The Software Behind Silicon (with Synopsys Founder Aart de Geus and CEO Sassine Ghazi)

Transcript: (disclaimer: may contain unintentionally confusing, inaccurate and/or amusing transcription errors)

Ben: Hello, Acquired listeners. Today is a very special treat. Our guests are the founding CEO of Synopsis, Aart de Geus, and the CEO today, Sassine Ghazi.

Synopsis is the $80-billion company that makes the software that chip designers rely on to do their jobs. It is one of the two big players. Along with Cadence Design Systems, the field is called Electronic Design Automation or EDA.

It’s a crude analogy, but you can think about it as the productivity software for chip designers, like the Microsoft Excel or Figma for that profession. So much of the complexity of chip design these days has been baked into the EDA software, that it makes entirely new types of chips possible that you couldn’t do without them.

They’re the essential infrastructure behind the AI era and all the semiconductor innovation that we are experiencing today. No AI applications would be possible without EDA and the incredible optimizations that the software does for chip designers. In fact, in a full circle moment, Synopsis even uses AI now to design the software to design chips. With that, on to the interview with Aart and Sassine.

Aart and Sassine, welcome to ACQ2.

Sassine: Thank you for having us.

Ben: We wanted to do a deep dive for listeners. It's been a while since we were in the land of semiconductors. We've covered NVIDIA, TSMC, Apple, ARM, so many of your customers and companies that you work with, but we have never hit the world of EDA directly.

Listeners, Aart de Geus is one of the storied pioneers of the semiconductor industry, 37 years ago founded the company, and really evolved to become an essential part with Synopsys of the semiconductor value chain and whole ecosystem today. Recently, Sassine, you transitioned and took the helm going from COO to CEO. We have the unbelievable privilege of having both of you with us here today.

Aart: Pleasure to be here.

Sassine: Yes, thank you.

Ben: All right. This is Acquired and we love history. I think to ground the current state of the semiconductor ecosystem, why don't we wind it back to the beginning of Synopsis? What did the lay of the land look like then? And how crazy was the idea for what would become EDA when you were first getting started?

Aart: What we're talking about here is the mid-80s. Just to put a little stake in the ground, I was at General Electric designing at about four-micron. I know you don't remember that that existed, but they were this big. That also says that General Electric was actually in semiconductors. At some point in time, they had invested in the factory of the future. This was the future. The same as AI is now everybody needs to have it, well, then it was semiconductors.

Things went pretty well until they didn't go so well. Didn't go so well meant, in hindsight, that in 1985 was the worst downturn in the history of semiconductors in the 80s and 90s. I think it hit General Electric hard because that was a very large company. They had a very steady state, dividend-driven investor group. These ups and downs in the semiconductor industry turned out to not be really their thing.

Long story short, we're going to be laid off. It was completely accidental, but it was also accidental that in the five years roughly that I've worked there, and especially the last three, we had developed a number of design tools that actually were very innovative. One of those was synthesis. We were somewhat known because of that.

While on one hand, I literally actually interviewed for a job—we're going to be laid off after all—simultaneously, we had this rebellious idea of, what if we took the technology and looked if we could do a startup? We did it with (I think) great care of thinking because we decided very quickly that we're going to do that in full light of General Electric, meaning tell them about it and not take anything.

It was a great company. They had taken very good care of us. Actually, it was just the right thing to do. There was an opportunity to advocate this spin out with the technology, which all was going to be lost. They were going to get out of this field. With a small group of a total of seven people, we essentially got out of GE with their support, both some financial support and the transaction on the technology for what's the equivalent of a million dollars of value.

Fast forwarding by the time we went public, GE pocketed $23 million. I'm still proud of that because they really deserved it. It was the right outcome, but it was all pretty accidental how that came about.

David: Wow. It's so rare that a corporate spin out into a venture-style venture goes well. That's amazing. Actually, what were you all designing microprocessors? Was this more specialized chips?

Aart: No, this was so-called gate arrays. These were essentially chips for other customers. It's hard to remember a gate array, but a gate array was essentially a long series of transistors that have been prefabricated. They are sitting in rows, and you use the first layer of metal to essentially take those transistors and make certain gates out of it—the NAND gate, NOR gate, an inverter. That was pretty much the choice set. You use the rest of the layers to connect those to actually create the actual circuit. GE would do that for their customers and then manufacture these chips.

Ben: Gate array, that's the GA in FPGA, right?

Sassine: Yes.

Aart: Yeah, I think so, actually, because it's the same concept.

Sassine: Yes, it is exactly the same concept, yes.

Aart: I've never heard that question. It's pretty good.

Ben: It's funny you said gate array. I haven't thought about that in a long time, but FPGA is the current hotness.

Aart: You're absolutely right. We're hot already then, I guess.

Ben: This idea that you had that you could turn it into its own company, was there a blueprint that already existed for chip designers needing great software to do their jobs well, or was that a novel idea that could be an independent company?

Aart: You really have two questions at the same time here. Why did we get into synthesis in the first place? It's a technical question, and then how to get to a company. It's funky how we got there because while it was there, there was a guy at GE who had explained in some seminar that if you used multiplexers, you could actually create circuits that would be denser than just NAND/NOR inverts.

I talked to one of my designer friends and said, can you put together the footprints that you need to make a multiplexer? And he did and put that in the library. We'll design with those, and it gets smaller circuits. The problem is none of the designers knew how to use them.

After some reflections, we thought, why don't we just automatically design that? Somehow we managed to write a program called Socrates that actually did that, and got quite good results, although it turned out in the long-term that multiplexers were not a good idea because multiplexers are not restoring logic, meaning if you put three in a row, your signal degrades. Whereas with all the others, the signal stays a very square wave, which is what you wanted.

In the process, we became quickly known through some papers published as being on the frontier of this thing called synthesis. Of course, by that time, we discovered that IBM had worked on it for a long time, so did Fujitsu and Toshiba, and a whole bunch of large companies. But at the same time, we knew we had something because the results were astoundingly good compared to what the manual design did before. And within GE, they used it on the gate arrays with great results.

The whole notion of, well, now suddenly we're going to be laid off and all of that is gone, it gradually morphed into talking to some VCs, talking to some designers and saying, well, what about creating a company?

