Episode IV: EDA Dies
The IEEE/ACM's International Conference on Computer Aided Design Sunday afternoon, after 3 hours of mini-tutorials – 2 each hour running side-by-side, 6 tutorials in all – those giving the tutorials gathered together on the main ICCAD stage to participate in a panel. The panel was called: IC Design in 65 nm and Beyond: Evolution or Revolution? You've undoubtedly heard this question posed before. It's standard fare at cocktail parties and mixers whenever designers and CAD tool developers stand around chatting. But I would argue, no matter what you heard over a gin fizz, it weren't never so interesting as what you might have heard if you'd been with the other 125+ people facing the main stage at 5:30 last Sunday afternoon in San Jose. ****************** Rob Rutenbar from CMU was moderating. He was standing in for IBM's John Cohn, who had apparently been called away to create a PowerPoint presentation for his boss. Rob was more than adequate. It was a gift to the audience that he was there. Rob said as CMOS gets further and further stressed, it's going to be a battle between Mr. M and Mr. D. The photos he had on his slide at that point, he said looked suspiciously like John Cohn. Historical purists would insist they were Marx/Darwin. Revolution/Evolution? Get it? Rob said scaling issues, power issues, manufacturing issues are sending product variations through the roof – both in quantity and quality. He also said that in the future, either we're facing even more relentless uptrending in the Moore's law thing, or we're facing an unavoidable plateau. Leave it to an academic to see both sides of an argument. Rob said we're either facing an uncorrelated smear of probability as a first principle in design, or it's deterministic. Yep, definitely an academic. Then, he put up a slide that was nothing short of a product endorsement. It was the soon to be released Microsoft RTLsignoff Professional. Nobody in the audience had even been to happy hour yet, but they were already rolling on the floor. Then Rob told us about EDA tools moving with marked determination to the bottom of the value chain. He put a page from Amazon.com up on the screen. One of the tabs on the GUI said: "Misc. EDA & Lawn Tools" If you clicked on that link, you could learn that people who bought back-end tools, also bought ... ****************** Then Sani Nassif from IBM had a shot at the crowd. He said he was going to explain why revolution was bad. "It hurts," he said. He could have stopped there, but he didn't. He said revolution isn't guaranteed to produce progress, it's costly in terms of careers and lives, and revolution is blind. It's a process whereby you're hoping that in your groping you'll get to the door. Sani said we have to stop using band-aids. We have to stop thinking of a transistor as a switch, and he had a slide with a pendulum to prove it. There was a switch at one of the pendulum arc, and a transistor at the other. "It's not a switch," he said. Let's get the pendulum to swing back from that extreme thinking. Sani also said a wire's not a capacitor, parasitics matter, power supplies are not ideal, design rules are not step functions, and the world is most distinctly analog. It's full of variability. Leakage and defects happen. We must embrace those concepts back into the EDA tools because designers can't hope to cope without automation. Let's stop using rules and let's move to models. He was pretty adamant. He was also adamant that's there enough life left in silicon to last through to the end of his career. We're over designing and losing out in every generation node, because we're not working to hone our understanding of the physics at that stopping point. We've got to slow down, he insisted. We've got to give EDA time to catch up. I was moved almost to tears. He said, "Let's take what we know and put it into the tools." For a second I thought to raise my hand and ask, "But isn't that what we are doing?" But, Sani sat down, and I forgot my question immediately. ****************** Wayne Wolf from Princeton stood up. He's seems like a pretty serious guy. He said, "Revolution is inevitable." And that's because of SoC integration. We've got to understand these applications; these are complex functions on complex chips. We've got big design challenges. There are large amounts of memory, which are requiring large runs to understand the system balance. SoCs today, Wayne said, are built primarily from IP and we've got to verify it at the interface. Software IP may exercise hardware IP in unexpected ways, so we need application-specific failure info and tests. It's harder to define the hardware that's going to be expected to handle yet more dynamic tasks, the design simply must understand the application. Software must be managed with the hardware. It's all in the methodology. So, be prepared to be performing massing simulations. And get with program - Transaction-level modeling just isn't going to cut it. It's just never going to be sufficient. Then Wayne sat down. ****************** Virage Logic's Yervant Zorian stepped up to the mike. "It's evolution," he said. We're going to see gradual improvement in the tools as needed. We'll address the complexity, the time to market, cost and yield. Today's challenges, Yervant said, look like yesterday's challenges. Then he drilled down. Complexity requires a grasp of hierarchy, especially with 1-10 levels of metal. Proper tools are needed. SIP's require new 3-D packaging technology. We need gradual evolution to ease