HIGH-SPEED SERIAL DATA LINK DESIGN AND

advertisement
HIGH-SPEED SERIAL DATA LINK DESIGN AND SIMULATION
BY
EDWARD W. LEE
THESIS
Submitted in partial fulfillment of the requirements
for the degree of Master of Science in Electrical and Computer Engineering
in the Graduate College of the
University of Illinois at Urbana-Champaign, 2009
Urbana, Illinois
Adviser:
Professor José E. Schutt-Ainé
ABSTRACT
This thesis describes the modeling and simulation of 10 Gb/s serial data link architectures. The first architecture uses a continuous-time receiver equalizer to equalize
a channel with known frequency response. The second incorporates an adaptive decision feedback equalizer (DFE) for use in unknown or time-varying channels.
ii
To my Father and Mother
iii
ACKNOWLEDGMENTS
I would like to thank my adviser, Prof. José Schutt-Ainé, for his guidance and selection of this thesis topic. I would also like to thank Prof. Hyeon-Min Bae, without
whom this effort would not have been possible. Much appreciation goes to my colleague Johnson Liu for providing assistance with the compilation of this document
in LATEX. I am grateful for my brother Young and my dear friend Shinae and their
endless encouragement, especially on those tough days. Finally, my parents deserve
my utmost gratitude for their tireless prayers and support in my endeavors. This
thesis, for all it is worth, is dedicated to them.
iv
TABLE OF CONTENTS
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vi
LIST OF ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . .
vii
CHAPTER 1 INTRODUCTION . . . . . . . . . . . . . . .
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 High Speed Data Links . . . . . . . . . . . . . . . . . . . .
1.3 Challenges of Multi-Gigabit Backplane Data Transmission
1.4 Pre-Emphasis and Equalization . . . . . . . . . . . . . . .
1.5 Receiver Design for AWGN ISI Channels using DFE . . . .
1.6 Adaptive Equalization . . . . . . . . . . . . . . . . . . . .
1.7 Thesis Organization . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
2
2
3
6
7
8
DESIGN
. . . . . .
. . . . . .
. . . . . .
. . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
10
10
11
CHAPTER 3 DISCRETE-TIME ADAPTIVE EQUALIZER
DESIGN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Receiver Design . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
17
17
CONCLUSION . . . . . . . . . . . . . . . . . . . . . . .
19
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
CHAPTER 2 CONTINUOUS-TIME EQUALIZER
2.1 Methodology . . . . . . . . . . . . . . . . . . . . .
2.2 Channel Model . . . . . . . . . . . . . . . . . . . .
2.3 Receiver Model . . . . . . . . . . . . . . . . . . . .
2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . .
CHAPTER 4
v
.
.
.
.
.
.
.
.
LIST OF FIGURES
1.1
1.2
1.3
1.4
1.5
1.6
1.8
1.9
Backplane link diagram. . . . . . . . . . . . . . . . . . . . . . . . . .
Communication link. . . . . . . . . . . . . . . . . . . . . . . . . . . .
Simple pre-emphasis implementation. . . . . . . . . . . . . . . . . . .
(a) Simple DFE system. (b) DFE removing postcursor ISI. . . . . . .
Optimum receiver design with equalizer. . . . . . . . . . . . . . . . .
Example pulse shapes in receiver. (a) Output of sampled matched filter
is always symmetric. (b) Output of noise whitening filter resulting in
only postcursor ISI. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
DFE feedback equalizer combined with a feed-forward linear transversal equalizer, which mitigates precursor ISI. . . . . . . . . . . . . . .
Adaptive receiver block diagram. . . . . . . . . . . . . . . . . . . . .
Alternative training system. . . . . . . . . . . . . . . . . . . . . . . .
2.1
2.2
2.3
2.4
2.5
2.6
Design flow. . . . . . . . . . . . . . . . . . .
Channel model. . . . . . . . . . . . . . . . .
Receiver model. . . . . . . . . . . . . . . . .
Channel frequency response and group delay.
Frequency response after receiver. . . . . . .