You have to understand that at that point in time, I was a very young person. I had a bunch of way younger people. Because six out of the seven had all been summer students, that's the cast of characters. The notion of writing a business plan was an interesting concept. I still have a couple of the books that I bought in the local Barnes & Noble of how to write a business plan.

David: Jensen did the same thing when he had to write a business plan for NVIDIA.

Aart: Although, he was already closer to the business side when he worked at LSI Logic. It's the same concept fundamentally. The one thing I just couldn't figure out was, what is the difference between orders, revenues, and sales? To this day, I don't quite understand the difference between sales, revenue, and orders. For that, we have people now, as they say.

Ben: You invent a great product, and those things will figure themselves out.

David: Different revenue, bookings, billings. One thing I want to understand quickly before we get to the company, before synthesis and software, how was chip design done? Was floor planning done with drafting boards? Was it pen and paper?

Aart: You're right with so many of those pieces. The first thing to understand is there are fundamentally two layers. There's the functional layer and then there's the physical layer. When you talk about the layout, you already have an understanding of what the function is and what the building blocks are. Now you actually have to physically design them and physically connect them.

We were working at the functional level. The notion is you have a complicated digital math function that you want to implement. You need to choose the right gates. There are a number of methods to simplify that, but ultimately you build a set of building blocks that you then connect. You typically did it on paper or gradually on a schematics entry type thing.

Then comes the question, how good is it? Fewer gates is better. Area was not really used. The substitute at that point in time was just the number of gates. If you knew that, the rest was determined.

The other thing that was important and will turn out to be absolutely crucial in how we differentiate it is we understood that speed was key. Speed is determined by whatever is the longest path through your design. We could judge if the circuit was getting better. Not only was it getting smaller, but also was it getting faster? That combination turned out to be the key differentiator.

Ben: Fascinating. Sassine, we have not yet gotten to your role in the story, so I want to start from your beginning with Synopsys. You joined the company in 1998, but I'm sure that in your jobs at Intel and elsewhere, you came across synopsis before. Do you remember your first experience?

Sassine: Yeah. As Aart is describing to you the synthesis, the gates, the function, then the place and route, my first experience with Synopsis, I was doing my master's in electrical engineering. Actually, I was more on the control system side, so I did not touch Synopsis at all.

After I finished my master's degree, I realized that's not the field I want to be in, because most of the job opportunities at the time were controlling massive mechanical stuff, be it oil diggers, a giant satellite, or what have you. Then I started my PhD in VLSI design. This is my first introduction to synthesis. As you're describing it, Aart, how do you build the library, the building block you synthesize?

Back to your question, David, the largest design at the time that a single engineer could do was very limited by the number of gates, because the actual software from a capacity point of view could not manage. The clock time to run and synthesize will limit how much you can design in terms of size of design.

Ben: Wait, so the physical on-chip limitations were actually not the bottleneck? It was the ability for the design software to handle the complexity?

Sassine: It's both. First, you need how much the software can handle the complexity and still meet your performance target. Aart is right. At that time, performance target was the key. Power area was so secondary, which 10 years later, it became performance power, then area. Now you optimize to the end all at the same time in order to make your requirement.

When I started my career at Intel, believe it or not, a lot of the stuff that the synthesis creates were manually verified. You lay out the transistors, you make sure you have the right width, right length and how you connect them together to create the actual cells.

My experience with synopsis was grad school. Of course, at Intel, I used many of the Synopsis products. That's when the opportunity to join Synopsys came along. I was super familiar with the company and the support, the product, R&D, and the rest are history.

Aart: You joined because we were the only company that had no bugs, is that right? Can you confirm that?

Sassine: Exactly.

David: Doing this synthesis class of problem is really, really hard to do, right?

Aart: What made it particularly hard was there were, of course, techniques to optimize just the functionality. Many of those were algorithmic. We added to what we then called an expert system, which was to look at certain situations in the circuit and say, this doesn't look good, but I know here's a better version. You would add so-called rules to make it better. You add a rule, it gets better. You add five rules, it gets better. You add one more rule, it gets worse because now you need a rule to manage the rules.

I always like to highlight that it was an expert system because that makes us kings of AI 30 years later, but the fact is it was limited in its capability. But it was dramatically better than humans.

By the time, and this was not even the first product, it was a prototype of the first product that we had. The minute we became a company, we talked to customers. They would give us one of their circuits that had maybe max, a couple of hundred gates. They had worked on it for many, many, many weeks. Then in a matter of a few hours, we could literally give it back to them 30% smaller, and smaller meant 30% fewer gates, and 30% faster meaning shorter critical paths.

They would look at it, and then it's impossible. There's no way you did that, and then they would go away literally for two weeks. Then they would come back and say, well, I've checked and I checked. It's actually doing it. Their expectations, of course, were immediately way higher than what we could do because they had just encountered magic. Out of that interaction, something very profound happened. The majority became our friends' customers because they could say, yeah, but you know what you did here, that's not that great.

By being able to look at our circuit and say, that's not that great, it makes them great. But they gave us a gift of feedback that two weeks later, we had fixed based on their input. Therefore, they become parents of the tool, too. Everybody had added something. That whole first generation of two dozen, three dozen companies over time, all had the same behavior, which is they rooted for us because they could see it happening on their own circuits.

David: It doesn't matter what kind of circuit you're making. Whether it's a microprocessor, an analog system, or a gate array, you need this technology, you need this optimization. Intel's happy that you are getting better, even though that's also serving TI or whoever.

Aart: You open multiple boxes here. For starters, we were strictly a digital company. Today, we do a variety of things on analog circuits automatically, but that was far away. Plus, this was a cornerstone to really the digital age.

Before Synopsys, and in all fairness, I should say before synthesis and before place and routes—the two go hand in hand—the field was called computer aided design. You did stuff on the computer, but the computer essentially helped you do stuff that you did yourself.

What was so great about synthesis, it actually created something. We were part of the transformation of computer aided design to electronic design automation. I remember we had a moniker for that. We're the only ones that have license to kill or one of 007, because license to kill means we can actually change a circuit. That was completely taboo before. If a tool did that, it means they had put some bugs in it.

