PAM4 isn't the perfect solution to the bandwidth problem posed by NRZ, but it's the best one we have, at least for now.
In January, DesignCon's "Case of the closed eye" panel has met for its 2017 edition, which marked 15 years since the panel took shape.
Pavel Zivny from Tektronix also took to the podium and continued the pattern discussion. "The test people were whining that PRBS31 is too long of a pattern," he said. "We tried to convince the industry to drop it. We got our wish." He noted that SSPRQ patterns can produce sufficient stress on systems, but with shorter patterns than PRBS31. "SSPRQ has the right properties. It has the right stress points and has similar statistical properties as PRBS31 (over 2 billion symbols), but is shorter. Unfortunately, SSPRQ patterns break everything."
My chip broke so it (the test pattern) can't be right, he's heard from chip makers. But, SSPRQ is now part of communications standards and Zivny said engineers aren't sure what to do. Zivny argued that some are advocating for PRBS13Q as opposed to PRBS17Q, but as Marty Miller from Teledyne LeCroy showed, the results can differ. But, just as the industry overcame difference in test results from different equipment, this problem will also work itself out. Perhaps we'll get an update next year.
Zivny then moved on to mask testing, asking engineers in the audience if they have tried correlating the results of mask testing to a link's performance. Few raised their hands. "There is some correlation between mask testing and BER, but it's not great," he said. "We need different approaches." One such approach is impulse testing, which is used to derive channel operating margin (COM). The advantage of COM testing is that you don't need equalisation and error correction. But, he noted that we still don't know how well COM testing correlates to system performance. His other problem with COM is that it's hard to look at the test result and understand the problem. For example, COM doesn't tell you if a signal's rise time is too slow, but eye diagrams do.
Zivny brought up TDECQ measurements (Transmitter Dispersion Eye Closure for PAM4), which come from the optical side of serial communications. He added that while a closed eye results in catastrophic failures and wide open eyes result in good links, a problem occurs when trying to correlate eye openings to link performance somewhere in between.
Stephens wrapped up by reiterating that having the best eye opening doesn’t always give the best BER after applying FEC. Ideal equalisation might give best raw BER but not the best BER when using FEC. Thus, using PAM4 and FEC may solve the bandwidth problem, but it opens new problems.
He then asked for questions. "Are there PAM mask tests?" asked one engineer. LeCheminant responded by noting that eye masks were shot down by optical standards groups, citing no correlation between eye masks and system performance. Instead, fiber-optics engineers rely heavily on TDECQ.
"How do you know if TX signal is good?" asked another attendee. "How do you handle the nuances? That's how we stay employed," replied Marlett. "The system is sensitive to channel loss deviation. As a result, these measurements are still in their infancy and are in flux." Zivny added, "Differences are unavoidable." Standards should not define how to design receivers. You could have two that comply but function differently.
It's clear from the 2017 closed eye panel that PAM4 isn't the perfect solution to the bandwidth problem posed by NRZ, but it's the best one we have, at least for now. This discussion will likely continue at DesignCon 2018.