Reading time ( words)
During DesignCon, I met with Cristian Filip, a senior product architect with Mentor, a Siemens business. Cristian had just received word that his paper had won a DesignCon Best Paper award—his second such award in three years. I asked Cristian to discuss his paper and how this technology can help improve manufacturing yields at high volumes.
Andy Shaughnessy: Congratulations, Cristian. You’ve written a paper that just won a DesignCon Best Paper Award. Can you tell us about yourself and the paper?
Cristian Filip: Sure. I joined Mentor in 2014. I work in the product architecture group and am involved mostly with analysis solutions. Over the last few years, I’ve co-authored about six DesignCon papers—two of them winning the Best Paper Award. One paper won in 2016 jointly with Vladimir Zdorov—one of our developers. In 2018, it was a collaboration between Hamburg University, Cisco, and Mentor.
Both of those papers that won the Best Paper Award centered around high-speed serial links, running at 28 gigabits per second. And they used channel operating margins as a figure of merit for estimating channels performance. In 2017, I had another DesignCon paper on the same area of interest, and we showed how to optimize channels while accounting for manufacturing reliability. At that time, we used design of experiments (DOE) as a method for optimization along with artificial neural networks. As a next step in the paper from 2018, we explored some improvements to the method. This time, we introduced the polynomial chaos expansion with the stochastic testing method to improve our way to optimize those links.
In this paper, "Efficient Sensitivity-Aware Assessment of High-Speed Links Using PCE and Implications for COM," we basically compared three different methods. One was based on Monte Carlo sampling, the second on DOE, using the response surface method, and the third on PCE. We showed that this new method not only reduces the number of required simulations from thousands to about 50 but also the simulation run time.
There is an additional benefit to this method because it provides better insight into what can affect the performance of the channels. This new method based on the polynomial chaos expansion with stochastic testing has the advantage of reducing the number of experiments or simulations required for channel modeling and channel optimization. It also provides additional information about the performance of various components of the channels, which are very useful in the debugging process and improving the overall performance of the channel.
Shaughnessy: How does this all work?
Filip: We took an example to show how we can apply this method based on a 100GBASE-KP4 for a channel, which is basically a backplane and two different blank cards, and we showed how various routing parameters, material parameters, and cheap I/O characteristics affect the performance of the channel. We also showed how we could optimize the link by using the proposed method.
Further, we compared the results from channel operating margin against the traditional diagrams based on eye height and eye width opening and showed that they actually correlate very well. As a basis for this new method, we started by using an open-source, publicly available library, and we made some improvements to that method. This method can be applied by anyone; it’s not related to specific tools or flow, so it’s very flexible and proved to be very efficient in our experience.
Shaughnessy: So, you’ve optimized a way for people to deal with high-speed links?
Filip: Correct, especially in high-volume manufacturing where we define the failures per million. We’re always looking at minimizing the number of failures per million. If you’re going to build a million units, you need to figure out how many will probably fail in high volume and minimize that so you can improve the yield.
Shaughnessy: And the variability seems to be a big thing too.
Filip: Yes, it is. The good thing is we typically know the distribution for the various manufacturing processes. In general, they have either Gaussian or uniform distribution, and we can apply the distribution to predict the performance of the channels. We create a model that allows us to apply a Monte Carlo type of analysis to further estimate that performance of the channel in high volume statistically.
Shaughnessy: Cool. That sounds like a good paper. Is there anything else you want to add?
Filip: I want to give a special thanks for our collaborators from Hamburg University and Cisco. It was a great experience working together with seven people from three different organizations spread across the globe.
Shaughnessy: Being that this is your second award since 2016, you seem to know what you’re talking about.
Filip: I hope so. I also want to thank Mentor for providing the opportunity to participate in this type of competition. DesignCon a great conference where so many talented people come together and exchange lots of interesting technical experiences.
Shaughnessy: Thanks so much, Cristian.
Filip: Thank you, Andy.