into testability, particularly when it comes to dealing with IP. Multi-chip modules are up to 3 or 4 layers. IBM predicted that 20 years ago. Hence, it's always been evolution. It will always be evolution. Time to market must be short. Therefore debugging must be quick, efficient and accurate. We need to debug using on-chip structures. Cost considerations are going to require that reconfigurability be the name of the game. (That's good, because that's what I think too.) You're going to need to be able to adjust the chip to the task at hand – either at the moment of manufacturing, at the moment of power-up, or at in the field. You gotta be nimble. Yervant didn't exactly use "nimble," but that's what he meant. Yervant said, evolution requires co-designing with function and manufacturability all blended into one thoroughly streamlined design methodology. So, repeat after me ... we've just got to optimize that front-end to back-end flow. We've just got to get seamless integration between logic-and-memory and test-and-diagnosis. Before Yervant sat back down, he said we need more test engineers. We need more memories that can be repaired. And, we need yield learning. Yervant endorsed evolution, but I agreed with what somebody said later on. The guys pushing evolution sure seemed to be making an articulate argument for revolution. And Yervant is one of them. ****************** IBM's Uwe Fassnacht took the stage. Straddling the fence, he said, "We need broad evolution with occasional revolution." I thought to myself, who can argue with Family Values? Uwe then said we need change to repair the probabilistic design techniques that exist in timing analysis. We've got IP and libraries that are just going to have to be characterized for variability. We need new modeling formats. We need new and/or better standards. And, how amazing to hear a guy from IBM say (and I paraphrase), "IDMs will have the advantage here. IDMs can promote and organize close collaborative work between the EDA tools/vendors, the designers, and the fabs." Before Uwe sat down, he said we need better design services, better IP, better design methodologies, better tools, better characterizations, and better power tools. He's right, you know. Help me understand, though. Was Uwe's plea one for revolution, or one for evolution? ****************** Jamil Kawa from Synopsys actually went last, but I'm in charge here, so I'm making him go second-to-last. Jamil stepped up to the mike and said, look at the span of history. We've gone from vacuum tubes to nano-tubes. Look at the automobile industry. It was fragmented. Then Henry Ford showed up. Now, we're almost to point of automobiles that are assembled on demand. This is evolution. If you're were listening, you might have thought, "Assembled on demand is very cool!" Jamil said we've got process worries, we've got memory worries, we've got analog and RF worries. How can we possibly know what will emerge from the pressure to solve these things? He said that revolutions in design are evolution. At least, that's what I heard him say. Jamil said engineers are innovators, but mostly engineers think as 'evolutionaries,' not as revolutionaries. And besides, even when we think we're revolutionary, we're still always building on the past. CMOS has got issues. There are high-k oxides. There are low-k dielectrics. These are problems. He ended with the well-known sentiment: If people had 5 fingers, 7 would be 12. I would have chosen that as the best single statement of the hour except for what came next – actually it came just before. Because after Jamil sat down – actually before he stood up – Ralph Otten had the floor. ****************** TU Eindhoven's Ralph Otten said, let's observe the chip industry over long periods. If we do that, we can see certain rules of thumb that are process-generation independent. We can have deep insights into the thing. Ralph had some slides with equations that invoked Moore, et al. That was comforting to the crowd. I would put the equations here, but it's too hard to track down the symbols I would need to do that in Word. So Ralph concluded. If the only way to solve the terrible problems we're facing today is to bring the problem solvers and their algorithms in house, into the big IDMs, then so be it. And when that happens, EDA (as an industry) will die. If the only way to solve the terrible problems we're facing today is to out-source the solutions, because it's just too much for a single IDM to solve in-house, then the problems will fall to the third-party tool providers to resolve to completion. But, the problems are so intense, there are no possible solutions. The third-party tool providers will be incapable of solving the problems, and no one will want to buy their tools. And when that happens, EDA (as an industry) will die. To be fair, Ralph's arguments leading to this conclusion were slightly different than the ones I lay out here. I'm hoping he'll see this bit of text and respond with a more accurate recounting of his reasoning. For now, suffice it to say, that's the reasoning I heard. In any case, if death is a step function, and revolution is a step function, Ralph is a revolutionary. If A=B and A=C, then B=C Ralph's last statement, I'm unequivocal about. Ralph said, no matter if it's evolution or revolution, "Either way: EDA dies." **************************** Next week ... Episode V: WID and D2D ****************************
Peggy Aycinena owns and operates EDA Confidential. She can be reached at peggy@aycinena.com
|