COM – It's a Fine Line

Thursday, February 25, 2021

Our previous look at COM established how, why, and where COM fits into serial link SI analysis. Furthermore, we explained how COM defines a "line" to help evaluate our channels - much in the same way the lines on a tennis court help define the game. While debate might occur when the ball lands on the line, it never occurs to us to ask if the line is in the right place. Nevertheless, regarding COM, in this post we ask that question: "Is COM's line in the right place?".

Personally, I prefer to simulate the actual system phenomenon that determines signal performance, such as the eye opening at the Rx latch or the active channel's BER. Does COM simulate that? Kind of, but not really. As we discussed last time, COM blends elements of both Passive and Active analyses to develop a meaningful result - a single number even. And a single number is helpful because our binary minds like pass/fail and yes/no answers better than SI's more common answer "it depends".

This post compares COM and simulated eye openings and BERs against two system variables: (1) Backplane Length, and (2) Crosstalk Aggressor Clearance. We hope and expect to see decent correlation, or at least tracking. Let's have a look.

A Relevant System Model

Though COM had bounced around Intel previously, it entered the public domain through the IEEE 802.3bj 100GBASE Specification. As such, for our analysis we construct the 25 Gbps channel shown in Figure 1. The channel includes a backplane at center with line cards on each end, interconnected through coupled trace and connector models. The channel model is also configured with sweep variables to simulate variations in route length and crosstalk clearance.

Figure 1: 25 Gbps Backplane Channel

Figure 1: 25 Gbps Backplane Channel

 

SiSoft QCD is used for drawing the circuit, running the simulation, and post-processing the results. QCD integrates with COM in 3 different ways: (1) QCD GUI drives the IEEE Matlab COM code, (2) IEEE COM code calls QCD's S-parameter files from Matlab, or (3) QCD's native COM. Watch the video at the bottom of this post to learn more about these three options. Here we use option 1 to invoke the industry-standard implementation. Let's have a look at how it performs.

Channel Length Simulation Results

SI Engineers are commonly asked "how long can this channel be?", so we begin there. Figure 2 shows that both Rx latch eye height (upper, blue) and COM (lower, red) decrease with backplane length as it ranges from 8 to 20 inches (X axes).

 

Figure 2: Backplane Length versus Eye Height (upper) and COM (lower)

Figure 2: Backplane Length versus Eye Height (upper) and COM (lower)

 

Figure 2 further illustrates the difference we're focused on: only the COM plot has a "line" to help decide what works and what doesn't. COM declares values above 3 dB to be "good" while values below 3 dB are "bad". Based on that assertion, and our engineering judgment, we might conclude 16" is our maximum backplane length. Is that a good conclusion? ...or overly constrained? We're not sure yet, but at least COM is helping us move towards an answer.

While Figure 2 shows eye height is decreasing with length, we are left wondering when and where the point of failure might be. Knowing that modern receivers fail somewhere between 50 and 10 mV doesn't help because the plot translates that into a wide range from 10" to 30". And so we query the Rx component's supplier for a milli-Volt "line", yet often no answer emerges. Worse yet, sometimes it seems like no one knows. Hmmm, no "line" to use except experience and judgment. This scenario leaves us with a similar uncertainty using either COM or Rx latch eye height, but, again, at least COM provided a line.

Digging deeper, Figure 3 shows a plot of BER versus COM. Although some of the pleasing linearities of Figure 2 are gone, we observe our 1e-12 BER "line" generally agrees with COM's 3 dB - within ~10%, which is rather stunning given all the variables at play. Now we're inclined to ask "yes, but how good is that BER simulation?". A great question, which is further addressed in this paper (pages 7 and 8) and industry debate.

 

Figure 3: COM versus BER

Figure 3: COM versus BER

 

The plots in Figures 2 and 3 demonstrate COM's ability to provide a meaningful result. While we'd expect COM - or any serial link metric - to handle the length/loss trade-off decently, how well does it handle crosstalk? The dynamics of coupling seem more complex; perhaps beyond the capacity of a simple 3 dB metric.

COM and Crosstalk

Our next simulation tests COM's ability to capture and quantify crosstalk variation by varying the Rx Line Card's (Figure 1, at right) aggressor clearance from 20 to 40 mils. For this scenario, Figure 4 reveals an almost exact correlation between Rx latch eye height (upper, red) and COM (lower, black); it appears only the Y axes have changed. This correlation is almost too good, and we're left wondering if COM's mathematics somehow implemented an active analysis simulator. Furthermore, COM's line agrees with the industry-accepted 25 mil clearance required to constrain crosstalk to acceptable levels. The BER plot (not shown) is again not quite as linear, yet further confirms performance fails below 25 mils of clearance. Perhaps it's going too far to use Figures 2 and 4 to reverse engineer acceptable Rx eye height to be 40 mils. Or is it? Regardless, this data reveals that COM does a decent job of handling coupling.

 

Figure 4: Aggressor Clearance versus Eye Height (upper) and COM (lower)

Figure 4: Aggressor Clearance versus Eye Height (upper) and COM (lower)

 

In Conclusion

The data shown suggests COM's authors have baked in a significant amount of serial link variables and achieved a relevant Figure Of Merit (FOM). COM's on-going relevance as serial link technology advances is also ensured because it continues to get updated with new features. While it's my nature to be skeptical of simplistic answers, even a cursory review of COM's Matlab code and variable spreadsheet reveals COM is far from simplistic. Indeed, the only thing that's "simple" about COM is its answer. Bravo COM, and thanks for drawing a fine line.

But I am using the term "fine line" to reflect its double meaning. Is COM so robust I can abandon my proven methodology of simulating with AMI models against many other channel metrics? I will not be doing that anytime soon. While this post has shown a couple compelling correlation examples between COM and full-channel simulation, the engineer in me will need to see a few hundred (thousand?) more spanning the matrix of variables I care about. Particularly with SI tasks, there's a fine line involved in knowing what to use when. Nevertheless, COM is clearly on the list of relevant channel metrics. And QCD's integration simplifies the challenge of implementing COM correctly.

For details related to running COM within SiSoft QCD and additional simulation results, including the crosstalk BER plot mentioned but not shown here, download this AppNote. And here's a video demonstration of QCD's Matlab COM integration in use on a CEI 56 Gbps LR PAM4 backplane channel.

COM (Channel Operating Margin) Video Series

  • Learn about COM, the Channel Operating Margin metric and industry standard MATLAB® COM script
  • Perform design space exploration using QCD and COM
  • Perform COM analysis with QCD and IBIS-AMI models

Please enter your contact information below to continue to view this video.

Work or university email:

Donald Telian, SiGuys - Guest Blogger 2/25/2021



Did you like this post?

Sign up for our mailing list and we'll let you when we post a new one.