Can we avoid COMmitting the same mistakes we made with jitter?
COM (channel operating margin) combines multiple measurements into one signal-to-noise-like figure of merit, analogous to ENOB (effective number of bits), used to characterize analog-to-digital converters. With COM, the greater the margin, the better the channel. Because COM is built from different measurements and includes the results of models, there are many ways for it to fail.
As we introduce this new observable, it's timely to look back on the legacy of how we handled a similar situation almost 15 years ago: jitter.
Compared to COM, jitter seems simple: the variation of the timing of signal transitions with respect to their ideals. It's easy to think of the distribution as a histogram of these timing variations. Figure 1 shows how an oscilloscope displays the jitter in an eye diagram now it uses a histogram to characterize the jitter.
Figure 1. Oscilloscopes can provide you with a jitter distribution (Source: Teledyne LeCroy).
The mistakes of total jitter
Peak-to-peak jitter, which has appeared on clock data sheets for generations, turned out to be inadequate. Jitter from random processes—predominantly from phase noise within the SerDes’ reference clock—varies over time; the longer you measure peak-to-peak jitter, the larger it gets. At this point, those of us driving the standards for high speed serial data technology followed Yogi Berra's advice: "When you come to a fork in the road, take it."
To get away from poorly defined, impossible-to-reproduce peak-to-peak jitter, we made the well-reasoned choice to incorporate BER( bit error ratio) in the definition of a new quantity, TJ(BER) (total jitter defined at a BER). TJ measures the eye closure at a given BER, that is, if TJ(BER) is less than the bit period for the BER specified, then you have some jitter margin and you should be okay—which is the desirable feature of a peak-to-peak measurement.
Sounds great! Drinks all around, right?
Well, because the BERs we cared about were very low, 1E-12 to 1E-18, TJ(BER) turned out to take a very long time to measure and the only equipment that could measure it, BERTs (bit error ratio testers) were really expensive and not all that useful for diagnosing other problems in the lab. So, we developed techniques to estimate TJ(BER) from measurements that could be made quickly, but this led to a huge problem; what you might call a clusterjitter. The extrapolation techniques relied on the abilities of oscilloscopes to measure separate components of jitter: RJ, DJ, ISI, PJ, DCD, and several abbreviations that you either know or wish you could forget.
Different test-and-measurement companies developed different approaches and the measurements disagreed badly. From 2000 to well into 2006, equipment from differing T&M companies—companies you know and rely on and who make great equipment—differed by at least 30% and frequently more than 100%. It wasn't until 2004 that anyone assembled a system that could accurately distinguish which results were right and which were wrong.
The problem came from how we chose to stir RJ, DJ, ISI, PJ etc. into the goulash from which we estimate TJ(BER). By building TJ(BER) from a goulash of inter-dependent quantities (Figure 2) we made it unreasonably difficult to determine where mistakes were being made in the measurements that went into the goulash. You see, if you change the amount of ISI, you also change RJ. Add crosstalk, and all bets are off.
Figure 2. Debugging goulashes of inter-dependent variables is hard and COM is a goulash.
As we advance from a few Gbits/s to 10+ Gbits/s, ISI (intersymbol interference) is the biggest problem. ISI is caused by the frequency response of a channel; it shifts a signal’s amplitude and timing by amounts that depend on the sequence of transmitted symbols. Which brings us to COM and the possibility that history is repeating itself.
Back in 2003-2004, a group of us at what was then Agilent Technologies built a precision jitter transmitter and put all the jitter analyzers available from all the test & measurement companies to work. We spooned precise amounts of RJ, PJ, ISI, and DCD into the goulash (random jitter, periodic jitter, intersymbol interference, and duty cycle distortion), made hundreds of measurements and then… I sat down in the peace of my living room to determine the best techniques. The biggest problem I had was that there were too many ingredients. Even though I knew with excellent precision how much jitter of each type went into each signal, there was no way to determine why different techniques failed (and they all failed!).
To make sense of the measurements, I had to do basic science: start with the cleanest, lowest jitter system we could build and then inject one type of jitter at a time, compare results, and then add a second type of jitter, and so on. It took six full time weeks to analyze the data and discover which techniques were accurate and why. It took over two years before the T&M industry started to converge to results within 10-15% of each other—the entire industry was confused for more than five years and at least one company went out of business for its failure to produce accurate results.
Turning back to COM, having a single figure of merit that emerges from several causes can allow design flexibility without sacrificing interoperability. That combination—flexibility and interoperability—is a technology standard's Holy Grail. No one wants a standard to resemble technological socialism, but everyone wants it to assure interoperability. Bingo! COM.
Unless, we run into the same problems that we did with jitter.
COM is given by the ratio of the signal amplitude to the collective signal impairments. It includes impairments from the signal channel, all crosstalk aggressor channels, and every other source of impairment that the standards committee could think of.
Like TJ(BER), the accuracy of a COM measurement depends on the precision of its ingredients. Unlike TJ(BER), measurements of COM also depend on models. Here's a list of the COM-goulash's ingredients:
- Channel S-parameters for the channel in question and all crosstalk aggressors
- Transmitter and receiver (SerDes) package models
- Receiver 3 dB bandwidth
- Max/min values for transmitter equalizer coefficients
- CTLE (continuous time linear equalizer) gain
- Peak-to-peak differential output voltages for the victim, near-end, and far-end aggressors
- Level Separation mismatch ratio (for PAM4 applications)
- Transmitter signal-to-noise ratio
- DFE (decision feedback equalizer) length and limits on coefficients
- Dual-Dirac model-dependent peak-to-peak DJ
- One-sided noise spectral density
This list makes the recipe for TJ(BER) look comparatively sparse—yikes!