Channel Operating Margin Isn’t So Bad

February 26, 2019 Jason J. Ellison

 

What is a COM anyway?

Channel Operating Margin, or COM, isn’t well understood. Since it isn’t well understood, many people doubt it really means anything. After all, how can channel quality be represented by just one number in decibels? It turns out, COM is actually the latest evolutionary step in a long progression of channel validation techniques using eye patterns. This blog will trace the evolution of COM back to its roots and put meaning to the infamous COM metric.

 

The First Channel Operating Margin: Eye Patterns

Let’s start with eye patterns. Eye patterns are a way to look at a long stream of serial data. Back before Keysight ADS and PyBERT [1] [2], an eye pattern was measured with a digital sampling oscilloscope or a real-time scope. In the eye pattern window, the y-axis units are voltage and the x-axis units are time spanning two unit intervals. A unit interval, or UI, is the amount of time for one bit to pass. Thus within two UI of time, you can center one bit data on the screen with half of a bit of margin on either side. However instead of viewing just one bit, all of the bits are overlapped, one at a time, until the entire stream of serial data is on the screen. The signal quality is quantified by the size of the hole in the middle. If the eye pattern looks really good, you might hear an engineer say, “You can drive a truck through that eye!” The most common ways of quantifying the opening are width, height, or area. The crossing of the eye at the DC point is the jitter, and the jitter is typically measured statistically with a histogram.

Figure 1. Example of a serial bit stream.

Early channel specifications, and in some cases passive component specifications, used something called an eye mask for the pass/fail criteria. An eye mask is usually a diamond-shaped area defined by an eye width and height. A passing eye has only so many detected samples, or hits, inside the eye mask. The patterns of ones and zeros are dictated by the standard and are usually pseudo-random bit sequence or PRBS pattern. You can basically bank the patterns into two categories: before 10 Gb/s and after 10 Gb/s. Before 10 Gb/s 8b10b encoding was used in most systems and PRBS 7 was the appropriate pattern. When 10 Gb/s was introduced by the IEEE in 802.3ba, the encoding switched to a 64b66b scrambler and PRBS 31 took over. Even today at 112 Gb/s, PRBS 31, or QPRBS 31, is still the standard pattern most use.  

 

Statistically Speaking

Chronologically after measured eye patterns, StatEye is the next method of qualifying passive channels, and it was heavily used by the OIF. The idea behind StatEye is explained in detail here: [3] In brief, StatEye predicts eye patterns using a pulse response of a system. A pulse response is the time-domain response of a system excited with a one-UI square pulse, and the system is a passive channel including equalization. The equalization technologies available in StatEye are FFE, CTLA, and DFE. The transfer function of a system is gathered from S-parameters. Since channel S-parameters can be simulated, StatEye is an efficient way to try many channels and equalization settings to see what works. All the while, the eye mask is the pass/fail criteria using the statistically predicted eye opening.

Somewhere between StatEye and COM, peak distortion analysis (PDA) became somewhat common. The method is well documented by Heck and Hall in Advanced Signal Integrity for High Speed Digital Designs [4]. In summary, it uses the same pulse response as StatEye, but it’s output is simply the so-called worst case eye opening. PDA doesn’t make up any data, and that is the reason I personally like it. I’ve implemented it myself and found PDA predicts worst case eye patterns with high confidence. However, PDA and StatEye do not include the impact of the transmitter and receiver in the channel, and you need to find the best equalization setting manually.

Figure 2: Example of an Eye Pattern in blue and PDA in dashed black.

 

Enter COM

COM was developed as part of IEEE 802.3bj, 100GBASE Ethernet, and added the IC imperfections to the simulated channel. It is easier to use and more widely adopted than StatEye, and is the defacto channel quality prediction tool today. As I already mentioned, COM builds from StatEye and adds several new noise sources. Specifically, the noise sources are  loss from the IC, IC package reflections, IC related jitter, and a lumped Gaussian noise source for all other stuff happening in the IC, like crosstalk. The implementation of COM is found in IEEE 802.3 Annex 93A [5].