Ben: You wouldn't want software intervening in your own workflow. The creativity was reserved for the human, for the designer. Please only aid me, do not automate for me.

Aart: It was more than creativity, it was the trust that it actually would work. People really could not trust that tools would change it for the better.

Sassine: And it's amazing. Even in 2018, when we introduced AI for the synthesis and place and route, believe it or not, the resistance from our users was I want to know what the AI changed. But that's the idea, you cannot. How many parameters do you want to understand?

There was a lack of trust for about the first two years, even though the results were always better using the AI system, meaning the users could not trust it or use it because they wanted to answer the question, engineers. But I need to understand, what did it do? Of course, right now is a different story. AI is so well accepted that that question is gone.

Aart: What I love about what Sassine just said is, we are essentially a company that has repeated its own history over and over again. I was almost tempted to say that we learned it all from Ronald Reagan. Trust, but verify. Here's this AI stuff, and then you still need to simulate a lot to make damn sure that you didn't have an error in it.

The value of trust is extremely high, but the necessity for verification is also, because the cost of going to manufacturing something that has a bug is, whoa, you made a big decision there.

Ben: Often, I don't trust AI tools because when I do look at the output, I'm like, it's not clear to me that this is better than me doing it. There are many situations where it is and many where it isn't. Do you feel like EDA is uniquely well-suited toward a designer just letting go and saying, I trust the machine, and I don't have to understand every little input?

Sassine: Maybe. What you're referring to, Ben, is generative AI when it's generating something through a natural language and you say, this is 90% accurate, not 100% accurate. I'm assuming that's what you're referring to.

In the EDA world, we're all about massive optimization problems. We're talking about many billions of transistors that you're jamming in a small silicon area. You're trying to optimize, where do you place it? How do you route it? How do you get to the performance, the power? You know this is no way for a human to do it, so our industry has been very much in the space of using technology to optimize.

What we do for AI—we're doing generative AI but put that aside for a moment—is using machine learning and AI algorithms to optimize for an outcome. You always have, as Aart mentioned, many steps before you commit to manufacturing and say it's going to work. You don't just say, oh, that's what synthesis or AI-generated is going to work, I'll go to whomever your foundry. You spend many, many millions and the chip does not work. You have many checks of verification.

What we have pioneered with AI for EDA starting in 2017 right now is used by dozens of customers in production, meaning they're trusting it. They're trusting the outcome, but there are all the other checks you have to go through to verify that it's going to work once you manufacture it.

Ben: To be clear, there are basically specific goals around size, around power efficiency, and around overall performance. It's basically optimizing within that set of constraints. You can prove when it's done that it's better than what happened before the AI came in.

Aart: Those are the outcome metrics. You just mentioned three or four, and that's what we optimize for. You know what that forgets? It's not only that. There's the other 10 trillion constraints that you have to meet that tolerate zero error.

The very big difference between many of the AI optimization things that we see in the world, and some hallucinate more than others, there are many very good ones, is that we have a constraint that is much, much harsher, which is absolute correctness in functionality.

By the way, Sassine jumped quickly 25 years. In those 25 years, there've been 25 years of revolutionary techniques and enhancements, not only to what we do, but to the circuits that we do with our customers.

Sitting on an exponential rate of change, that is a large number of revolutionary changes. Every single one has been delivered in an evolutionary way. In other words, you forget one lesson learned in 1997, crosswall capacitance, you forget that, nothing works today. While he made it sound so simple, we're now AI, the heck out of it, we do. That AI itself works on an unbelievable number of parameters, and the rules specifically for the layout have no tolerance for error. You violate one of those, the yield will go zip down the drain.

Sassine: That's why when the question often comes up, why aren't there EDA startups, why is the market consolidated to just two, it's that exact point that Aart just made. The learning is not just, hey, can I train a model, then create an output, and I'm there? The cumulative knowledge to get to the current state before you look at the future state is massive.

David: It strikes me that there are actually a lot of parallels to the foundry business. Just on the software side, you can't just go recreate TSMC, obviously, as we are seeing. It's all those cumulative years of learning about how to do this. This is the same thing in the software.

Aart: It's interesting because TSMC was founded three months after Synopsys. That's, in hindsight, super interesting because that was simultaneously a change in the industry where the focus was going to go towards fabless design and then foundries that would do the manufacturing. Remember, before that time many companies were IDMs. They had their own foundry.

David: Real men have fabs, I think was the quote?

Aart: Today, we wouldn't dare say it like that rightfully so, but it was a very macho attitude. Still to this day, people that spend a lot of capital are really proud of spending the capital, but they also have no choice.

There's a slight little fun anecdote, which is Morris Chang founded the company, but the first CEO was a guy by name, Jim Dykes, who happens to be the general manager I worked under at GE. When they closed, he went there. It's one big family enterprise here.

The other anecdote I'd like to bring up where we talked about those customers that use our stuff and we learned from, I don't know if you recognize the names, Chris Malachowsky and Curtis Priem.

David: Of course.

Aart: They were at Sun Microsystems. They were among our first customers and actually very good guys to work with. We know them extremely well. A few years later, Jensen showed up because he was the caretaker from LSA Logic, because that is where Sun manufactured its chips. Then they teamed up, I forget which year, but 1993 or something like that. Of those three companies, we are proud to say that we are the ones that have survived the longest.

David: So good.

Ben: Sassine, you got to tell us the story, and then we'll come back to everything else that we're talking about here. I was watching the keynote, and it appeared that maybe Jensen arrived literally seconds before he was about to run on stage, because it was maybe concurrent with the week of GTC.

Sassine: Yes. It's funny. My team thought I was joking like I was setting it up this way. I'm like, no, I mean it. When I went on stage, he was not there yet. He was not in the building, but he was texting.

He's like, I'm in the parking lot. I'm like, okay, great, I'm about to hop on stage. Initially, the idea was about 4–5 minutes, and then I bring him on stage. I'm looking at the clock ticking, I'm like, let me burn more time. Then I'm like, all right, should I continue with my presentation or wait? Yeah, it was live.

