Crosstalk problems are back

-May 29, 2012

The inexorable demand for electronic systems with increasing bandwidth and decreasing size puts more high-speed channels in ever-closer proximity. Technologies such as 40-Gbps and 100-Gbps Ethernet employ up to 10 channels at 10 Gbps each or four channels at 25 Gbps. When so many high-speed serial lanes reside in a single system, they're bound to interfere with each other.

Serial buses such as Ethernet, Fibre Channel, and PCI Express capitalize on the robust nature of serial technology, with its interference-canceling differential signaling and jitter-canceling embedded clocking. To achieve their data rates, these buses employ multiple serial lanes that operate in parallel. With each additional lane, a bus scales to a higher data rate (Ref. 1).

Unfortunately, every lane is both an aggressor and a victim. Differential signaling can only partially cancel crosstalk at these high data rates. After a decade spent developing serial data technologies and dealing with jitter, closed-eye diagrams, and pre-emphasis and de-emphasis, and then having equalization improve signal quality and reduce bit errors, engineers have realized that crosstalk has come back to haunt them.

Marty Miller, chief scientist at LeCroy, put it like this: "When we started switching from parallel to serial interconnects, crosstalk stopped being a big concern, but now we're moving to serial channels in parallel. We have systems with dozens of SerDes on one chip operating at 50 times the data rate [of parallel buses]. We're looking at a crosstalk nightmare."

What is crosstalk, anyway?

Crosstalk is the electromagnetic interference of multiple signals that occurs when radiation from an aggressor channel is picked up by a victim channel. Maxwell's equations describe electromagnetic radiation only when electric and magnetic fields change. Crosstalk is generated during logic transitions. With rise and fall times of 20–30 ps at 10 Gbps or 5–10 ps at 25 Gbps, the emerging high-rate standards have rapidly changing electric fields that can couple from one lane to another. The faster the change, the louder the crosstalk.

High-speed serial systems use differential signaling to beat back all types of EMI (electromagnetic interference). Differential pairs consist of two conductors. One carries the signal, and the other carries the inverse of the signal, or its complement. If the conductors are arbitrarily close together and are of precisely the same length, then they carry identical EMI. The receiver takes the difference of the pair, canceling the interference and reinforcing the differential signals. This is called common-mode rejection, because signals common to both elements of the pair are rejected.

Unfortunately, real differential pairs have non-zero separation, are not exactly the same length, and suffer other asymmetries such as impedance variations at contact points and variations in trace widths, thickness, and roughness. These imperfections limit common-mode rejection.

Engineers developing multilane serial links anticipated crosstalk's return, but that's not the whole story. Pavel Zivny, domain expert on high-speed serial data with Tektronix, said, "Intersignal crosstalk was expected, but there's another source of crosstalk that we didn't anticipate. As the transmitters on a SerDes chip draw current from the power plane, they generate power-supply variations that have different properties." Current draw in one device can cause variations in the power delivered to an adjacent device, a condition known as "chip crosstalk."

Another form of crosstalk, intersignal crosstalk, appears as sporadic jolts on the signal at the victim whenever an aggressor makes a logic transition. The interference lasts for the duration of that transition, whereas chip crosstalk has a longer duration.

Crosstalk is exacerbated by pre-emphasis and de-emphasis, which enable high-speed serial data receivers to reach a low BER (bit-error rate), even when signals are so degraded their eye diagrams are closed. At the transmitter, pre-emphasis enhances the high-frequency content of signals, which fends off the low-pass nature of the transmission lane's frequency response. Unfortunately, that high-frequency content generates the most crosstalk.

It's worse at the receiver. As Miller said, "Two of the three standard equalization techniques CTLE (continuous-time linear equalization) and FFE (feed-forward equalization), amplify crosstalk noise. The third technique, DFE (decision-feedback equalization), is the only one that doesn't make it worse, but nothing we have now makes it better."

Another complication arises in systems with parallel buses. Because the channels in parallel buses are frequency locked, they have fixed-phase relationships. With fixed-phase relationships, the time position of intersignal crosstalk in the eye diagram is also fixed. If crosstalk jolts occur at the crossing point in an eye diagram, they have a smaller impact on the BER than if the crosstalk jolts are offset by half a bit period and occur in the center of the eye at the victim.

