Bigger, Faster, More Diverse And Expensive

Aart de Geus, chairman and co-CEO of Synopsys, sat down with Semiconductor Engineering to talk about the race toward AI everywhere, how splintering markets are affecting design, and why software is now such a critical component of hardware development.

SE: We’re seeing big advances in compute performance due to advanced packaging and heterogeneous architectures. Is that sustainable?

de Geus: Yes, and the reason we’re getting all of this massive compute power is that it’s being focused on a narrow set of problems. That’s happening because architecture is the only way to get a substantial increase in throughput or calculation. Another way to look at this is that massive data multiplied by machine learning and communication is an economic driver of such magnitude that you will only be happy with 100X or more compute power. Then you can do real stuff, and the economics support it. So there’s a big push on silicon. For the past 50 years or so, we’ve had a Moore’s Law push, meaning the technology has evolved to make things possible. It’s now grown to the point where we can have smarts in anything, and that’s driving an economic pull. So the Moore’s Law push has turned into an economic pull, where 10X to 100X is just the starting point. That’s why you’re seeing this evolution. Enormous amounts of data have to be coalesced through a network to get it ready for utilization. Then you pull out massive amounts of learning — and send back those learnings, with interpretation — to the edge AI. That go-around, compute-wise, is extremely demanding. And if it could be more demanding, we could get better results. There is no field that is not looking at this today.

SE: So in effect, the splintering of end markets is driving the demand for more compute power, while the computing is driving a further splintering of those markets?

de Geus: Yes, but it’s not clear if the splintering in markets has always existed, and that the only way to get benefits out of this new solution is to splinter the whole electronics paradigm for efficiency. Solving some problem for agriculture is very different from a data standpoint than solving a problem for health care. And while there are similar principles around how you collect the data and the rules around it, these things are all very domain-specific.

SE: Would these markets have benefited from very specific solutions earlier if they were available?

de Geus: Not necessarily, because in the first go-round those markets were enabled by computation and mobility. In the 1950s, everything was done on paper. So computation already changed the world, and massive networking changed it again. Making that networking mobile changed it yet again. I look at those as technology pushes, meaning you get a smaller node, higher density, higher speed, lower power. That has worked well. But now we are in a phase where in order to continue to do that, it takes creativity on the architectural side, and it creates economic push to be able to innovate. Because the economic opportunity is so large, we’re going to see more money going into it and more money going into the architectures.

SE: Given that, and the rollout of the edge as a key piece of the infrastructure, what does the edge look like? Is it a highly customized piece for splintered markets?

de Geus: There is no question that each domain will have very different solutions, because the requirements for these areas are different and the efficiency is different. We all know that in the car, for decades, one of the key things was the legislation of safety and how to verify chips and systems with a variety of self-test capabilities. Now, when you add machine learning to that, all of these things are just as important as before, but the car can be breached, and so can safety. If a car can be stopped by a cell phone, you have some serious problems. That’s different from a game where if a game gets hacked you either lose or win. So each domain will have its own requirements. In health and medical, for example, you have much more concern about privacy. Big learning will be more centralized, and the interpretation will be more about edge devices. If there is some learning that can be done in the edge device to minimize what has to go through the network, that helps. But at the same time, we’re seeing the networks going through a fairly radical shift with a 5G push, and the reliability of that is being hurried for a good reason.

SE: If you look at the cloud today, despite differences in chips and implementations, most of these implementations are still largely doing the same thing. Will the edge be that cohesive?

de Geus: I don’t think so. The edge is going to be completely optimized for the lowest possible cost. If you are collecting temperature variations, that’s a very different data stream than if you’re collecting facial recognition images. So with images the first thing you think about is how to compress that data, while with temperature data there’s so little that you don’t worry about it. Techonomics governs everything, and sometimes it’s the ‘onomics’ part that is the driver and at other times it’s the tech. Historically, tech was the push. Now the ‘onomics’ part is the pull. But they are an inextricable pair.

SE: How do you see this AI revolution playing out? We’re at the phase now where it’s new and cool, but as it moves forward it changes and optimizes — and it trains other systems. So now you have something that’s different from where you started, and you may have bugs you cannot find.

de Geus: We’ve had those problems forever in high tech. There is effort on any of these perspectives to make things work, and work better. And precisely because of the initial, ‘How did you get that result?’ feels a little murky and mysterious, there’s going to be a lot of push in that direction. So if the answer is 42, how did you get there?

SE: Given that, what new markets are on the horizon?

de Geus: It’s hard to predict. For awhile, we were in the phase of, ‘AI solves all problems.’ AI was able to beat the Go champion, which was really good, but that is still far away from any real-world problems. On the other hand, there is unbelievable promise in the Google cars that have been driving around Silicon Valley for the past 10 years or so. We are constantly closer to autonomous driving, except that every so often a car causes injuries. The progress is unbelievable, but I don’t know if there’s a really good assessment of whether a field is at 3% of solving problems or 5%, or whether it’s only 0.3% because all fields are so deep.

SE: They’re also evolving individually. In the past we could tell how much faster an x86 processor was running a particular application. We don’t have those kinds of metrics anymore, right?

de Geus: Metrics work really well if you have a really simple issue. So if you are going to measure all the people that do really well in the 100-meter dash, that’s an extremely clear metric. If you are asking how much AI has helped people be happier, then you look at everything from suicide rates to how long people live. Therein lies an unbelievable opportunity. AI is attempting to replace human intelligence, which is much more developed, but in some places slow. In many ways, the computer did the same thing. The computer was not able to write better letters, but it did make it possible to change letters when you made an error without having to use whiteout. Sharpness of metrics will become the definition of success, and right now most of the AI algorithms do some teraops/flops metric. That is the metric for the hardware, but the software will be domain-specific.

