SiSoft Discusses Signal Integrity Drivers and Challenges
SiSoft has been developing cutting-edge signal integrity simulation and analysis software for over 20 years now. In the last few years, the company has launched a variety of innovative SI tools that help automate tasks that used to take weeks or months to complete. I recently interviewed CEO Barry Katz via email, and he detailed their customers’ challenges, and some of the market drivers in the world of signal integrity.
ANDY SHAUGHNESSY: SiSoft has rolled out a lot of new features in the 2016 software suite. Tell us about some of your customers’ challenges that lead to these new features. Let’s start with OptimEye.
BARRY KATZ: Integration and GHz signaling is moving signal integrity inside the chips, and radically changing the problem my customer is solving. While it’s still important to carefully route a PCB, achieving clean signaling is increasingly dependent on correctly configuring IC equalization. And that’s no simple task, because as we passed through 10 Gbps, customers were confronted with millions of equalization options. With so many choices, how can you determine which one is optimal? This is a very different problem than figuring out the value and placement of termination resistors (which have also been integrated). Because customers didn’t have a fast, automated way to optimize equalization configurations, they were adding extraneous components, buying expensive materials, and unnecessarily shortening connections and lowering data rates. There were a lot of growing pains.
OptimEye tackles the configuration optimization problem head-on. Our DesignCon 2016 paper[1] explains how one customer used OptimEye to remove dozens of components while making their links 25% longer with 60% better performance. This illustrates how important optimal equalization implemented as IC register values can be. SiSoft thrives on solving problems like this. It is perfect for us because link optimization is a system-level problem and we hold all the pieces of the puzzle to get it right; it can’t be solved efficiently by the designer of the transmitter, receiver, or passive channel in isolation of the others. Indeed, the millions of configuration choices illustrate that the best thing each designer could do was offer a range of options because they did not know how the connection would eventually be implemented in a system. So, because we see the whole system, it is our job to resolve the settings in an optimal way. And when you do that, a typical link can achieve 100% better performance, as we demonstrated in our OptimEye Webinar, with only a software change. So, you see what I mean when I say SI is moving inside the chips.
SHAUGHNESSY: What sort of time does this save a typical SI engineer?
KATZ: Months. Particularly when designing systems with dozens or even thousands of links. Think of it: thousands of links and each one has millions of configuration options. How are you going to solve that? Sounds like a great place for compute power and advanced algorithms to step in, right? Today’s SI engineer is sweeping through thousands of setting options, measuring the results, and eventually converging on an acceptable choice. That approach requires hours of labor and simulation to resolve a single link while OptimEye finds the answer in seconds. Increasing compute horsepower has enabled a few significant breakthroughs in the history of SI, and this is one of them. Unfortunately, in absence of an automated solution like OptimEye, about half the industry isn’t even trying to solve the problem and letting the ICs use power-up defaults which, because they’re slanted towards worst-case scenarios, deliver terrible performance on a typical link and sometimes don’t even work. So you see why we’ve made significant investment in OptimEye and are excited to make this technology available.
SHAUGHNESSY: Let’s talk about STATify. What were some of the drivers behind it?
KATZ: When we analyze serial links with OptimEye or otherwise, the active Tx and Rx models are typically implemented using the IBIS-AMI standard, which allows a variety of model styles that support different types of analysis. So, interoperability gets interesting when interconnected Tx and Rx models are of different styles, causing some analysis types to yield unreliable results, as we detailed in a paper at DesignCon 2015[2]. So you had to know what you were doing. But now with STATify, everything “just works.” That’s the good news at a high level. Customers can now run statistical analysis with models that don’t support it, and that is quite a few of them, because we “STATifyd” them. This not only enhances usability, but also has advantages when determining equalization schemes because equalized pulse responses are always available. There’s a lot you can tell from a pulse response when you know what you’re looking at. There’s a lot we could say about STATify, like the way it helps predict low-probability errors without simulating lots of bits, but that’s the basics. It’s a unique capability that we add to our list of firsts in the IBIS-AMI space.
SHAUGHNESSY: Your customers are dealing with huge amounts of data when sweeping a pre-layout solution space or simulating thousands of post-layout nets. Why don’t you explain how that’s traditionally been done, and then how it’s done with the data mining/management capability?
KATZ: In the past, SI engineers became pretty good at exporting data to Excel or writing custom scripts to filter, sort and visualize their results. And, when it makes sense, we output CSV data directly to your spreadsheet of choice. But for many SI tasks that doesn’t make sense. So we’ve taken a two-pronged approach: First, allow customers to customize which data is saved and in what order, and second, incorporate advanced sort, filter, and roll-up capabilities directly within our tools. My view is that if you can imagine ways you’d like to visualize your data, then you should be able to make it happen. The tools should even
challenge you to think about the data in new ways, and I think we’ve achieved that too. I really enjoy watching customers find new ways to plot and visualize data because that’s when important discoveries and breakthroughs are made. Of course, it’s great to see them performing more efficiently in their routine tasks too. And because a picture is worth a thousand words, we’ve created a series of on-line videos so customers can see how this (and our other new features) works. A video is worth a billion words, and so is data visualization and trend analysis done right.
SHAUGHNESSY: Tabbed routing is another issue we hear a lot about from engineers working with DDR designs. Why has tabbed routing been so tough to model, and how do you address that?
KATZ: That’s an interesting question, because our team took an alternative approach like we did with via modeling. In both situations, you might be inclined to think these structures require generalized 3D solutions. But, if you look closer you see fairly consistent structures getting designed and manufactured. We find that once we comprehend and model the relevant pieces of the structure we arrive at solutions that are orders of magnitude faster than conventional 3D solvers and correlate well with measurement. My team has actually published quite a bit in this space, including the paper, “Fast, efficient and accurate: via models that correlate to 20GHz,” and I view it as an arena where SiSoft really shines. Customers like the approach because analysis is fast enough to sweep through layout alternatives and derive the best solution for both signal performance and manufacturing. Our approach also enables large batches of post-route simulations to complete without a perceptible change when tabbed routing is used.
SHAUGHNESSY: What do you have coming up in the next few years or so?
KATZ: We have a lot of excitement over the future of our products. Enabling our customers to run more simulations smarter and faster with optimization and supporting advanced visualization and data mining techniques is key. SiSoft will continue to make significant investments in electromagnetic and algorithmic modeling as well as leverage our technology and expertise amongst our products. We see serial links moving towards speeds of 56Gbps with PAM4 modulation and detailed modeling of vias and crosstalk being critical to accurate analysis. We see statistical analysis becoming a requirement on both serial and parallel interfaces. One thing we will continue to do better than any of our competition is work closely and collaboratively with our customers and their suppliers to ensure we bring the capability they need for their next generation designs.
SHAUGHNESSY: I appreciate your time.
KATZ: Thank you, Andy.
References
- “New SI Techniques for Large System Performance Tuning,” by Donald Telian, SiGuys, and Michael Steinberger and Barry Katz, SiSoft, DesignCon 2016.
- “Understanding IBIS-AMI Simulations,” by Douglas Burns, John Madsen, Todd Westerhoff, and Walter Katz, SiSoft, DesignCon 2015.