Ben: It's wild. I'm sure this exists in other industries, but it has to be a very special thing in your industry that sometimes there's a company that becomes an unbelievably important company in the world, and you get to be a huge partner to them in that success. It's absolutely fair to say that there's no chance NVIDIA could do what they do without Synopsys software. How do you think about the role and the importance in the world that you've really become?

Sassine: As I mentioned in the keynote and I want to say even in Jensen's words, because I was not putting words in his mouth, the Synopsis is mission critical to NVIDIA's success. We don't take that lightly. When we know we are mission-critical to many billions of dollars of our customers' revenue, it's a huge responsibility.

It's a responsibility to continue on innovating because they're aspiring to build the bigger product, the next big thing. If our software and our support, our ecosystem engagement is not able to stay ahead, I don't want to say with them, ahead so when they're ready, we have it, it won't happen.

Right now, where it's happening is with our chip customers as they're architecting the future product and with foundry. Those two are becoming so important. There's a triangle always, us, customer, foundry, that we are working on architecture, on physics, manufacturing, and software to bring it all together.

I want to say, that has accelerated in terms of being interconnected in the FinFET world when the transistors start moving to more complicated manufacturing. Of course, the last five, six, seven years, when we say it's impossible to design and manufacture those chips without our contribution as an industry, it's not an overstatement.

Aart: There's also a historical perspective that, yeah, we can only be thankful for that we were part of this. While now it feels old, Moore's Law essentially was the exhibit of what an exponential is. Exponential is easily the toughest mother of mathematical function because staying on that sucker, damn, it's going fast.

We were lucky that we had seminal technology at a moment, where seminal technology was needed again to move forward and continue to move at an unbelievable speed, not necessarily exactly the exponential that Gordon had predicted, but still the exponential that changed mankind.

Ben: I think this is an important point that I don't want to just gloss over here. People take Moore's law as if it's derived from the natural universe property. It's the same way of F=ma or something. It's not. It literally relies on companies like Synopsis getting clever again. Every time Moore's law happens, it's because somebody got clever again and, oh, my God, we just barely made it. You're the cause of it, not the result of it.

Aart: It's interesting because Moore's law, of course, started just as an observation. He had seen this curve as moving up rapidly, and then he made some prediction that it probably would continue for a while. That prediction then became, well, you have to do it because otherwise you're not not with the team here. The race is on, and then that race itself started to self-time itself against that. By the way, that includes also the different switches to the different sizes of wafers, the ability to manufacture.

It was not always the perfect exponential, but the gestalt of it was absolute. This is not something new in humankind's history. The printing press had exactly the same characteristic, how in essentially 50 years from virtually zero books, it went to 20 million and changed the world. Of course, in many ways, the industrial aides have the same characteristic again.

What is so exceptional about this one is if we look at what we have done, Synopsis, we've contributed about 10 million X in productivity. You say, well, are you going to do another 0.5 now? Hell no, we need to do another 10 to 100 to 1000X. Of course, that's not going to be possible with just doing that on one chip. I'm sure we'll get to the whole notion of SysMoore and how that is changing things.

What is important are two things. One is that in order to stay on that exponential, you need to race like crazy. The way you do that is you race with people that are crazier than you. In other words, you go to those customers that are even more paranoid of not being successful and that are thankful but never happy. That's a polite version. They drive you like crazy, and we've had the good fortune that 75% of our products always be the state of the art for all this time. We've been there.

I like to compare it sometimes to the Tour de France. You see all these guys biking like crazy, and then suddenly the three guys that move away from the peloton. By the time those three guys are a couple of hundred yards away, the others will never catch up. There's a reason for that. The others cannot team up well enough, whereas those three guys, by necessity and by scale, every 15 seconds or whatever it is minutes, they change the guys up front. They chase each other until the last hundred yards, and then everybody's on their own.

In our case, the race never finishes. Our Tour de France is now 37 years. You need to keep going at it. That combination has been unique in this industry.

Ben: Obviously, for these advancements, for Moore's Law to stay true, it requires incredible collaboration between (say) TSMC to figure out how to reduce the number of atoms in between the two transistors or whatever it is that is measured as three nanometers. It requires ASML to make an even better laser, and it requires the EDA software to become even better, it requires the cleverness of the chip designer, and everything has to work together in the same generation at the same time betting on each other's dependencies. Does it feel to you that eking out that next generation performance gain is harder than ever, or has it always felt this hard?

Aart: I would say it always felt this hard. I think it's different than it was. In the description that you gave, you have a relationship between foundry and the equipment vendors. You mentioned ASML, applied materials, those would be the ones that essentially focus on how many atoms, specifically. ASML would be focusing on how many photons you need to get at which frequency in order to get really small lines.

By the way, we are in that domain, too, because Synopsys is the leader in TCAD, Technology Computer Aided Design, which is yet another simulation or modeling, if you like, of the truly miniscule.

At the same time, we alluded to the fact that you have to align how you design with the building blocks that you have available. You can say, well, let's optimize the building blocks for what you're designing, or let's optimize what you're designing for the building blocks that you have. This is often called DTC or design technology co-optimization, where Synopsis is a leader in.

What you notice in this story is they do the designs, they do the manufacturing, but we make it all happen. Somebody builds the Lego block, somebody does the castle, but we make sure that all these things hang together.

Sassine: I agree Aart. It's always been difficult, but I'll say the last six plus years, it's been much more difficult. The reason I'm saying that, if you look at Synopsis' relationship with foundry six or seven years ago, we used to get input from foundry called enablement. You enable whatever they create in our product, and you provide the product to the customer. You become the bridge between foundry customer semiconductor company designing on that foundry through enablement. You take whatever they created, then you put it in your product, and you give it to the customer.

The last five or six years, it's impossible any longer to just do enablement. We sit with hundreds of engineers at TSMC, at Samsung, at Intel, at GF, sitting during the process development. The technology development is no longer enablement because enablement is impossible. You have to invent stuff with them to see, will your physics, the way you're pushing it, handle the design you're aspiring to design on it? That's a big change that happened that is different from before.

Aart: Actually, I buy what you're saying. I was coming from the, it's the same because we have always worked as hard as we could. That does not change the pressure for driving. But I think what you're introducing is actually the notion that beyond scale complexity, we've now entered systemic complexity for every level.

