Today I have the privilege of having Mr. Prithviraj Banerjee here. He is currently the CTO at Ansys, and welcome, Prit.
Thank you, thank you, Raj. Very, very good morning to you and everybody else. Yes, thank you.
And where are you calling in from today, Prit?
I'm calling from the Ansys office of San Jose, California.
Awesome, wonderful, Prit. We're just going to dive into this podcast here. Just tell us about yourself. Who are you, why are you there, and what do you do?
Sure. So I am Chief Technology Officer at Ansys. Ansys is a modeling and simulation company. We help our customers in building the most amazing products, and my role as CTO is to advise the company about the long-term technology strategy: where is simulation headed, where is Ansys headed, where should Ansys invest in terms of new technologies. And I've been doing this for the last six years. Prior to Ansys, I have spent some time in academia. I spent about 20 years in academia. I was involved with a couple of startups, and I've also worked in several large companies like Hewlett-Packard, ABB, and Schneider Electric.
Awesome. So I love the word simulation in technology. We use a lot of simulation. Virtual machine, you know, goes back to a long time ago when all of this started. Tell us in the real world where do we see your product? Is it like simulating a car race or simulating crashes?
So actually, I'm glad you asked that question. Right, so simulation is used by designers across different industries, right, in designing and simulating amazing products. So, for example, we serve customers in the automotive industry, and there we are helping people design cars. Right, and so when you say design cars, right, there is the OEM like a Ford or a General Motors or a BMW who will use our simulation, but the simulation tools will also be used by the tier-one suppliers that supply parts to Ford and GM, for example, Aptiv, Bosch, Continental. It will also be used by the tier-two suppliers like Infineon and NXP that make those ECU chips that go inside those parts that go inside the car. So the entire supply chain of the automotive industry uses our simulation tools. Or let's look at the aerospace industry, right? The OEMs are like Boeing and Airbus, right? They will use our simulation tools to figure out how that airplane will fly. But not only that, the tier-one suppliers, the engine suppliers, the General Electric, Rolls-Royce, Pratt & Whitney, they use our simulation tools to simulate those engines. And then the software platform companies like Honeywell, right, that use to control all the things on the airplane, they are also using our simulation. Next vertical, say energy, industrials, right? All the energy companies, the TOTAs and the Chevrons of the world, right, as they're sort of extracting oil, they use our products. The ABBs and Schneider Electrics of the world that make power and automation equipment, they use our products. Next vertical, high-tech industry, right? So all the chip, electronic chip companies at the sort of Intel and AMD and Broadcom and Qualcomm, and all the cloud companies, the Microsoft Azure and Amazon AWS, they are our customers. Final, final sort of things is healthcare, and like companies like Medtronic and Johnson & Johnson that make medical device equipment, or companies like pharma companies like Bayer and Novo Nordisk and so on like that make pharma drugs. So we have more than 15,000 customers across these five verticals, and they're all using it to design the most amazing products.
That's insightful. I had no idea that there was a need and someone is fulfilling this need. So walk us through how did you build this technology? What does your team look like? I mean, sounds like there's a lot of developers. Tell us how many transactions you're dealing with. Give us the complexity and how you're solving this problem.
Yeah, so the way we solve the problem is we look at the world around us, which is governed by the laws of physics, right? So when you look at this glass that I'm holding here, right, I can say I want to apply this much pressure, and when I apply this pressure, the cup bends. This physics, the physics of structures, is governed by what are called second-order partial differential equations called Euler equations. Or there is coffee inside this, right, which is governed by fluid flow that is governed by Navier-Stokes equations. Or this electronic thing that we are talking to each other, right, is governed by Maxwell's laws, electromagnetic equations. So the world around us is governed by Newton's law, Maxwell's law, Navier-Stokes equations, and so on. When we take those fundamental differential equations and we solve it numerically, that is what we do, right? So Ansys got started more than 50 years ago by John Swanson, who used to work at Westinghouse Corporation in Pittsburgh, and he says, well, I want to solve this structural problem using a technique called finite element analysis. So that was the beginning of Ansys, and then over the years we have moved into different physics like fluids physics and electromagnetics physics and so on, right, through organic R&D and through acquisitions. And today we are a company of about 6,500 employees. We have about $2.1 billion in revenue. We have more than 15,000 customers. We have more than 10,000 transactions of sort of software sales that we do to our customers across these five verticals, and the software is sold sort of full perpetual license or lease or on the cloud.
So talk to me about the technical implementation. How does that look? Without, of course, we don't want any secrets, but talk to me about the transaction volume. Talk to me in terms of response time. Like I drive an electric car, and it has, you know, full self-driving capability, and it's able to think as soon as it sees an object, and the response time is incredible. Talk to me how you deal with such complexities in tech.
