Thinking Big: From Chips To Systems

Multi-die systems require new tools, technologies, and some very different approaches to design automation.


Semiconductor Engineering sat down with Aart de Geus, executive chair and founder of Synopsys, to talk about the shift from chips to systems, next-generation transistors, and what’s required to build multi-die devices in the context of rapid change and other systems.

SE: What are the biggest changes you’re seeing in the chip industry these days, and why now?

de Geus: It’s not just the size of the transistors anymore. Angstroms are a really cool nomenclatures and they represent a whole new age. But it’s things like backside power delivery that are really interesting. And you might ask, ‘Why didn’t we do this before?’ The answer is that it was really hard and expensive.

SE: We also have glass interposers and substrates, which require new ways of handling different materials. All of this is happening very fast, too.

de Geus: Yes, and once you say that this multi-die thing has a shot, the race is on to produce the first ones. Intel mentioned Ponte Vecchio as the first substantial one, and we were involved in making that happen. Out of that came many lessons, but the number one lesson is that it’s possible. It’s not any different than, “Oh, can you get to the moon?” Once you get there, it’s possible, and now many other things can happen in space. This is true for many human endeavors. What determines the moment it becomes possible is techonomics. It’s this intersection of whether the technology is good enough and possible, and is it affordable. You also can go from, ‘It’s affordable, but we don’t quite know how to do it.’ Once that horizon is visible, the ingenuity and the engineering advances are remarkably fast.

SE: There are a lot of new ideas out there, and a lot of uncertainty about which ones will take off.

de Geus: That’s a normal state when you arrive at the critical-mass zone. People smell, ‘There’s a there there.’ And now the question is which there is actually there or not there, and that race is on.

SE: Do we have to agree to some 80/20 rule or 90/10 for this to work with sufficient economies of scale?

de Geus: I don’t think you have to agree. You have to win. Winning means you have solved the techonomic equation so that it’s great and affordable. Synopsys was a pioneer four or five years ago with AI, because we were able to optimize substantial pieces of chips automatically through a whole set of steps. The design flow itself got optimized, not just an individual tool. It’s not just that the layout has gotten better. We’ve worked on that for the past 40 years, and for the last 15 years with machine learning. But suddenly it’s now multiple steps, including the timing and the power and the form factors. I remember the first design where I saw it work, because we started, and at some point it got more and more versions. And the versions got better and better and better and better until it petered out, which was the best we could do. That was the same moment and feeling and graphic representation as 1984, which was our first synthesis, where we suddenly could put on the map a solution that was better than what a human could do in a fraction of the time.

SE: To some extent EDA has been working with some form of AI for years, right?

de Geus: Yes, and the product code name at the time was Socrates, and the ‘es’ was expert systems. And the reality was the expert system was a little bit of an expert. It was really a good amateur system. But it was this notion that certain human steps can be replaced with something better. Then automate it, and self-automate it going forward. That’s was when the papers were published on it, which is 40 years ago. Synopsys is 37 years old, but the technology we started with is Socrates.

SE: One of the areas that has resisted this kind of automation is analog. What’s happening there? Is it mostly digital. Is it still mostly analog?

de Geus: Analog chips are always a bit of both. They get segregated very well, and they are optimized in different fashions. What we’re talking about now is finally doing analog better and better and better. And the challenge with analog is that there is a configuration that’s a schematic of which transistor talks to which transistor and how big are they. And then there is, of course, the layout, which is not as automatic as in digital. You have to optimize both of those things in parallel. In digital you have a couple of questions. Is the function correct? How fast is it? And how much power does it use? End of story. In analog, every signal has to meet all kinds of other constraints in terms of the amount of fiddling — up and down voltage, sensitivity, temperature, and so on — and that’s a complex set of physical equations. Optimizing that initially was much harder, but now we can do that amazingly well — even better than a human.

SE: If we can accelerate analog design, we save an entire step, right?

de Geus: Yes, and we’ve been able to accelerate analog design by well over 2X. When we develop a new library of all the interface blocks, those are mostly digital. But they all have a PHY, which is the physical part, and that’s analog. Those tend to be the bottleneck. We have the digital part nailed down and we know what it’s supposed to do. We can synthesize and automate place-and-route for the file. The challenge is tuning it, and that’s why these tools are so powerful. If you can accelerate that, you can move much faster to the next node.