The systemic complexity at the manufacturing side is that it's not sufficient to just understand sub pieces. You need to understand how the pieces work together. I think that we benefit a lot from the understanding from the foundries, but they benefit a lot from the optimizations we do that now help them get faster and better transistors.

Sassine: It's even to continue scaling. If you look at it now, it's really unbelievable to think that the industry is talking about 18 angstrom, 14 angstrom, no longer nanometer. Those are not only from a physics limitation you're hitting the limit. Once you want to put them into a production on a chip, the latest announcement from NVIDIA was Blackwell, 208 billion transistors, imagine the heat those transistors are generating, so just from a thermal.

In physics, when you're designing that transistor, you say, oh, it will work. But once you jam them together and you run the software at the full workload, the heat that is generating…. Even though the physics from manufacturing, it works, there are many aspects that you're hitting the wall that you need to plan, design, architect for.

Ben: Is it reasonable to say that before, you just had manufacturing limitations working against the edge of what's possible and now, we're actually bumping up against the edges of physics governing what's possible?

Sassine: Completely. When you think of the new wave of designing chips—this is what Aart was talking about—you start designing through advanced packaging, multiple dies sitting in a package. Electronically, you can design it to function correctly, and then can manufacture it and package it to function correctly.

Once you start running it in the field with the software workload, then you run into all kinds of physics issues. Thermal is the biggest one. But thermal, when it creates heat, it may create warpage. It may create cracking, so things will start cracking mechanically. You need to take into account all these physics effects during the design stage and during the process technology development. It's absolutely a multidimensional design factor you have to take into account.

Ben: If I could just entertain a thought exercise, what do we need to do to get 4X, 8X, 16X more performance from here? What are the innovations that need to happen for that to be possible?

Aart: There are two things. The first one is well-understood from many years ago, which is to develop the hardware for the specific workload. Somewhat overly simplified, the first years of Moore's Law is, here are more transistors, better circuitry, write your software, and people say, wow, I can now do so many more. And then it's, oh, you need more memory, here's more memory. Write your software and make the world happen.

Then came gradually this conundrum of, well, yeah, but can you not make it a lot faster? Especially all that visual stuff on the screen, it's so slow. Out of nowhere, somebody says, well, why don't we not use a general-purpose processor to do pixels? And then the thing becomes called the GPU. What does the GPU do? It loves pixels. It does only pixels. It can do them forward, backwards, sideways, and so on. Out of that is essentially a specialized accelerator.

Of course, they discover that, well, it would be better to have two, four, or sixteen, actually multi core, even smaller processors. Essentially, what you have is now a workload that has determined the hardware that you need. Advance that to 15–20 years later, and say the workload is driving a car without accidents. You can imagine that by saying, well, you have to take your old 386 and see what you can do with that, you're going to go nowhere. You need actually a whole bunch of specialized machines from anything that takes the many sensors data and compresses it or transports it and so on to ultimately the AI algorithms that can run preferably real time to drive the car.

One of the statements on that is called software-defined architectures. I show it as this V from top down because you're starting with high level functionality, drive it correctly, and get there. At the same time, and it's literally at the same time, you come to the conclusion that chips that are more than 1½ inch square, and I know there are some people who do whole wafers, but it quickly gets to an end. Adding another zero in the number of transistors is going to be really long, long haul.

You say, well, what if we split functionality into multiple chips? What if we brought them really close together? And therein lies the essence, the words close together. Because the notion of having multiple die maybe on an interposer, which is itself a chip, is not new but (a) it was difficult, it was expensive, and (b) it was slow.

If you look at the evolution of the last 20 years, the single thing in my opinion that is empowering multi die is connectivity. Meaning we have improved dramatically not only reducing the distance, but the bandwidth, meaning how many pins you can do, how small these pins are, and how little energy they need to flip a bit or to pass a bit from one chip to another. Still way more than keeping the same chip. If we could just keep it on the same chip, that would be cool, but that's not going to be possible.

Ben: This is, of course, the Blackwell. An example is the new NVIDIA Blackwell chip, a silicon interposer between two dies that enables super fast information to flow between the two dies.

Aart: Intel and AMD have very similar constructs, and they all increasingly now look like they're 12-=20 or so chips. By the way, these chips don't have all B processors, actually need memories. The cool thing with memories, you gradually can stack them, and they stack potentially better because they don't create the heat that Sassine was creating in his processors.

Thermal is absolutely one of the big killers in all of this and a few others, but the enabler is connectivity. If you now look at a picture of bottom-up from physics, you come to this whole new architecture that's really connectivity-driven. You come down from software, as in software-driven, the word architecture has a functional perspective and has a physical perspective.

That opens an entire new age. We call it SysMoore, so systemic complexity with a Moore's law exponential ambition. I like to use the word exponential because I'm a strong believer that what we see happening is another 20 years of additional complexity. Speed may have to be redefined as well, we do a whole bunch of things in parallel. That's a different form of speed, but any speed you can improve is still valuable.

Ben: With SysMoore, what you're basically saying is we're going to abstract up one level what the notion of the system is. We're not measuring Moore's Law specifically on this one chip anymore, we're measuring it for your whole system where the goal might be to drive this car safely. Are we able to optimize  more components of it to work together harmoniously to continue to achieve Moore's Law-like outcomes?

Aart: What I would add is it's not abstracting one level. We've been abstracting more levels already for many years, and I think that includes the embedded software, the software that connects to other pieces, then ultimately the various forms of AI optimizations, and then still the domain-specific knowledge of that.

A great example of this is if you were to ask us, hey, if you really wanted to cut another 20% of the power, which layer would you start with? I can tell you it would not be the transistor, it would be the software somewhere.

Ben: It's like whenever I'm tempted to buy a lighter carbon bicycle, I realize that instead of spending $3000 to shave an ounce, I could probably lose a pound and it would be nothing but advantageous.

Aart: Yeah. Example in case, you are the software and you're a little too soft here.

Ben: Of course. What's the old phrase about bicycles—n plus one is the right number of bikes to have.

Sassine: Yeah, exactly.

Aart: I have a t-shirt that says, just one more guitar. It's just saying one more guitar and you're going to be a great musician.