Eye diagrams of receiver. . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
10
11
12
13
14
3.1
3.2
3.3
3.4
Channel model. . . . . . . . . . . .
Channel frequency response. . . . .
System model. . . . . . . . . . . . .
Adaptive DFE equalization results.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
16
16
17
18
1.7
vi
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2
2
4
5
6
7
7
8
8
LIST OF ABBREVIATIONS
BER
bit error rate
EMI
electromagnetic interference
FIR
finite-impulse response
DFE
decision feedback equalizer
ISI
intersymbol interference
LE
linear transversal equalizer
LMS
least mean square
MLSE
maximum-likelihood sequence estimation
PCB
printed circuit board
VGA
variable gain amplifier
vii
CHAPTER 1
INTRODUCTION
1.1
Motivation
The ever-increasing need for high data rate in communication systems is driving data
transmission within networking equipment to multi-gigabit per second rates. Serial
I/O interfaces are being widely adopted into backplane applications, short and longhaul communications, and chip-to-chip links for computing applications. A typical
backplane system is shown in Fig. 1.1, consisting of chip packages soldered onto
daughter cards that are then plugged into the backplane through connectors.
Looking at the many recent emerging industrial standards gives us insight into
the speeds of interest. Ethernet data rates are advancing from 100 Mb/s onward
to 10 Gb/s. In computing applications, serial ATA is increasing from the current
1.5 Gb/s and 3 Gb/s, and targeting 6 Gb/s. PCI Express 2 is 5 Gb/s and going on
to 8 Gb/s in PCI Express 3.
Copper/FR4 backplane based serial transmission is one of the most commonly
used techniques in high-speed transmission due to its lower cost, reduced complexity
and high reliability. Due to these advantages, copper/FR4 PCBs are expected to
continue as the material of choice for telecomm and computing applications. However,
these PCBs do face some technical challenges, mainly because of distortion due to
skin effect, dielectric loss and reflections. This results in reduction of transmission
bandwidth, leading to closure of the transmission eye and, ultimately, high BER at
the receive end.
1
Package
Daughter Card
Trace
Connector
Backplane Trace
Figure 1.1: Backplane link diagram.
1.2
High Speed Data Links
A communication link typically consists of a transmitter, the communication channel,
and a receiver (Fig. 1.2). The transmitter takes the digital data and converts it
to analog waveforms on the channel. The channel is the communication medium
between the transmitter and the receiver, and can have physical realizations such as
free space for wireless communication, optical fibers for optical communication and
PCB traces, coaxial cables or twisted pair wires for off-chip electrical communication.
When a channel is implemented using electrical technologies, the implementation is
often referred to as an interconnect.
Transmitter
Channel
Receiver
Figure 1.2: Communication link.
1.3
Challenges of Multi-Gigabit Backplane Data
Transmission
The challenges of multi-gigabit data transmission can largely be divided into two
categories, the first being structural issues and the second being frequency dependent
2
loss due to material properties. One of the greatest structural challenges is overcoming
impedance mismatches due to connector-via transitions. Improved laminates and
optimized connector-via architectures can alleviate these issues, but as data rates
steadily rise, the frequency dependent loss of the channel which usually manifests as
low-pass nature becomes a serious threat to signal integrity.
Skin effect describes how high frequency currents tend to travel on the surface of
the conductor, rather than the whole cross section of the conductor. This reduces the
effective conductive area of the trace, increasing resistance which causes the signal
to be attenuated. As frequencies increase into the multi-gigabit per second rates,
dielectric loss emerges as the dominant factor in high frequency attenuation, since its
effect is proportional to frequency, whereas skin effect is proportional to the square
root of frequency. Dielectric loss can be defined as a ratio of conductivity to frequency,
known as loss tangent. Materials such as FR-4 with a high loss tangent will see a large
attenuation of signal at high frequency. The combined attenuation and dispersion due
to these effects causes the signal to spread over to adjacent signals. This is called
intersymbol interference (ISI).