Most of the math behind COM is simplified by the standard body as much as possible. For instance, S-parameter concatenation is reduced to algebra instead of conversions From S-parameters to ABCD-parameters or T-parameters and matrix multiplication. The most difficult equation is calculating the probability density function (PDF) of ISI related noise, but after a few tries it’s really not too bad. There are some omissions that are considered implementation specific, such as how to ensure 32 sample points within every UI of data, but these details can be found within the open-source code provided freely by the IEEE [5].

COM finds the best case scenario for a given channel using a set of possible equalization settings. This is accomplished by sweeping all equalization settings and calculating something called the Figure of Merit (FOM). The equalization setting that produces the best FOM is used for the remainder of the calculations. Once the PDFs of all noise sources are calculated, the noise at a detected error rate (DER) is identified. The DER is the desired bit error rate (BER) for the system, and is determined by what Forward Error Corrections (FEC) technique, if any, is being considered. The available signal is determined by the pulse response voltage at a specific sampling point. The available signal is divided by the noise at the detected error rate (signal to noise ratio), and this number is converted into decibels. Voilla! COM! See, it actually does have meaning. 😃

The settings used for COM are determined by the available IC technology. The level of IC technology is agreed upon by industry leaders such as Intel, Broadcom, Mellanox, Fujitsu, and many others. In other words, an IC using technology implemented in COM should be able to function in working channels as predicted by COM. Obviously, this is very powerful since the standard has now (finally) put some of the channel ownership onto the IC vendors.

Even though it sounds like COM is this utopia of channel prediction, it does have limitations. Since it is one set of settings for all systems being considered by the standard, it does not predict the performance of any one IC by itself. To get measurement correlation, you need to adjust the COM settings for each individual IC. Further, COM neglects any noise contribution from skew. Fortunately, a DesignCon paper by Jason Chan addresses this deficiency, and I hope to see updated COM scripts utilizing his ideas in the future [6].  

 

Conclusion

To sum it up, COM is not so bad. It is a very logical next step in the evolution of channel analysis, and it makes channel evaluation relatively easy. I’m very grateful that the authors of COM were kind enough to freely release and support the MATLAB code. I hope to see COM implemented and improved by other signal integrity engineers in the future. Who knows, maybe we’ll see a Python or Octave implementation someday.

All figures were created with GNU Octave, https://www.gnu.org/software/octave/.

 

References:

[1] Keysight ADS landing page, https://www.keysight.com/en/pc-1297113/advanced-design-system-ads?&cc=US&lc=eng

[2] PyBERT landing page, https://pypi.org/project/PyBERT/

[3] A. Sanders, M. Resso, J. Ambrosia, Channel Compliance Testing Utilizing Novel Statistical Eye Methodology, DesignCon 2004, http://www.ece.tamu.edu/~spalermo/ecen689/stateye_theory_sanders_designcon_2004.pdf

[4] S. Hall, H. Heck, Advanced Signal Integrity for High Speed Digital Designs, Wiley 2011

[5] IEEE 802.3 Ethernet Working Group landing page, http://www.ieee802.org/3/

[6] J. Chan , G. Zheoff, Mode Conversion and Its Impact on 112-Gbps PAM4 Systems, DesignCon 2019.

About the Author

Jason J. Ellison


Jason J Ellison received his Masters of Science in Electrical Engineering from Penn State University in December 2017.

He is employed as a signal integrity engineer and develops high-speed interconnects, lab automation technology, and calibration technology. His interests are signal integrity, power integrity and embedded system design. He also writes technical publications for journals such as “The Signal Integrity Journal”.

Mr. Ellison is an active IEEE member and a DesignCon technical program committee member.

More Content by Jason J. Ellison
Previous Article
The Benefits of Using a Professional PCB Design Service Bureau
The Benefits of Using a Professional PCB Design Service Bureau

Learn about the Benefits of Using a Professional PCB Design Service Bureau straight from the experts.

Next Article
End-to-End Tracking of Your PCB Design
End-to-End Tracking of Your PCB Design

Learn how to track your design changes and keep them transparent for the whole team.

×

Trade In Your Outdated Design Tool & Unlock 45% Off Altium Designer. Sign up and an Altium representative will be in contact shortly.

First Name
Last Name
Company Name
Phone Number
Country
State
Postal Code
Acknowledging Altium’s Privacy Policy, I consent that Altium processes my Personal Data to send me communications, including for marketing purposes, via email and to contact me by phone.
!
Thank you!
Error - something went wrong!