (Prentice Hall, 1998) national Solid-State Circuits Conference (ISSCC) from 1993 to 2002 and VLSI
(translated to Chinese and Japanese), Design of Analog CMOS Integrated Cir- Circuits Symposium from 1998 to 2002. He has also served as Guest Editor
cuits (McGraw-Hill, 2001) (translated to Chinese and Japanese), Design of Inte- and Associate Editor of the IEEE JOURNAL OF SOLID-STATE CIRCUITS, IEEE
grated Circuits for Optical Communications (McGraw-Hill, 2003), and Funda- TRANSACTIONS ON CIRCUITS AND SYSTEMS, and International Journal of High
mentals of Microelectronics (Wiley 2006), and the editor of Monolithic Phase- Speed Electronics. He is an IEEE Distinguished Lecturer and a Fellow of IEEE.
Locked Loops and Clock Recovery Circuits (IEEE Press, 1996) and Phase-
Locking in High-Performance Systems (IEEE Press, 2003). His current research