From a circuit point of view, the limited bandwidth of the channel increases the
rise-time/fall-time of the signal, which causes the voltage level of a single bit to depend
on the bit pattern it resides in. The degradation of the signal due to ISI results in
two main issues on the receiver side: reduced voltage margin and timing errors due
to jitter in zero crossings. In order to reduce BER, transmitter pre-emphasis, receiver
equalization, or a combination of the two is generally used to compensate the channel.
1.4
Pre-Emphasis and Equalization
Pre-emphasis boosts the high frequency components of the signal at the transmitter,
before the signal is sent to the channel. One simple implementation of pre-emphasis
3
is a 2-tap finite impulse response (FIR) filter, resulting in a boost of signal every time
there is a signal transition (Fig. 1.3).
Some of the disadvantages of pre-emphasis include higher power requirements
to boost the signal, aggravated crosstalk, and increased electromagnetic interference
(EMI) due to the overshoot and undershoot. Also, as the channel is usually not known
a priori, more focus is aimed at equalization design, while a simpler pre-emphasis
implementation is used.
X
Z
D
Y
X
Y
a
Z
Figure 1.3: Simple pre-emphasis implementation.
Equalization is used at the receiver to flatten the channel frequency response
to overcome the high frequency signal loss during transmission. Receivers can be
implemented in discrete-time or continuous-time. Continuous-time receivers can be
designed using an analog equalizer, while discrete-time equalizers can be designed
using digital filters.
Several types of discrete-time designs can be considered. As introduced in the
literature [1], using a maximum-likelihood sequence estimation (MLSE) receiver is
optimum according to a probability of error criterion. However MLSE has a complexity that grows exponentially with the length of the ISI channel time dispersion,
which makes implementation expensive. Other sub-optimum methods include using
a linear transversal equalizer (LE) or a decision feedback equalizer (DFE).
The linear transversal equalizer (LE), which can be viewed as being the same as a
FIR filter, is very versatile in that an infinite length LE can equalize any channel H(ω)
as long as H(ω) 6= 0, |ω| ≤
π
.
T
Also, it can be used to implement matched filtering
as well if the taps are spaced at Nyquist intervals instead of symbol intervals [2, 3].
4
One of the main disadvantages of using a common LE design is that the equalizer
usually has a high-pass characteristic due to the low-pass nature of the channel. This
can result in severe amplification of high frequency noise components, especially with
channels with spectral nulls.
The decision feedback equalizer (DFE) is a non-linear filter which allows some of
the noise problems associated with linear equalizers to be overcome. By using a linear
combination of past decisions, the DFE compensates for the channel response while
eliminating noise amplification. While the DFE is vulnerable to ‘error propagation’
where an initial decision error will feed back and cause a sequence of errors, the system
will recover from the error events in relatively short order if the data is insured to be
random [4].
Figure 1.4 shows the operation of a DFE canceling postcursor ISI, in which the
pulse response of a channel is assumed to be a simple RC low-pass filter [5].
X
Y
D
a1
D
a2
D
a3
D
1
a1
a2
a3
a4
a4
(a)
(b)
Figure 1.4: (a) Simple DFE system. (b) DFE removing postcursor ISI.
5
1.5
Receiver Design for AWGN ISI Channels
using DFE
Figure 1.5 shows the optimum equalizer receiver design process for digital transmission
channels corrupted by AWGN and ISI [4, 6]. A matched filter is used for maximum
SNR before the sampler, and a noise whitening filter is selected so that the resulting
response is minimum phase. This creates a causal and stable output F (z), effectively
eliminating precursor ISI, as shown in Fig. 1.6. The matched filter, symbol-rate
sampler and noise whitening filter are collectively called the whitened matched filter
(WMF). As can be seen, the DFE front-end is equivalent to the WMF [7], which
ultimately operates as the precursor equalizer for a DFE (Fig. 1.7).