Ben: The difference between you and greatness is right there. That's awesome.

Okay, if I could perhaps paraphrase the two things that you said, it's this idea that, hey, what if we admit that density is going to be really, really hard from here to get even more density on a chip? So either (a) we can stop making so many trade-offs in the hardware to accommodate general purpose computing and just make specialized hardware, and (b) we can horizontally scale, we can just connect more dies together, so we basically have more compute. It's going to take twice as much space for twice as much compute, but at least we get twice as much compute.

Aart: The only thing I would slightly tune in to what you said is all these things multiply. You don't care at which layer of abstraction you can have an improvement. If you can improve the transistors by 5%, that's still 5% that applies to a lot of things. Not all things benefit equally.

It's been interesting to listen to some of the people that manage big compute centers. They would say, I don't care how much power you use on the processor because if you make the processor faster, I can leverage that on all the other chips that are expensive to buy or to run.

Systemic complexity is fundamentally defined in the simple math of multiplications, whereas scale complexity is mostly additions. We like to have more transistors, but it's the multiplicative effect that changes what you can do.

Sassine: In my mind, there are other factors too. Since you listened to my keynote, I called it on two vectors. One is the march to angstrom. There is always the opportunity to advance on Moore's Law. Then there is the march to the trillion transistors, which will only happen through multi die architecture decisions you need to make.

Technically, they're both doable. The decisions our customers are making are financial decisions. Does it make sense (let’s say) for your next phone to have a chip that may cost $15,000? The answer is no. But hey, if you can run AI on the edge, it's going to be very cool, it's going to be super fast. Yeah, sure, but you cannot afford it.

Some of those chips we're talking about, they're selling for $25,000–$30,000 a pop for certain applications because the yield is horrible, because you're pushing the limit of everything. The architecture is not only a technical decision. It's an economical decision you have to make. How much do you go down the angstrom? How much do you go up the architecture for that trade-off?

Ben: Put another way, this is an end dimensional space, where the dimensions are actually different for every customer and use case. And yet you are still trying to optimize them and produce the best product suite that you can.

Sassine: Exactly. That's why right now, when we are talking to our customers, we're not talking about, hey, we have this product design, whatever you want with it. We have end market-specific. We talk to automotive customers in a very different conversation than the mobile, than the data center for all those reasons.

That changed. Again, six or seven years ago, we did not have those end market focus discussions, because it's the same product you can develop on the same rhythm of Moore's Law, and life is great. Now there are all these trade-offs that you need to take into account.

Aart: The foundation is the same tech-onomics. Every technical decision is simultaneously an economic decision, be it for the build or for the use side of things. If you go back to the very point we started which was here's a synthesizer that creates functionality, which is the value, and it does that with the performance side and a number of gates, and the number of gates is essentially the -onomics that determines how expensive it's going to be to manufacture that.

That has now taken so many dimensions. If you look at the manufacturing side, these expansion boosts have been almost an order of magnitude over time because, okay, are you going to do a 300-millimeter fab now? We did 200 millimeters. The entire industry has to retool for that. You bet, it becomes much more expensive.

Right now, there's no visibility to do 400-millimeters, partially because it's too hard to coordinate an entire industry to get to that point. The economics, at some point in time, taper off. This is where innovation comes in, of course. Can you do it differently? Multi-die is an answer to that.

Ben: Fascinating. This question always seems to be divisive for people in this industry. Is EUV lithography a technology that's going to get us through over the next decade or two, or do we need to find a new better way?

Aart: Yes.

David: To both.

Aart: Yes and yes. There's still much to do. New generations of these machines are coming. At the same time, there's also a lot of development in the manufacturing from a material side with as much as possible self-aligning devices, where you don't need a mask for every layer that you depose and also for places where you can actually erode material under other things in a sideways fashion.

The reason I'm on purpose a little bit open-ended on this is because we have learned many, many times that saying no always turns out to be wrong. Being at an advanced semiconductor conference, as an undergrad student in 1978, the leaders in the field were all unanimous. Electronics is going to be big, and one micron, of course, is the physical limit.

Many years later, I had the opportunity to give a medal to one of the guys that said that, and of course, I couldn't resist bringing up what he had said. At the same time, it's so great to give the medal to the very person who had predicted the impossible and then was an engineer and made it happen to get around it. This is happening in the core, angstrom race, as Sassine mentioned. It is happening right there.

Remember, 15 years ago, FinFET, it will never happen. These vertical, wobbly things, it will never happen. For sure, they will never be in cars. And here we are. Engineering is very different from science. We work around science.

Sassine: EUV still has a lot of mileage that can serve. I don't know. I'm not too familiar with what will come next after it, but it's still an early adoption from a process technology point of view and when the transition happened.

A number of foundries resisted it for cost reasons, and fell behind. Then you're like, you know what, forget the cost. If I want to stay relevant, I need to be on it. What's next? I don't know.

Aart: The ASML folks, the technology leaders say they see another decade of delivery. We all think so, right?

Sassine: Yeah, it has mileage.

Aart: If they can't, Synopsys will help work around it.

Ben: Engineering.

Aart: Engineering, yes.

David: On that front, I'm curious. Synopsys is a wonderful company, great revenues, incredible market cap, all these things, but at some point along the way, you became something more too. You are one of a few linchpins in the system. Did you see that in the beginning of, if you look at any exponential function and it goes as long as Moore's Law effectively has, it's going to undergird the world eventually. When did this become apparent to you?

Aart: I think Sassine alluded to to one of the aspects on the technology side, which is when the relationship with the top foundries started to change, because suddenly they had touched some boundaries that they couldn't get beyond, we needed to get their information in order to be able to model what the circuits actually would be able to do. I will put every one of those under the notion of systemic complexity.

By the way, systemic complexity is not a last 30 years, not a future yes. Whatever you do, once you reach some boundaries, systemic complexity becomes the thing that you have to handle around that thing. The systemic complexity of a single transistor today is unbelievable, but we wouldn't have thought of it as super simple Lego blocks. This has happened.

