Steinberger Talks PAM4, the Next Generation of Modulation

Estimated reading time: 8 minutes

This year at DesignCon, I interviewed our old friend Michael Steinberger, SiSoft’s lead architect for serial channel products. Steinberger is always a great interview; he breaks down complex signal integrity simulation technology in ways that are simple and often humorous. I caught Steinberger in between panel sessions and asked him to discuss a paper he presented at the show, and some of the challenges his customers are facing.

Andy Shaughnessy: Michael, tell me about the paper you presented at DesignCon.

Michael Steinberger: Well, we are talking about choosing a combination of transmit and receive equalization for a given channel. This year, one of my primary assignments has been to develop to maturity a capability to automatically determine the optimum combination of transmit and receive equalization for a given channel, which is actually a problem I started working on back in March of 2002.

Shaughnessy: It's not always overnight.

Steinberger: No, this one wasn't. If somebody wants to try to reproduce it, all I want to do is watch because I found it to be quite a challenging assignment. What we were talking about yesterday was the results we got from automated equalization; but what we actually presented in the paper was more walking people through how to choose equalization manually, so that they had some sense for what the tradeoffs were and they could understand, from an engineering judgment point of view, what kind of results they were getting out of the optimizer.

We talked mostly about choosing transmit equalization taps. That was fairly straightforward and had some waveforms associated with it. We did have a little bit of fun, with a hardware demonstration in the middle of that. We have an algorithm for choosing the recovered clock time from a pulse response. We call it the hula hoop algorithm. What does that mean? Well, what it means is that, if you have a hula hoop of the diameter of one UI, you can drop it down over the pulse response. Where the hula hoop settles level, the center of the hula hoop is your recovered clock. It's as simple as that.

I brought with me from home a coil of copper tubing and formed that in the shape of a pulse response and I brought with me a hoop about 18 inches in diameter. So I had Barry Katz, president of SiSoft, hold up the pulse response and Donald Telian, the consultant with SI Guys, dropped the hoop on top of the pulse response. Here in a DesignCon presentation, we had a hardware demonstration. I think that went fairly well and it certainly is information that engineers should have in their gut when they're looking at these results to make sure they understand what's going on and understand what trade-offs they're making.

Shaughnessy: What sort of challenges are these attendees facing? What's driving them to try to find out more about this?

Steinberger: PAM4.

Shaughnessy: I've seen a lot about PAM4 at DesignCon this year.

Steinberger: It's got everybody's heart going pitty-pat. It has a different set of challenges. One set of challenges is that the eyes are a lot smaller and so it's an awful lot harder to tune the equalization in to where you can start opening the eye and see whether or not you're getting a wider eye or not. Especially when we introduced PAM4 in our product and we said, "Oh, by the way, we've got optimization." It wasn't quite as mature as it is now. A bunch of customers were just all over it because they had PAM4, but what were they going to do with it?

Well, with this feature we've developed, you enable the feature and it will automatically choose the equalization for your PAM4 and then you can do the performance analysis. Now you're doing the performance analysis with PAM4 with an open eye. You can actually see something. PAM4 has really driven the interest in this particular capability.

Shaughnessy: What are some of the bigger changes you're seeing from your viewpoint?

Steinberger: I have a list of technologies that I'm putting in place, and so I look at change from the point of view of those technologies. Certainly this optimization/equalization is something that's new to most people here and so that would have to count as a change. PAM4 of course is a change. One of the changes that's coming up, and it isn’t just an opinion, I'm confident about this, is that PAM4 is going to end up being married to error correction.

In the studies I did comparing PAM4 to NRZ, when I had enough margin that both PAM4 and NRZ could achieve a low bit error rate, NRZ had a definite advantage even if I doubled the symbol rate, etc., etc. As I reduced the amount of margin available to where I started getting an error rate, PAM4 started developing an advantage. That was new to me. I hadn't expected that at all.

Now what that means is that in order to capitalize on the advantages of PAM4, you're going to need error correction codes. They're going to end up being married to each other by the time you're done. Trying to do PAM4 at a raw bit error rate of 10 to the minus 12th or below is probably not going to be terribly attractive. Doing PAM4 at a raw bit error rate of 10 to the minus 9th, a whole lot more attractive and maybe at that point it's going to prove in for a number of applications.

If you're at 10 to the minus 12th, 10 to the minus 9th, or 10 to the minus 7th, and you say, "I want to get down to the 10 to the minus 15th so I'm just going to pop this error correction code on it." Well, not so fast, because a lot of the analyses that have been done on error correction for serial channels use assumptions that are just plain wrong. They are assumptions that came from band limited transmission with added white Gaussian noise and that simply doesn't apply to high-speed serial channels.

With a high-speed serial channel, the dominant impairment is inter-symbol interference; and error correcting codes behave a lot differently in the face of inter-symbol interference. This is something that was brought to our attention by a customer who was experiencing this in the lab. I'm now able to reproduce those results. I demonstrated it in our FEC tutorial last year. People aren't really ready to hear that yet. That's one of the next bombs that's going to hit. Now, when it hits, call me because I know how to fix it. I'm not going to fix your error correcting codes necessarily, but I am going to be able to give you accurate predictions of what the performance actually will be. That's one of the big bombs that's coming.

Another thing I’m working on is via modeling. I've been working on it for a long time. I've made some breakthroughs recently, so my capability is going to be coming online in the next year or so. That capability is going to allow you to analyze vias in orders of magnitude less time than what you're doing now, with results that match measured data. I don't know whether that's going to be a big change, but it's a change.

Our ability to scan the crosstalk in interconnect networks is going to be dramatically improved both because of some of the work I've been doing and because of some work that one of our customers has done. We're collaborating with them. That's going to bring us some very much improved capabilities. I guess that's enough to look forward to for one year.

Shaughnessy: Have you caught any papers or sessions that really impressed you?

Steinberger: I've seen a lot of brilliant work, most of which I didn't fully understand. I just got blown away by a presentation on conductor roughness. That's another area where we're prepared to support whatever work is being done, but that's not an area where right now we're in a position to contribute directly. I thought that was an interesting piece of work, even though it was too much for me to absorb in the time given. I think that's probably the standout of what I've seen, but then I've been in orbit bouncing between papers and there are a lot of great papers that I haven't had a chance to even poke my head into. There's a lot of great work being presented here.

Shaughnessy: How about the show floor? It seems to have expanded over time.

Steinberger: Unfortunately, the show floor is something I hardly ever get around. I've got a well-respected colleague at one of the connector vendors who was showing me a 56 gigabit PAM4 right there on the show floor. Not a bunch of external test equipment. There it was on a PC board running across the backplane, of course with his connector. For years, people have been telling me what couldn't be done and I've been saying it could. Now I go to the show floor and there it is.

My history with serial channels started back in 1995 when I told a room full of people at Bell Labs that we needed to be designing serial channels at 622 megabits and the room full of people told me all they were ever going to need was 155 megabits per second. So I'd love them to see that 56 Gbps I was just looking at the other day. [Laughter]

Shaughnessy: Thanks for your time, Michael. Great talking to you, as always.

Steinberger: Thank you, Andy.