AWGN ISI Channel
...10001001
In
Transmitter
pulse shape
Channel freq
response
g (t )
c(t )
Received
Signal
r (t )
Matched filter
h* (t )
Symbol-spaced
sampler
t  kT
z (t )
h(t )
AWGN
Discrete-time model
of channel with ISI
...10001001
In
1
F ( z 1 )
X ( z)
*
Noise
whitening filter
vn
Non-white
Gaussian Noise
Whitened-Matched Filtered
ISI channel model
...10001001
In
1
F ( z)
F ( z)
n
Equalizer
AWGN
Figure 1.5: Optimum receiver design with equalizer.
6
n
-3 -2 -1 0 1 2 3
Precursor ISI
Postcursor ISI
n
n
0 1 2 3
-3 -2 -1 0 1 2 3
Precursor ISI
Postcursor ISI
Postcursor ISI
(a)
(b)
Figure 1.6: Example pulse shapes in receiver. (a) Output of sampled matched filter
is always symmetric. (b) Output of noise whitening filter resulting in only postcursor
ISI.
n
Channel
In
DFE
Iˆn
0 1 2 3
H ( z)
E( z)
Postcursor ISI
zn
Precursor
Equalizer
Postcursor
Equalizer
H ( z)E( z) 1
Figure 1.7: DFE feedback equalizer combined with a feed-forward linear transversal
equalizer, which mitigates precursor ISI.
1.6
Adaptive Equalization
Adaptive equalization [2] must be considered because we cannot assume to know
beforehand what the channel response is, and we are limited by implementation complexity. Coefficients can be decided using MMSE criterion [8], and the LMS algorithm
is widely used due to its ease of implementation.
The simplest form of adaptive equalizer is shown in Fig. 1.8. In steady state, the
decisions at the slicer output are assumed correct and are used to generate the error
signal (decision directed mode). However, the decision directed mode is only effective
in tracking slow variations that result in a low error rate at the slicer, and requires a
training sequence to initially acquire the equalizer coefficients (training mode).
7
Received
Signal
Sampler
Receive
Filter
Slicer
Decisions
Adaptive
Equalizer
Decision
directed
mode
Error Signal
Training
mode
Training Signal
Figure 1.8: Adaptive receiver block diagram.
To avoid singularities from the zero of the channel transfer function inside the unit
circle, we can select an equalizer such that W (z)H(z) ∼
= z −L , where the result is a
time-shifted version of the desired signal, as shown in Fig. 1.9. A more general W (z) =
Y (z)/H(z) can also be used to reduce noise enhancement or to ease implementation,
in what is known as a partial-response signaling system [9].
zL
I ( n)
...10001001
I ( n  L )  d ( n)
Channel freq
response
Adaptive LMS
equalizer
H ( z)
W ( z)
e( n )
 ( n)
Figure 1.9: Alternative training system.
1.7
Thesis Organization
This thesis documents all the effort devoted to design and improve upon a high-speed
serial link receiver. Chapter 2 describes the design of a continuous-time receiver
equalizer for use in 10 Gb/s data links. A discrete-time adaptive DFE receiver is
simulated using ADS and measured channel S-parameters in Chapter 3.
8
CHAPTER 2
CONTINUOUS-TIME
EQUALIZER DESIGN
2.1
Methodology
Compared with discrete-time equalizers, continuous-time equalizers have some attractive properties in terms of noise, jitter and potential power dissipation. The use of
passive networks is also gaining attention due to the low power and smaller size of
implementation. In this chapter, an analog equalizer is designed for use in a 10 Gb/s
data receiver.
The design flow is shown in Fig. 2.1. The channel is first modeled using the
W-element in HSPICE where the channel AC response and transient output are
simulated. The AC response is used to obtain the single zero and pole of the analog
equalizer. The transient channel output is fed into MATLAB Simulink where the
receiver is modeled.
Figure 2.1: Design flow.
9
2.2
Channel Model
The channel is modeled as a differential line with a FR-4 coupled microstrip. The
2000 traces on the line are copper with width 508 µm, thickness 35 µm, and thickness
of the dielectric being 254 µm. The width between the traces is set to be 700 µm.