SE: So have you convinced the analog designers to use these tools?

de Geus: The key is to get people to recognize, ‘If I use this I can do something better.’ It was the same story in 1987, when we rolled out synthesis. Only about 3% of the customers complained we were taking their job away, and when you think like that your job is going away. But the majority thought, ‘Oh, this is fantastic, because now I can do something bigger and faster.’

SE: Who’s going to push these capabilities. Will it be the chipmakers, or systems companies, or some new type of company.

de Geus: It will be the chip companies that are here today, because they know how difficult it is. A brand new company doesn’t know exactly how they’re going to be differentiated with that. The adoption comes from the existing users. It’s not that you have to suddenly push a whole new methodology. It’s incremental. It extends what you can do yourself.

SE: Are we getting to the stage where computing is pervasive, making the box almost irrelevant?

de Geus: There’s always the question is whether what is possible is desirable, and is what is desirable possible? The second category is actually the harder one, because that’s the one that moves mankind forward. It sounds a little grandiose, but that can literally move humanity forward. There are so many areas that are opening up because there are technical discontinuities that happened with a 10X improvement in performance that you couldn’t get with 2X.

SE: But designs also are becoming more complicated and expensive, so people are expecting them to last longer. At the same time, we have a much shorter window on how fast the innovation is happening.

de Geus: That was gold for the mobile industry all through the 2000s. You barely had figured out the five features on your phone that you needed before you really needed the next one, because that was ‘so yesterday.’ That slowed down a little bit, but there are still more features. I just saw a demo of a phone that can do live translation of language, so I can say something in French and hear it in English. That’s pretty amazing.

SE: But with the rising cost, you still want to keep that device longer, right? And you don’t want to trade in your car after three years.

de Geus: That’s right, and this is a very big challenge. It begs the question, ‘To what extent can you create a car that is modular in its electronics?’ The car manufacturer is very much struggling with that, because they need to decide how big a jump do they take to the next node. If they make a big jump, they have a longer time, but it’s also much harder. And cars need to be really safe and reliable. Mostly, they’re counting on upgrading the software. And we know what that feels like on our phone, because every time you get an upgrade it’s doesn’t quite work.

SE: How about a model where you take your car in for service and they replace the chiplets?

de Geus: With that comes oodles of software to write. Verification of all of that for safety reasons is going to be a big challenge. That’s been our job as a high-tech industry since the very beginning. People want to continue believing that chips will always be faultless, and the reality is that it’s not like that.

SE: So does the chip industry need to shift to resilience rather than bug-free?

de Geus: Resilience is good, but redundancy is expensive.

SE: In the past, you’ve pointed to adjacencies as the logical expansion point for EDA vendors. Is that still the case?

de Geus: The challenge is to find commonalities. Our job is to drive technology forward, but at the same time drive it for partners that are a little different. For well over a decade we have developed an IP collection first and foremost for TSMC, because they were by far the largest one. But it’s the same collection now used for Intel and Samsung. This is very important for the end customers, because they can design with the same cores and decide where they’re going to go in silicon, and this is risk-reducing. So quietly we’re both a point of commonality and also an instrument for their differentiation. Take backside power delivery, for example. Intel led on that, and we did some optimizations about 24 months ago. I saw 24 months ago, ‘Oh, this is going to be cool.’

SE: You bypass a whole bunch of metal layers with that approach, not to mention reduce the noise, right?

de Geus: Exactly. It’s like instead of routing a water line through your living room, you can put it in the ground. It’s all techonomics.

SE: It’s also an evolutionary industry, right? Each new step is built on something else.

de Geus: We forget the single lesson learned in 1997, which is that nothing works across capacitance. And so all our customers want the most advanced thing — but just don’t change anything.

SE: Given all of this, how does Synopsys see itself these days? Is it an EDA company, or something else?

de Geus: We’re still at the heart of high tech, because we touch the parts of the exponential that are changing the most rapidly, while being both a player and a customer of the equivalent on the software side, which is all the super AI development. We’ve moved from scale complexity to systemic complexity, because systemic complexity is more multi-dimensional and involves more parties. This is the difference between additive complexity versus multiplicative complexity. With multiplicative, if there is a single zero we all get zero. It doesn’t matter if it’s the machine that puts the chip down on the board and bends a pin. No matter how many angstroms you’re working at, the bent pin is not connecting. You have to become better at team play. We’re moving from a small band to a bigger orchestra, but the trumpets still need to come in when the drums come in.

Leave a Reply

(Note: This name will be displayed publicly)