The $35B Bet on Physical AI: Synopsys SVP Prith Banerjee on Agentic AI & The Future of Chips

Ansys was the leading developer of systems simulation tools. And in each of these areas, you have to design a system on the computer before you actually make the product. How do you do that? With the combined power of Synopsys EDA tools and the system simulation.

In the past, you designed a new chip every maybe three years. Now, Nvidia, Meta, they're all talking about designing a brand new chip every year. So, the complexity of the chip has increased and the pace of innovation has increased.

Innovation is defined as something that is novel that the world has not seen before, but it also has to have impact.

Here is a grand theory of innovation that I would like to say. Right. So...

Thank you very much.

Absolutely. Thank you very much for inviting me here. So my career actually has had three phases. I spent my, again as you mentioned, I did my undergrad from Kharagpur, PhD from Illinois, but the first 20 years I was in academia. I was a professor, a department head, and a dean at the University of Illinois and Northwestern. So that was phase one. Phase two was I did a couple of startups from the University of Illinois, Northwestern, Exel and Banachip, which I sold to other companies. And then the last 18 years I've been in large R&D labs of companies like Hewlett Packard, HP Labs, I was CTO at ABB, Schneider Electric, CTO at Ansys, and now Synopsys. So I feel fortunate to have been part of innovation in three different areas: in academia, startups, and large companies. But what I'm really excited about is in my current role, right, Synopsys used to be the leading electronic design automation company in the world, right? Where you do a chip design, you design a chip with a billion transistors, 10 billion transistors, and so on. So there was electronic design automation tools that Synopsys did. They also were the leader in IP, intellectual property. When you have a large chip, you have actually IP blocks like PCIe Express blocks or Ethernet controller blocks and so on. So who creates those blocks? Synopsys does. But that was the world of chips.

But what was happening is the world of chips is becoming the world of products, intelligent products. When you have a car, the car will have a chip inside the car. That is the brains of the car that drives the autonomous driving of it. But there's a chip, there's software running on the chip. But then the chip is actually controlling the actual physical stuff, which is the world of the physics and physical AI that we'll talk about. So Ansys was the leading developer of systems simulation tools, the structural simulation tools, the fluid simulation tools, the electromagnetic simulations. And that was the world I used to work in. I was a CTO of Ansys. So what we now see is this revolution of chips to systems design, right? This intelligent products in every industry, be it automotive, aerospace, high-tech, energy, healthcare. And in each of these areas, you have to design a system on the computer before you actually make the product. How do you do that?

With the combined power of the Synopsys EDA tools, the IP, and the system simulation. So I am fascinated to be in that leading edge of innovation, driving that innovation for our customers in this world of chips to systems. And the world we kind of talk about has sort of three parts. These things, they are silicon designed, right, because you have to have a chip which has a billion to 10 billion, 100 billion transistors. It is software driven, the software-defined vehicles, software-defined airplanes, and so on. And it is AI enabled. It's all powered by AI. That's the world that we're talking about.

Absolutely. By way of acquisition. So as I said, Synopsys earlier had only the moat around chips, the silicon chip itself. But as I said, the chip is actually now going inside an intelligent product, which is a physical system now, the physical system that we see around us. For example, the car. When you're designing the car, that car has to, once you sit in the car, it has to be crash safe. Means you bang the car, and if you had to sit in the car, you will not get hurt. So the way people used to design cars in the past was to build a model of a car, and they actually crash tested the car, right? And I don't know if you've seen videos of a crash testing, it's just incredible. And they put a dummy inside the car and they measure whether the dummy will survive or not. Each crash testing costs like $23 million. You make a prototype of a car, you bang it. Now that you don't have to do anymore, this crash testing is done with Ansys tools called LS-DYNA, full crash simulation. You can do thousands of simulations of a car. So without actually doing a hardware prototype, do it through simulation. Or I know you fly on airplanes. The Boeing airplane or Airbus airplane that you fly on, right? In the past, they used to build physical prototypes of an airplane wing, put it in a wind tunnel, and test that wing at different speeds and so on, whether the wing will lift or not. Each wind tunnel test was $10 million, right? So you do a shape of a wing, you test it, it doesn't fly, or there's a drag and you fall down. Oops, mistake. $10 million down the drain. You can do that through physics simulation, fluid simulation with Ansys, right? You are designing a very complicated antenna design with all these 3G, 5G, 6G antennas and so on. And it's very, very complicated. You don't have to do that with a physical prototype with all those five bars that you see on your phone. You can simulate the electromagnetics with HFSS. So the power of Ansys is all those things that people had to do with hardware prototyping, you can now do on the computer with detailed physics simulation because the world around us is governed by physics, and physics doesn't lie. And we at Ansys and now at Synopsys, our solvers model those things with numerical methods, partial differential equations, with the accuracy that you don't have to do a physical testing. So now what we are enabling is the reason this was a $35 billion thing. We were the leading provider of these tools, and our revenue was about $3 billion at the time, right?