The second thing is when the architectures started to somehow have a wishlist on physical behavior, which was far away from where they were. Suddenly they had certain desires of how fast to access the memories in order to be certain computations and vice-versa to domains that had been nicely separate for good reasons. Once they get closer and closer, suddenly they are one. That is a moment of systemic complexity.

There was a movement coming down from that perspective. In a whole different camp was the notion of globalization had been successful. Suddenly you dealt with parties literally all over the world that could only be successful by having a chain of participation and collaborations. If there's a singular skill that matters more in systemic complexity than anything else, it's a combination of trust and collaboration.

I think Synopsis emerged as, hopefully, trustfully good enough, but also needing and intending on collaboration. That was fantastic. Of course, the fact that there's deglobalization in the world in the last seven or eight years, it complicates things for many people. But at the same time, it's a skill set that's still relevant for the future. It's going to be way more relevant, not less.

Sassine: Maybe another way to answer your question, David, if you go back maybe 15 years ago, our industry was not that exciting. It was so hard to recruit. It was so difficult to bring in young, fresh blood out of school into not only EDA but EDA and semiconductor.

I remember when I was a GM at the R&D Product Development, one of the initiatives was, how do we excite the next generation to study electrical engineering, to study computer engineering? Because it was like, it's not exciting, I want to study more on the software side. Maybe that's what you were in your background.

David: When I started in Venture in 2010, we had a startup EDA-type company that was in the portfolio, and it was like the black sheep. [...]

Sassine: That's right. Exactly.

Ben: Who now I believe is a customer of Synopsis and designs their own ships.

Sassine: That's right. Exactly. Now it's very different, because there's a recognition that in order to drive that ambition of software of applications, et cetera, you can for sure buy a general-purpose chip, but you're not going to be competitive. How do you customize from the silicon all the way up to the system, to the application that you're designing?

That's why many companies who can afford it are trying to develop their own silicon or architect their silicon, because they know the importance of the silicon in the context of the software and the apps they're building.

If you ask me 15 years ago, do I envision we're going to be at this point, I didn't see it. We could see that we're going to march down Moore's Law. But now with AI as a huge opportunity to disrupt every market, then every market needs to go through its own transformation at the software level, system level, the way they're designing their end-product. What's powering it is the silicon, but not by itself in isolation, silicon in the context for each and market application.

Aart: Sassine just triggered a thought by bringing in the vertical market and making this vertical movement. What has changed is we started in a technology where it was a technology push, and then there was an economic success of the people applying it to software, whatever it was.

What has happened now is that technology push continues, but there's an end markets pull and having a technology push and an end market pull accelerate things substantially. Of course, meanwhile, everybody's inundated by big data. What the hell do you do with that? You need to process it somehow. By the way, it's going to change your business. Those are very big statements.

They come to the semiconductor world and say, I need something much faster, much bigger. We're racing forward, but we have direct impact in their P&L, on the profit part, on their differentiation. Whereas in the past, they partially looked at us, well, yeah, expensive tools. Now it's like we open the door with them and for them.

Ben: It's a heck of a tailwind as the world has this pull that you're talking about, but also as specialization of computing becomes more and more important to eke out that next frontier, you used to just sell to a handful of companies. Now there's a strong incentive for many more companies to design their own silicon.

Specifically, I think it's true that eight of the top 10 market cap companies in the world design their own chips. The only companies that don't, to my knowledge, maybe it's some secret project, are Berkshire Hathaway and Saudi Aramco. Your customer base has exploded.

David: Berkshire is half of Apple. Half of Berkshire is Apple, so they do.

Aart: Do you know anybody there we could call to help them?

Ben: Your customer base has exploded and the level of importance of silicon in their business has also exploded. You have this dual access tailwind that's helping you.

Sassine: Yes. Actually, the number that I typically share, that people get big eyes when they hear it, 15 years ago, pretty much 100% of Synopsis revenue was semiconductor companies. Today, 45% of our revenue—of course, we went from a $1.5 billion to $6 billion in revenue so the base got much bigger—are system companies. System companies meaning those are end-market OEMs that they develop. They don't sell chips, they're selling a product. That gives you a sense of exactly the point you're making.

David: Fifteen years ago, I can't imagine the CEO of Toyota would come see you guys, but today they are.

Ben: Or Ford, I think, literally is an example. Ford designs their own chips. Maybe every car company does now, but that was always a thing that they bought through intermediaries.

Sassine: Exactly. The key point though, even if you're not designing your own chip, right now, you're talking to Synopsis. Say you are an automotive OEM that has no intention to design your chip. However, you need to architect your electronics, given the context of electronics is going to get bigger and bigger and bigger, given electrification, autonomy, etcetera.

You, the automotive OEM, are hiring chip architects without an intention to design a chip so you can architect your electronics in the car. Those are customers of Synopsis, because we have software that enables them to virtualize the entire electronic system. That's the exciting opportunity of the future.

Ben: Fascinating. Okay. I can't believe we've gotten this deep in the episode without asking the question. In January, the news broke that you are making quite a large acquisition of a company called Ansys. What is the logic there, and how does it all work together?

Sassine: We touched on number of the why and how the world is changing. I want to describe it in two reasons why we're doing it. Reason number one, deep in our core business. As I mentioned earlier, the challenge of going down the Moore's Law is no longer an electronics-only challenge. It's electronics and deep physics when it comes to putting these devices in a chip—thermal, structural, et cetera. Ansys is the leader in simulation and analysis in those spaces, so that's in our deep core business.

The other vector is what we just touched on as well, which is many system companies. Let's continue picking on automotive as a system OEM. They're trying to figure out, how do I design my whole car that has bunch of electronics, that is going to trigger a mechanical action, that's going to trigger a number of other physics action, they call it multi physics meaning different type of physics analysis that you need to do, how do I design the car with a way to simulate everything up front, i.e. a digital twin of a car, including the electronics, the mechanical, et cetera?

Again, answers is the leader in the simulation and analysis of that multi physics. We see the opportunity at the silicon level and at the system level. That's why we're describing our company as the design solution from silicon to system. We're looking forward when you bring two great companies to really deliver the engineering platform of the future. It's an awesome opportunity we're excited about.

Ben: I may be oversimplifying here, but there are some sets of things on the EDA side that Synopsis does really well, there are some things that cadence does really well. But in simulation, there's basically just Ansys. Everybody needs Ansys. Does that feel like it's a reasonable characterization?