Figure 1 shows an eye diagram with no crosstalk jolts. In Figure 2a, the crosstalk jolts appear at the crossings, and in Figure 2b, the jolts appear at the eye opening's widest point.

Crosstalk, Figure 1

FIGURE 1
. An eye diagram that has no crosstalk will have no impairments. Courtesy of Tektronix.


a)
Crosstalk, Figure 2a

b)
Crosstalk, Figure 2b

FIGURE 2. These eye diagrams display the effects of different victim-aggressor phase relationships. a) The aggressor and victim are in phase, and the crosstalk impairment is at the crossing point (two of these points are highlighted by red arrows). b) The aggressor and victim are half a bit out of phase, and the crosstalk impairment is at the center of the eye. Courtesy of Tektronix.
 
In many systems that use multiple serial lanes, including 40/100 GbE (Gigabit Ethernet), each lane operates at the same nominal frequency, but the lanes aren't locked. Because each channel operates on a clock recovered from the data, the relative phases are free to float. Figure 3 shows how unlocked, or "asynchronous," crosstalk smears the traces and is deceptively similar to the effect of random noise and jitter.

 
Crosstalk, Figure 3
FIGURE 3. When the aggressor and victim are asynchronous, crosstalk noise varies across the victim's eye diagram. Courtesy of Tektronix.

Oscilloscopes analyze crosstalk

If you're in complete control of a serial-data link's individual lanes, then you have the luxury of performing systematic, straightforward crosstalk analyses. Assume that you have a 100-Gbps link made from ten 10-Gbps lanes. Start by analyzing lane 1. Turn off lanes 2 through 9 and check the waveform and eye diagram of lane 1. Then, turn on a lane adjacent to 1. Any degradation you see is caused by crosstalk. Next, turn on each lane, and you can gauge the trouble. This is a simple diagnostic recipe for finding lanes that are more troublesome than others.

If the lanes are frequency locked, you should be able to isolate the problem. If they're not locked, then the degradation will look like a mix of random jitter and noise. In the unlocked case, it's better to examine the waveform, bit by bit, comparing the aggressor waveform to the victim waveform.

Unfortunately, these techniques won't help you estimate the eye closure corresponding to a given BER, which is what compliance testing requires, usually down to a BER of 10-12. Plus, you're not likely to have that much control of individual channels. Rob Sleigh, product marketing engineer for sampling scopes in Agilent Technologies' Digital Test Division, said, "At these speeds, the connectors and ICs have much smaller geometries and we should expect crosstalk from various sources. Additionally, 100 Gigabit Ethernet gearboxes switch between 4x25 Gbps and 10x10 Gbps, and since you can't shut off individual lanes, it makes crosstalk isolation even more challenging."

How Tektronix addressed the problem

Perhaps the most annoying problem with crosstalk is that jitter-analysis software frequently mistakes crosstalk for RJ (random jitter). Since RJ contributes disproportionately to the estimate of total jitter defined at a given BER, crosstalk can appear as a far more egregious problem than it really is.

In Zivny's view, "The main problem is to accurately predict total jitter and eye opening while providing a quantitative measure of the crosstalk problem." Tektronix accomplished these goals by implementing an approach that Zivny conceived.

In the early 2000s, manufacturers of test and measurement equipment were in a heated battle to determine the most accurate jitter-analysis technique (Refs. 2 and 3). At the time, measurements taken with different equipment often varied by more than 100%. One of the approaches that brought the problem under control was to use a spectral technique for measuring RJ. With this technique, engineers look at the spectrum of the timing of logic transitions; periodic components show up as spikes that can be removed.