The HSPICE field solver is used with the W-element to simulate the transmission
line. The driver supplies 50 W source termination to the channel and has 10 GHz
bandwidth. The receiver termination includes AC coupling capacitance, bond-wire
inductance, ball-pad capacitance, and die-pad capacitance. Input amplitude of the
voltage source is set to be 1 Vpp differential, with additive channel noise modeled as
having infinite bandwidth and 30 dB SNR. The complete channel model is shown in
Fig. 2.2.
Figure 2.2: Channel model.
2.3
Receiver Model
The receiver model consists of a single pole variable gain amplifier (VGA) and the
equalizer used to flatten the frequency response. The equalizer is modeled as a single
zero, single pole amplifier. The single real zero is needed to compensate the gain loss
and open the signal eye. The real pole is needed for several reasons. First, boosting
the gain above a certain frequency (usually the signal bandwidth) is unnecessary
because it merely amplifies more high frequency noise. Second, wide bandwidth
10
causes signal peaking in the middle of the eye, generating large jitter at zero crossing
points of the data signal. Signal peaking in the middle of the eye also decreases the
voltage margin at the middle of the eye, which increases BER. Finally, signal peaking
increases crosstalk between signal lines.
Thus the real pole is needed to limit the bandwidth. The pole location determines
the amount of boosting and the frequency range due to the zero. If the signal loss
is low, the zero and the pole can be close together. But, if the loss is high, the zero
and the pole should be separated further to get sufficient boosting. It is desirable to
have the zero and pole as close as possible to avoid adding excessive non-linear phase
response, for a symmetrical eye.
The receiver model which was simulated in MATLAB is shown in Fig. 2.3.
Figure 2.3: Receiver model.
2.4
Results
The channel model frequency response and group delay are shown in Fig. 2.4, displaying the channel, driver-channel, and driver-channel-termination, respectively. Ta11
ble 2.1 summarizes the channel characteristics. A -6 dB gain offset in the low frequency is due to the 50 W output impedance of the driver, and the channel and driver
losses at 10 GHz are -10 dB and -3 dB, respectively. At this frequency, the impedance
mismatch between channel and termination causes additional 2 dB signal loss. The
phase response is fairly linear from 100 MHz up to 10 GHz and group delay at this
frequency range is about 3 ns.
Figure 2.4: Channel frequency response and group delay.
Table 2.1: Channel characteristics.
Channel
Gain @ low freq. -6.07 dB
Gain @ 10 GHz
-16 dB
3 dB freq.
2.63 GHz
Group Delay
3.05 ns
Driver + Channel
-6.08 dB
-19 dB
2.36 GHz
3.07 ns
Drv. + Ch. + Termination
-6.08 dB
-21.1 dB
2.33 GHz
3.1 ns
The flattening of the frequency response after the receiver is shown in Fig. 2.5.
The pole of the VGA is at 7 GHz and the equalizer has a 2.1 GHz zero and 4 GHz pole.
The eye diagram of channel output has relatively good eye shape which shows that the
signal loss at high frequency in channel is not severe (Fig. 2.6(a)). Slight asymmetry
of the eye shows that phase response is not perfectly linear. The VGA filters high
12
Another real pole is inserted in the variable gain amplifier (VGA). The channel noise is
modeled after the channel with infinite bandwidth. The pole of VGA will attenuate the
channel noise.
B. Specifications
frequency
noise, increasing the SNR to 46.2 dB. However, the eye is more closed and
• crossing
20dB gain
range linear
VGA
has higher zero
variations
due to
the limited bandwidth (Fig. 2.6(b)).
o
Linearity : 25dB IMD worst case (250mVpp test tone at 5G and 6GHz)
The difference
between
the zero
and filter
the pole
of the
equalizer
about
GHz.
• high
frequency
boosting
(analog
equalizer)
forisskin
loss2compensation
o single real zero and single real pole
• <15ps spread in zero crossing after equalization
• than
Zero 2should
be tunable.causes a dent in the frequency response and
response. Less
GHz difference
More than 2 GHz difference causes overshoot in both eye diagram and frequency
closes
eye.modeling
The 3 dB frequency of the channel response is about 2.6 GHz, so the
C. the
AFE
zero should •be near
that
frequency.