Sassine: Remember the discussion we had in the beginning? There's the cumulative learning that you have in order to be the trusted simulator. Ansys in number of simulation, when I say they're the industry leader, meaning they're the trusted simulator because they've had a history of 40-plus years of cumulative evolution off their simulation.

To be clear though, in every space, same as you described in the EDA, there is Synopsis, and there are a number of other companies that we we compete with, their space is the same, but their their history of that cumulative learning, they have the leader and having that history and that sign-off trust. Once you do the simulation, can you sign-off that I can trust that outcome? That's the key in what they offer and what Synopsys offer.

Ben: Fascinating. Simulation is such an interesting area, because on the one hand it can help your customers do better or literally use your own existing software packages better. In addition to that, now that all of this new hardware complexity and AI exists, we are going to get better as a species of simulating way more things in the world. It also creates more demand for everything else that you make to the extent that the market for simulation broadly is going to grow.

Sassine: Exactly. One of the theses that we explained to our investors after announcing the acquisition is to picture the world five-plus years from now. Physical testing in the context of whatever that end device that you're physically testing is going to become more connected and smarter because it's going to have some chip in it, because it's interconnected and smarter. It gets too expensive, too long, and just not practical to physically test stuff. Simulation plays a huge role.

In that same context, when you think of simulation, you think of digital twinning stuff, you think of virtualizing stuff. This is where we see our core competency of what we've done at the silicon level because that's what you do. When you design a chip, you virtualize, you model, you simulate. You can do it now for much bigger systems than the chip.

Ben: Simulation has always been a good idea, it just wasn't technically possible for that many use cases before. It seems like we're now getting more and more fidelity in the physical world of simulating more complex projects.

Sassine: And you have accelerated compute. Before, it may take you weeks to simulate a very small function. Now with accelerated compute—one of the slides that was presented at GTC and in my keynote—was 10X, 15X, 20X speed up due to accelerated compute. That's a massive speed up.

Then you layer on top of it an AI for further acceleration. How do you get smarter? What do you simulate more effectively, efficiently, etcetera, using AI techniques? It's opening up the door exactly what you said, Ben, for more applications to simulate.

Aart: You're describing the company mostly through technical terms, but the reality is it is a group of people first and foremost. One of the things that has helped Synopsys precisely in this notion of all this learning that Sassine was talking about over the years accumulating, that has been enhanced greatly by having many people work here for a long time. Both Sassine and I are examples of that. At the same time, continually rejuvenate both with new people or different people, but also in our own learning of how you do things.

I think that is a fairly unique characteristic. Of course, there are some companies that we admire greatly because they showcase this. You mentioned NVIDIA, but I would certainly put TSMC also in that category of never quite being satisfied with yesterday, and yet tomorrow is impossible. But tomorrow is only 12 hours away, so you better get going.

There's a passion for making the future happen that is probably half grounded in utter paranoia, as in only the paranoia survive, and partially also in a belief that things are possible that we still have to invent. That is a very unique recipe for companies. I think that is certainly one of the things that characterizes Synopsys.

Ben: Well, thank you both for your time. My closing question that I've really been enjoying recently is, let's flash back all the way to where we started the episode, both of your first experiences with Synopsys. What is the same today as it was then? And what's something that couldn't be more different?

Sassine: The passion towards innovation has always been there from day one. Aart just made the last comment he made. It's always been what we're working on tomorrow is almost impossible, very difficult to do. That's an industry, that's a privilege in our industry because that's the key to innovate.

You talk to our engineers, they love the fact they're working on the most complex things known for humankind. That's not only not stopping the opportunities right now to monetize it. The opportunity to be at the center of what you're working on is so relevant to many inflection points that are happening in front of us. That's thrilling in my mind.

Aart: I probably, maybe not by accident, land on the same as Sassine has, which is we've touched the exponential and it's in our DNA. That sucker is not going away. It takes different forms, but 10X is still 10X. The next 10X is of the old 10X. That rate of change is just exceptional and having been, in some ways, somewhat central. A big piece of that roadmap is a privilege that is amazing.

At the same time, if you look at the rate of change for us as a company in terms of size and then of course, complexity but also of the world, we started this where the far east was not very important yet, and today it's one of the dominant parts of the high tech ecosystem. It's also part of one of the big stress fields in the world that adds an enormous amount of complexity.

Being now a company that is in the middle of these type of things, that's only needs to have opinions, but also careful actions of how you behave in a political minefield, how you behave in a situation where you see our industry is going to go to about 10X more energy utilization, and without all the the ramifications of touching what is happening in terms of climate change.

An industry that simultaneously we have had multiple countries in various states of doing well or deep war, and how do you deal with that brings a set of questions to us that as leaders, we have to learn just as much there as anywhere else.

It's interesting to what degree companies are now counted on as both, I don't want to say too strongly moral centers of gravity, but certainly value centers of gravity, as people are finding difficulty finding it in the political environments or in some cases, finding it really well, some cases not finding it at all in religion. Now the question is, what societal groupings matter?

We, for a long time, said that they who have the brains to understand should have the heart to help. In other words, we're co responsible for the communities that we're part of. By now, the community is humanity. We've started to modify it a little bit into they who have the brains to understand should have the courage to act. That is different than before.

David: I can't imagine that was part of the business plan that you were going to Barnes & Noble to figure out how to write—

Aart: We participated in a march in the support of people having AIDS in the 1990s. There were very strong opinions that said, well, that's not cool because, because, because. All things that today we think as, this was Middle Ages thinking. There's still a lot of Middle Ages thinking now.

I don't know where that leads. I think we have the great fortune to have not only the leadership that can give the next decade of moving it forward in a company that does well, but at the same time the question, what position do we take in the world that is more than Synopsis as a tech, maybe super tech company but as a tech company, as a human company? Those are interesting questions.

Ben: I can't imagine a better place to leave it. We're going to have to do another episode to explore all of that. Thank you both so much for your time.

Aart: Thank you for your great engagement.

Sassine: Yeah, that was fun. Thank you, both.

Ben: Listeners, we'll see you next time.