Figure 4 shows the effect that crosstalk has on the spectrum of uncorrelated jitter. Note the higher noise level and additional spikes in Figure 4b that are missing from Figure 4a. The RMS value of the remaining smooth continuum is then identified with RJ. RJ is assumed to follow a Gaussian distribution, and the amount of eye closure at a given BER can be estimated using the dual-Dirac model (Ref. 4). Crosstalk shows up in that continuum and, thus, the spectral techniques mistake crosstalk for RJ and overestimate eye closure, sometimes by more than an order of magnitude.

a)
Crosstalk, Figure 4a

b)
Crosstalk, Figure 4b
FIGURE 4. A spectral plot of RJ and PJ shows that crosstalk adds noise and spectral peaks that are missing in (a) and present in (b).
 
Zivny's idea was to first isolate RJ and crosstalk from the spikes of sinusoidal and periodic jitter in the spectrum. Because this data set doesn't contain any elements that are correlated to the data, like intersymbol interference, it's called uncorrelated jitter. The data is then transformed back into the time domain where the cumulative distribution function is used to isolate unbounded Gaussian RJ. The remaining bounded jitter is labeled NP-BUJ (nonperiodic bounded uncorrelated jitter). Whatever crosstalk that exists in the jitter distribution is isolated in NP-BUJ.

There is no a priori reason to believe that NP-BUJ is exclusively crosstalk. Engineers have a good idea of what's going on in their systems, so it's not much of a stretch to equate NP-BUJ with the horizontal eye-closure caused by crosstalk. This analysis is also performed in the vertical direction where noise plays the voltage equivalent role of jitter to get NP-BUN (nonperiodic bounded uncorrelated noise) (Ref. 5). Whether or not NP-BUJ and NP-BUN are exclusively caused by crosstalk doesn't affect the eye closure estimates at given BERs.

LeCroy extracts crosstalk residuals

At LeCroy, Marty Miller's approach is quite different. Miller said, "My belief is that the most important aspect an oscilloscope brings to the table is noise analysis. Crosstalk is signal-to-noise degradation-radiative effects with variations and fluctuations that you can see on an oscilloscope." Miller's approach is to automate a detailed examination of the victim waveform.

The approach starts with the test patterns. "It's important in any crosstalk analysis," Miller said, "that the patterns transmitted on the aggressors not only be distinct from the victim test pattern but have a different length." If the aggressor pattern is the same length as the victim pattern, then crosstalk will appear to be correlated to the data. If the aggressor pattern is precisely the same pattern as the victim pattern, not only will the crosstalk be correlated, but the crosstalk and the signal can exhibit constructive and destructive interference, or beats.

Miller likes a short test pattern on the victim and long patterns on aggressors. This makes it possible to fully evaluate each bit in the victim pattern subject to a representative sample of crosstalk impairments.

The next step involves evaluating the shape, or trajectory, of each symbol in the victim test pattern. By comparing each repetition of a symbol, Miller's technique separates the crosstalk impairment into a waveform of its own that he calls crosstalk residuals. The residuals are what you would see on the victim lane if you could turn off the victim signal and look at the crosstalk alone on that lane. The process is independent of any phase relationships between the victim and aggressors, so it doesn't matter whether or not the victims and aggressors are frequency locked.

The final result is a novel construct Miller calls a "signal-centric contour plot" (Figure 5). This plot is a combination of the trajectories of every symbol in the test pattern with all the jitter and noise as well as the actual crosstalk distribution portrayed as a fluid cumulative distribution function from which the probability of a deviation from the mean waveform can be calculated. It's like a generalized BER-contour plot and provides eye-closure estimates to any BER, with crosstalk properly included in the calculation.

Crosstalk, Figure 5

FIGURE 5. A signal-centric contour plot contains a superposition of the trajectories of all symbols in a test pattern, including all sources of noise and lines of equal BER, here at 10-6, 10-9, 10-12, and 10-15, with crosstalk (left) and without crosstalk (right). Courtesy of LeCroy.

 
Agilent opens the catalog

Agilent applications engineer Stephen Didde commented that, "The error detector of a BERT [bit-error-rate tester] has 100% data capture. It's the only test equipment that actually measures eye opening at very low BER. No matter what signal impairments there may be, crosstalk from different lanes, power-supply interference, or other unknown problems, the BERT measurement can be trusted."