SE: So what is the measure of a successful company in the future?

de Geus: It’s still about revenue. You try things, and you see what works. You learn by doing, and this is the measure of entrepreneurial success. If you have a car that gets significantly better mileage, but it looks like a dog — which is how many of the early high-gas-mileage cars looked — then you don’t sell a lot of those.

SE: Will we see more companies coming and going, and at a faster rate?

de Geus: Absolutely. We are supporting a radically large number of companies right now that all claim to have the best AI chip. We watch all of these, beginning as startups with great ideas. Many of them have great ideas, but do they emphasize the right thing? And even then, the minute it looks like something is promising, is the next generation available from someone who watched this evolution and executed fast? We’re going to see this for quite awhile. We now have proof that ML techniques truly can do better than a human in narrow problems. It’s like putting something in orbit after staring at space for 2,000 years, and now you have set the moon as the next target. We still face an unbelievably difficult journey — actually much more difficult than getting to the moon. But we definitely are seeing a big opportunity space. So now you have some proof of concept, combined with the human brain as the target.

SE: Not all of these startups have been successful in getting chips out the door. What are you seeing as the problem?

de Geus: These chips have two main characteristics. They have to deal with large quantities of data, so the whole pathway for that data is complex, both inside the chip and getting it in and out of the chip. And second, the transactions on the data are in many ways simpler than general computing, but there are many more of them. Therefore, all of these chips have duplication of the same core in astronomical numbers. The entrepreneurial push is in full swing right now. But this is a field that is unbelievably broad, and there are very different types of problems. The race is on.

SE: Because China has enormous amounts of data, the general thinking is that country has a lead in AI. Is that really an advantage?

de Geus: The real challenge is whether you can utilize that data effectively. China has some very proactive strategies for how to use new technology, and that’s not any different for AI or 5G or the notion of how you build a city where you can proactively manage traffic rather than predict where you will get stuck. If you believe technology can make those kinds of decisions, the key is how you authorize technology to do that to a degree that impacts the final result. There is no question that China sees this as a great technology space. There are people trying to do the same in the United States, but this country is less prescriptive about how to apply that technology. If you decide that we need to manage the amount of CO₂ that’s being produced, then in no time you come to the conclusion that we need to make market decisions about carbon pricing. That will manage it very quickly. But then you quickly get to the question about the role of government, and there are positives and negatives for each country.

SE: Given all of these changes, where does Synopsys fit?

de Geus: We’re right in the middle of all of this. Three years ago we changed the tagline under our logo to be ‘Silicon to Software.’ My sense was that silicon had arrived at a point of capabilities where it would completely change what software could do. AI is the most visible manifestation of that. So we have software-based prototyping, FPGA-based emulation machines, and these are being used to simulate software even though you don’t have the hardware yet. And simultaneously, we also put a bet down to extend the company into software quality and security. Last year, that surpassed 10% of the company’s revenue. Five years ago we literally had nothing in this area. That is how we are dealing with the evolution of this picture. We are in the middle of AI, software that doesn’t have systems yet, and software that can run on low power and have higher security.

SE: And now you can co-optimize on multiple fronts?

de Geus: Yes.

SE: What comes next? What pieces do you need that you don’t have?

de Geus: Next is a really long time. The desire for more computation will be unperturbed for multiple decades. We’re so far away from being able to do broad problems effectively that this will have to continue for quite a while. At the same time, it’s clear that silicon technology is hitting some of the economic challenges for developing the next thing. But the next thing is so valuable that I don’t think that’s really going to be a problem.

SE: Do you foresee the kind of tools you’re using being used to help develop algorithms?

de Geus: Absolutely. People like to know how much power algorithms are going to consume. We now have a reasonably good approximation and estimation methods that we can put in the tools for prototyping software where the hardware doesn’t exist yet. It’s essentially predictors for power utilization. That’s not any different than 50 years ago when people were using simulation to determine if something would work. Simulators are more complex today. They have to run blindingly fast, they have to run both the software and the mock up of the hardware, and they need approximations of power to be good enough to trust the results. And all of that is progressing rapidly, and it’s never done because it always needs to be faster.

SE: Scaling allows you to put more on a chip or into a package. Where are you with that?

de Geus: We moved into multiple dimensions a few years ago with timing and power utilization and testability. We always have worked at the very max of whatever abilities we have. The race is still on after 60 years of running a marathon. Now we can split these subsystems into multiple chips in a package. There are some challenges, but that’s what we do for a living. There is excellent progress, but we need to have more. We can truly claim that 80% of our products throughout our entire existence have been state-of-the-art. That’s the DNA of the company. Technology computer-aided design, which was based on Maxwell’s equation, has shifted to Ab Initio, which literally means from the beginning. We simulate at the atomic level. The level of detail has continued to go down into angstroms. If you look at the thickness of advanced gates, that’s 6 or 7 levels of atoms.

SE: Do you see a point where it’s not feasible anymore?

de Geus: The techonomic push has turned into a techonomic pull. So if you believe there is a high value that can be solved with difference intelligence in different fields, you quickly come to the conclusion that it’s worth it. So if semiconductors are $500 billion, and electronics are $2 trillion, the worldwide GDP is $85 trillion. Even a small percentage of change is a huge opportunity. If you just improve crops by 5%, that’s an enormous gain. If you do medical diagnostics radically better, that is very big money. The push on more advanced chips or systems of chips will continue unabated, and regardless of whether a chip costs $1 or $2, if it has impact it doesn’t matter that much.