Yeah. And the reason for this, after we were acquired by Synopsys, is now you look at the problem of these multi-chip systems, multi-chip modules, 3D ICs, chiplets, right? These large complex systems having, like for example, Cerebras just did an IPO, just 2.6 trillion transistors. These are complicated systems, right? So these have chips in them, chiplets in them. These chiplets are all connected together, 3D stacked dies and so on. And in the past, you would have to simulate them, design them individually. Okay, let me do the design of the chip with the EDA tools, and then, oh, you want to do the simulation, the thermal simulation, the electromagnetic simulation, you have to do a separate tool. Now that we are part of the combined system, we can do the full multiphysics fusion analysis all together with the integrated tools. That is why it is such an exciting combination, and that's why Synopsys paid so much money for it.

Fantastic. So the first large company that I was involved with was Hewlett Packard. HP built large servers, laptops, printers, and so on, right? And as you are designing those things, you have to actually think of the physical term of how a server is actually designed and so on, and you have to simulate them. Then I went to ABB. ABB is a leader of power and automation tools, and so you had these large robots that were being designed by ABB, the large power distribution, the transformers, breakers, etc. These are physical devices. Then I went to Schneider Electric, very similar area in power automation. In each of these areas, you have a physical product. Let's take the ABB robot, right? That robot has its sort of mechatronics, means you have mechanical parts and electronic parts all working together. When we used to build those robots, we used to model the physics of the mechanics of the robot, and then you have to program the robot to say how it would work. Right? That was the world.

So basically, when you have a human, right? The human hand moves around, right? And so you have this and it's kind of moving around. Now in a robot, you have to say the robot is, say, I have this thing. I have to pick up this thing, right? So the human hand goes and moves around and moves this thing.

So you have to do the same thing with a robot. The robot arm has an arm that will move like this. It'll move like this. It'll move like this, and so on, right? And then you have to do the appropriate controls to make it move like that. And then as you do it, you go and grab it and you pick it up, right? So you have to model the whole thing, and that's what we used to do when I was at ABB using these modeling simulation tools from Ansys to design this robot that would do those things. So suppose the robot has to lift a weight which is 10 pounds, right? Will it be able to do it or will it slip? So we were able to simulate the design, the actual operation of the robot purely through simulation.

Very, very.

And so, and I'm so glad you mentioned this. So when I, in my seven years as CTO at Ansys, I used to work on working with the R&D teams on taking something that is taking such a long time, how can you speed it up? There were a few approaches that we did. We looked at the sort of numerical methods itself just on a single processor to make it go faster and be more accurate, etc. The second thing was to use high-performance computing using parallel computing, shared memory, message passing, GPUs to speed things up. Then we also started looking at the use of AI to speed things up. So we used AI, we used cloud, we used GPUs, HPC, all of it to speed things up to get it to that level of accuracy. But to come back to the point of this agentic AI that you're talking about, right? So now the transition is now these systems. 10 years ago, you designed a chip with maybe a million transistors or 10 million transistors. Then it became 100 million transistors, a billion transistors, 10 billion transistors. And in the past, you designed a new chip every maybe 3 years. Now, Nvidia, Meta, they're all talking about designing a brand new chip every year. And it's not a chip with a million transistors, a chip with a billion transistors, with 100 million transistors. So the complexity of the chip has increased. And the pace of innovation has increased. So imagine that if one human for designing a one million transistor chip, right, it took maybe 100 engineers to do a thousand million, it would take 10,000 human engineers, and to do it twice as fast, so it would literally take millions, tens of millions of humans to do it. In fact, just the previous hour, I was in a cognitive silicon systems design workshop. These are our top customers, these are the Metas, the Googles, the Nvidias, the AMDs, Intels, all of them, right, working with the EDA providers. And I was representing Synopsys as to what the challenges are. And they were saying the cost of a new chip fab is about 700 to $800 million. You submit a fab, and that's what it costs, right? You cannot afford to make a mistake there. And these transistors, these chips have not a million transistors, they have 100 billion to a trillion transistors. As I said, Cerebras has 2.6 trillion transistors, right, in their system. So it's an incredible pace of innovation and incredible complexity to handle that. Instead of having, you cannot hire 10 million engineers, 100 million engineers. So we have therefore developed the technology of agentic AI. Agentic systems are used in all kinds of techniques, right? Anything that a human can do, an AI agent can do it. So we have created this whole concept of agent engineers at Synopsys, where every one of the tasks that a human does, using for example our Fusion Compiler, right, to do the actual chip design, or PrimeTime to do static timing analysis, or a TCAD thing, for any of these tasks, we are creating agents that will do the job on behalf of the human. And so the early part was simple agents, like let's write an RTL, right, a test bench agent or a verification agent or whatever. And now we have created this technology of level one through level four, just borrowing the concept from the autonomous driving world, right, where L1 driving is just assisting the driving to do the parking or maybe cruise control, and L4, L5 is fully autonomous driving. Similarly, we are imagining in the world of chip design and system design, fully autonomous design of these chips and systems.

With a little bit of intervention from humans, right? So there'll be a human in the loop to basically control the overall stuff to make sure things are going well, but it will make each of the human engineers super engineers.