In addition, Agilent is bringing more of its test equipment to home in on the problem. Sleigh said, "You shouldn't rely on one instrument to identify and determine root causes for crosstalk. In addition to time-domain waveform analysis using an oscilloscope, other tools such as vector network analyzers and simulation tools such as ADS [Agilent's Advanced Design System software] can provide insight into crosstalk issues early in the design process."

Mike Resso, signal integrity measurement specialist in Agilent's Component Test Division, added, "Crosstalk is ultimately a problem of common-to-differential mode conversion. The differential S-parameters of a victim lane indicate how susceptible it will be to electromagnetic radiation within the board, regardless of its source. In fact, not only can the magnitude of crosstalk be determined, but the position where crosstalk is picked up by the lane can be located" (Ref. 6).

Performing a complete calculation of crosstalk would require a scattering matrix with two elements for every trace, four for each differential pair of traces. A 10x10-Gbps differential system would require a 40x40 scattering matrix, which would require 1600 S-parameter measurements. No system is yet capable of making these measurements at that scale in an automated fashion. Resso said that Agilent engineers can measure six complete channels, a 24x24 scattering matrix that can account for a victim and up to five neighboring aggressors. He added, "The calibration takes about 15 minutes as well as the measurement, and the resulting file is 37 megabytes. To do more than 24 ports is definitely feasible, but we would need to think about managing the huge data sets."

In July, Agilent is releasing extensions on its EZJit jitter-measurement software. With that release, Agilent will provide amplitude-analysis and jitter-analysis algorithms that distinguish crosstalk from RJ and random noise. According to Agilent, these enhancements will improve total-jitter estimates in the presence of crosstalk.

Can crosstalk be corrected?

When high-speed serial buses reached 5 Gbps, it seemed impossible to accommodate BERs of 10-12 and lower without replacing standard printed-circuit-board materials with expensive new media that are less sensitive to intersymbol interference. But engineering innovation rose to the task, leading to the development of different equalization schemes that let receivers identify logic levels at the desired BERs in highly impaired signals whose eye diagrams are completely closed.

Will some sort of signal compensation technique save the day? Zivny said, "I've heard techniques discussed going back to telegraph days: echo cancelation and hybrid circuits." Sleigh added, "Once we've developed techniques to identify crosstalk, we'll have the understanding we need to create techniques for solving the problem. Right now, we're working on identification."

"Only compensation applied at the transmitter is likely to help, because that's where we know what's coming," Miller said. He concluded, "Crosstalk is going to provide a never-ending discussion as we learn how to perform these analyses, much like jitter analysis has done over the last 15 years." T&MW

REFERENCES
1. Derickson, D., and M. Müller, eds., Digital Communications Test and Measurement, Chapter 2, Prentice-Hall, 2008.

2. Stephens, R., and B. Fetz, et al., "Comparison of Different Jitter Analysis Techniques With a Precision Jitter Transmitter," DesignCon 2005, bit.ly/JqQJYa.

3. Chow, D., and R. Stephens, "Methodology for Jitter Measurement Correlation and Consistency," Proceedings from DesignCon 2005.

4. Stephens, R., "Jitter Analysis: The dual-Dirac Model, RJ/DJ, and Q-Scale," Agilent Technologies White Paper, 5989-3206EN, 2004. bit.ly/L8CxG6.

5. Agoston, M., and P. Zivny, "Accurate Analytical Model of Bounded Uncorrelated Jitter and Noise Improves the Accuracy of Crosstalk Impaired Link Evaluation: Theory, Validation, Practical Results," Proceedings from DesignCon 2012.

6. Resso, M., and E. Bogatin, Signal Integrity Characterization Techniques, International Engineering Consortium, 2009.

Ransom Stephens, PhD, specializes in the electrodynamics of high-speed serial technologies. After 15 years making precise measurements of rare signals buried in noise at labs across the US and in Europe, he left his position as a physics professor, directed advanced technology for a wireless web startup, worked for Agilent Technologies for five years, and for the last seven years has been a consultant, science writer, and novelist. ransom@ransomsnotes.com.

Loading comments...

Write a Comment

To comment please Log In

FEATURED RESOURCES