The
zeroamplifier
is tweaked
toa2.1
VGA
is modeled
with
ideal
with
realGHz
poleconsidering
and infinitethe
linearity.
Analog the
equalizer
s-domain
function
MATLAB, and
zero crossing• variation,
flatnessis ofmodeled
the eye with
center,
and thetransfer
symmetry
of theineye.
filter using ideal Rs and Cs in Verilog-A.
After equalization, zero crossing deviation is reduced to 4 ps and SNR is 43.9 dB
D. Simulation Results
(Fig. 2.6(c)).
The eye diagram of channel output has relatively good eye shape which shows the signal loss
The
SNR
is degraded
at the equalizer
output is
by not
2.6 severe.
dB due ZFE
to the
at
high
frequency
in channel
and termination
is equalizer
a good choice as an
equalizerthe
because
ofas
channel
output
is high,
30dB.
A little
asymmetry
of eye tells that
attenuating
signalSNR
as well
the noise
(Table
2.2). The
signal
attenuation
is more
phase response is not perfectly linear. High frequency noise is filtered by the low-pass
significant
due toofthe
signal
a VGA
low-frequency
compared
the noiseeye is closed
characteristic
VGA,
thushaving
SNR of
output is profile
increased
to 46dB.toHowever,
more and zero crossing variation is higher because of the limited bandwidth.
power which is distributed with a wide frequency range.
magnitude (dB)
frequency response
After receiver
-10
Before receiver
-20
-30
-40
8
10
9
10
10
frequency (log)
10
4
phase(degree)
0
x 10
-1
-2
-3
0
0.2
0.4
0.6
0.8
1
1.2
frequency (linear)
1.4
1.6
Figure
2.5:5.Frequency
after
receiver.
Figure
frequencyresponse
response
after
AFE.
13
1.8
2
10
x 10
SNR : 30.2dB
zero crossing variation : 10ps
SNR : 46.2dB
zero crossing variation : 15ps
(a) channel output.
(b) VGA output : 7GHz real pole.
(a) Channel output
: 30.2dB
SNRSNR
: 30.2dB
crossing
variation
: 10ps
zerozero
crossing
variation
: 10ps
(b) VGA output
: 46.2dB
SNRSNR
: 46.2dB
crossing
variation
: 15ps
zerozero
crossing
variation
: 15ps
(a) channel
output.
(a) channel
output.
(b) VGA
output
: 7GHz
(b) VGA
output
: 7GHz
realreal
pole.pole.
(c) Equalizer
output
zero : 2.1GHz,
pole : 4GHz
SNR : 43.9dB
Figure
2.6: Eye
diagrams
zero
crossing
variation
: 4psof receiver.
(c) eye diagram – equalizer output.
: 2.1GHz,
: 4GHz
zerozero
: 2.1GHz,
polepole
: 4GHz
:Figure
43.9dB6. Eye Diagram.
SNRSNR
: 43.9dB
crossing
variation
zerozero
crossing
variation
: 4ps: 4ps
(c) diagram
eye diagram
– equalizer
output.
(c)equalizer
eye
– equalizer
The zero and pole of the analog
are 2.1GHz
and output.
4GHz,
respectively, thus, the distance
Table 2.2: Receiver output
between the zero and the pole is about 2GHz. More than 2GHz distance causes overshoot in both
Figure
6.
Eye
Diagram.
Figure
6. Eye
Diagram.
eye diagram and frequency response.Channel
Less
than
2GHz
distance
the frequency
Output
VGA causes
Outputdent in
Equalizer
Output
response Signal
and closes
the
eye.
3dB
frequency
of
channel
response
is
about
2.6GHz,
so,
the
zero
Power Sum
0.0358 W·Hz
0.0312 W·Hz
0.049 W·Hz
The
zero
and
pole
of
the
analog
equalizer
are
2.1GHz
and
4GHz,
respectively,
thus,
the
distance
The
zero be
andaround
pole of
thefrequency.
analog equalizer
areof2.1GHz
and the
4GHz,
thus,considering
the distance
should
that
After lost
simulation,
zerorespectively,
is set
the
−7 to 2.1GHz
Noise
Power
3.442GHz.
× 10−5More
W·Hz
7.46 × 10
W·Hz
2.02
× 10−6 in
W·Hz
between
thevariation,
zero
andSum
theflatness
pole
is about
than
distance
causes
overshoot
both
between
the zero
and the
pole
is about
than
2GHz
distance
causes
overshoot
in both
zero
crossing
the
of2GHz.
the eyeMore
center,
and
the2GHz
symmetry
of the
eye.
Although
VGA
SNR
30.2
dB
46.2
dB
43.9
dB
eye
diagram
and
frequency
response.
Less
than
2GHz
distance
causes
dent
in
the
frequency
eye
andhave
frequency
response.phase
Less response,
than 2GHzthedistance
in the because
frequencythe
anddiagram
equalizer
the nonlinear
effect causes
is not dent
significant
Zero
Crossing
Variation
psof
15because
ps
4so,ps
response
and
closes
the
eye.
3dB
frequency
channel
response
is about
the zero
response
and
closes
the eye.
3dBequalizer
frequency
of
response
about
2.6GHz,
the
zero
maximum
phase
deviation
of
is 10
lesschannel
than
20°
(this
isis
the2.6GHz,
zeroso,and
pole
of
should
be
around
that
frequency.
After
lostsimulation,
of VGA
simulation,
the zero
set2.1GHz
to 2.1GHz
considering
should
be around
that and
frequency.
After
lost of
zero
is setisto
the isthe
equalizer
is close.)
the
phase
deviation
of
atthe
below
7GHz
is
45°
(theconsidering
pole
of VGA
zero
crossing
variation,
the
flatness
ofeye
the center,
eye
center,
and
the symmetry
of eye.
the eye.
VGA
zero
crossing
variation,
of the
and
the
symmetry
the
Although
VGA
7GHz).
Furthermore
,the
theflatness
phase
deviations
of VGA
and
equalizer
areofopposite
eachAlthough
other.
After
and
equalizer
nonlinear
phase
response,
effect
is not
significant
because
and
equalizer
havehave
the the
nonlinear
phase
response,
the the
effect
is not
significant
because
the the
equalization,
zero
crossing
deviation
is only
4ps
and SNR
is 43.9dB.
maximum
phase
deviation
of
equalizer
is
less
than
20°
(this
is
because
the
zero
and
pole
maximum phase deviation of equalizer is less than 20° (this is because the zero and pole of of
Fig.
7 showsisthe
discrete-time
response
with 10GS/s.
If 7GHz
the frequency
response of
is flat
in
equalizer
close.)
the frequency
phase
deviation
of VGA
at below
is 45°
VGA
equalizer
is close.)
and and
the phase
deviation
of VGA
at
below
7GHz
is 45°
(the (the
polepole
of VGA
is is
14
the7GHz).
entire frequency
(DC~10GHz)
the
signal does
not have
ISI at the are
sampling
points. Before
the
Furthermore
the phase
deviations
of VGA
equalizer
opposite
other.
7GHz).
Furthermore
, the, phase
deviations
of VGA
and and
equalizer
are opposite
eacheach
other.
AfterAfter
equalization,
the
frequency
response
has
the
maximum
2dB
difference
in
magnitude.
After
equalization,
crossing
deviation
is only
4ps SNR
and SNR
is 43.9dB.
equalization,
zerozero
crossing
deviation
is only
4ps and
is 43.9dB.
equalization, the frequency response is much flatter especially above 1GHz as expected in zero-
CHAPTER 3
DISCRETE-TIME ADAPTIVE
EQUALIZER DESIGN
3.1
Methodology
In many cases, the channel of interest is unknown or has characteristics that vary
over time and temperature. This is very common in backplanes where by design,
daughter cards are meant to be able to be installed or uninstalled according to the
user’s needs. In this case an adaptive method of updating the coefficients of the
equalizers is most desirable. Given the measured S-parameters from a backplane, we
design and simulate a discrete-time adaptive DFE to equalize a 10 Gb/s data link.
Using the Signal Integrity Workshop DesignGuide in Advanced Design System
(ADS), an adaptive DFE is modeled and simulated. The LMS algorithm [1, 9] is
used to adaptively find the equalizer coefficients for both precursor and postcursor
parts of the receiver. Unlike the system described in the previous chapter, this system
is simulated in discrete-time with a global clock, and so zero crossings are a non-issue
in this case.
The backplane of interest is a 1 m Nelco 4000 13-SI board having 2.500 traces
on each daughter card. The measured 4-port S-parameters of the backplane were
obtained from the IEEE 802.3ap Backplane Ethernet Task Force [10] (Fig. 3.1). The
frequency response of the channel is shown in Fig. 3.2.
15
Figure 3.1: Channel model.
Figure 3.2: Channel frequency response.
16
3.2
Receiver Design
The receiver is modeled as a 7-tap precursor and 7-tap postcursor symbol-spaced
equalizer (Fig. 3.3). Transmit bits are generated using a pseudo-random bit sequence.
An initial training period using the correct transmitted data is required to converge
the DFE before the receiver is able to begin coefficient tracking, where the slicer
decisions are used instead.
Figure 3.3: System model.
3.3
Results
The LMS algorithm is seen to be successfully converged and the channel equalized.
As symbol-spaced sampling is used, an open discrete-time eye diagram is obtained
using the data at the input of the decision slicer (Fig. 3.4).
17
(a) Precursor equalizer weights
(b) Postcursor equalizer weights
(c) Received eye diagram
(d) Opened eye diagram
Figure 3.4: Adaptive DFE equalization results.
18
CHAPTER 4
CONCLUSION
We have presented two different equalizer architectures suitable for 10 Gb/s serial
data links. Design trade-offs in both transmit-side and receive-side equalizers were
described. A continuous-time equalizer was designed to equalize a channel with known
characteristics. An adaptive DFE for use in unknown or time-varying channels was
also presented. Future work could be extended to parallel architectures, and the use
of pre-computation or pipelining architectures such as unfolding could be included
for higher performance.
19
REFERENCES
[1] J. Proakis and D. Manolakis, Digital Signal Processing: Principles, Algorithms,
and Applications. Upper Saddle River, NJ: Prentice-Hall, Inc., 1996.
[2] R. Lucky, J. Salz, and E. Weldon, Principles of Data Communication.
York, NY: McGraw-Hill, 1968.
New
[3] M. DiToro, “A new method of high-speed adaptive serial communication through
any time-variable and dispersive transmission medium,” in 1st IEEE Annual
Communication Convention, 1965, pp. 763–767.
[4] J. Barry, E. Lee, and D. Messerschmitt, Digital Communication. Norwell, MA:
Kluwer Academic Publishers, 2004.
[5] B. Song and D. Soo, “NRZ timing recovery technique for band-limited channels,”
IEEE Journal of Solid-State Circuits, vol. 32, no. 4, pp. 514–520, 1997.
[6] J. Proakis, Digital Communications. New York, NY: McGraw Hill International
Editions, 2001.
[7] R. Price, “Nonlinearly feedback-equalized PAM vs. capacity,” in Proceedings of
IEEE International Conference on Communications, Philadephia, PA, 1972, pp.
22.12–22.17.
[8] J. Cioffi, G. Eyuboglu, and M. Forney, “MMSE decision-feedback equalizers and
coding. Part I. Equalization results,” IEEE Transactions on Communications,
vol. 43, no. 10, pp. 2582–2594, 1995.
[9] A. Poularikas and Z. Ramadan, Adaptive Filtering Primer with MATLAB. Boca
Raton, FL: CRC Press, 2006.
[10] G. Oganessyan, “IEEE P802.3ap task force channel model material,”
2006. [Online]. Available: http://grouper.ieee.org/groups/802/3/ap/public/
channel model/index.html
20
Download