So when I said we will solve a physics problem, like for example, you are sitting in a car, right, and the car can crash. And so what automotive customers do is they'll put a dummy inside the car and they'll ram the car at 70 mph, hit a thing, right? So but when you do that crash testing, that car gets destroyed, right? So what if you could do that purely through simulation, virtually, right, without actually destroying the car? That's where we come in. Now, a problem as complicated as car crash, right, is so complicated, right, you would need maybe 1 million finite elements to solve that problem, and this problem may take just one sort of second of simulation. Physical simulation, physical thing, right, would require maybe 20,000 hours of compute time on a really sort of fantastic computer, right, to solve it accurately. So our simulations take a long time to run just to look at sort of one millisecond of real time that you're trying to deal with, right? Interesting. And so in order to get the sort of faster response times, we do look at things like using high-performance computing, HPC, and sort of shared memory multiprocessing, message passing, GPUs, all the stuff that you hear from Nvidia and Jensen, all those Blackwell GPUs. So we are using all of that technology to accelerate our simulation. We also use AI, machine learning, to accelerate our simulation. So essentially, these simulations are very accurate but take tens of thousands of hours to run for really complex designs, and we use high-performance computing or cloud or AI, machine learning technology to speed things up so that the designer will be able to evaluate multiple options and design the most optimal product.
That's powerful. So you touched on AI/ML, and that's my next question. Tell me how you're using AI/ML or deep learning or supervised, unsupervised learning in this context.
So AI/ML is used in simulation for two broad areas. One is to accelerate simulation, right? So as I mentioned, these simulations oftentimes take a thousand hours to run, 10,000 hours to run. So the first use of AI is to train an AI model with maybe a thousand different designs, right? Here is this design with this input condition, this boundary condition for this input. Here is the output, right? For this input, output, you train the AI model, this sort of thousands of neural networks, right? All the weights of those neurons, right? What should those weights be? And you do this over maybe 10,000 different designs. The neural network learns the designs, and next time you give it a new design, instead of running the full numerical methods that take sort of tens of thousands of hours to run, it would run in a matter of minutes. And that's called AI. And we have a tool that we have recently launched called SimAI that can literally accelerate simulation by factors of 100 or 1,000. So that's accelerating simulation. The other way is to make simulation easier to use, because you know, Rajiv, these tools are super accurate. They're really, really good, but they are very hard to use, right? You almost need to have a PhD in electrical engineering. I have a PhD in electrical engineering from the University of Illinois, right, to use our HFSS tool for electromagnetics, right? You almost need a PhD from the University of Illinois to actually know what settings to use to get the best quality and speed and convergence, right? So it is not that easy to use, right? The most advanced users know how to use it. So the role of AI is to sit and watch what an advanced user is doing, right? You have this problem. Oh, this is a multiple-input, multiple-output antenna problem. For this, you have to use tetrahedral meshing or tetrahedral meshing and so on, or you have to do it this way. So an AI will watch what a smart engineer is doing and then basically sort of do it automatically for this thing. And so we have just announced an AI tool called AnsysGPT, which is based on generative AI technologies to make the tools easier to use.
That's interesting. I used to joke around that you need a PhD to do this. Finally, I know that there is one problem where you truly need a PhD to do it. Talk to me about the generative AI product that you just launched. What is that?
Yeah, so again, it's a sequence of things that we're working on. So basically with GPT, as you know, right, you can ask a question, right? Hey, summarize, tell me about Prit Banerjee. Right, you will go through and find out all kinds of information on Prit Banerjee and give us saying, hey, what does Prit Banerjee do at Ansys, right, in 100 words or less. So what a GPT does is it scrapes through all kinds of information in the world. You ask a question and it'll give you an answer. So the first product that we have launched off this sort of GPT, ChatGPT technology is called AnsysGPT, which says we have, as I said, more than 15,000 customers, right? And they're always asking questions to our application engineers. Hey, I'm trying to run HFSS for this problem. What setting should I use? And then the Ansys engineer responds back saying, oh, you should set it this way. Then another question comes in, oh, I'm trying to run Fluent. What should I use? Oh, you should do it this way. So there's a question that comes in and there's a response from the engineer. So you have taken over the last sort of 15, 20 years, right, all the questions, all the answers, put them in ChatGPT, right? And it's our own on-prem version, not the one on the cloud, right? And using that, we have trained AnsysGPT to answer very accurately, right? All those RAG models and so on and so forth, right, to be with no hallucination, to give very, very accurate responses, right? And it's a 24-hour, 24/7 chatbot, right, which is super intelligent but based on all the questions and answers that have come from different customers to our Ansys engineers. And that's sort of the first thing that we have launched called AnsysGPT. So very contextual, on-prem.
And it's interesting to build that model. Do you use any third-party apps or you built it from scratch?
No, we built it from scratch. Well, we use the OpenAI sort of ChatGPT, but not OpenAI. We have a partnership with Microsoft, so the Microsoft version, I guess the Copilot, right, is what we use to build this. But we trained the model. But we didn't use any other third party. We just took sort of OpenAI ChatGPT, but the Microsoft version Copilot, and used it to create this thing. But where, so this is the first version, and then using this, what we're doing is, so this tool today, you get a question in English and then you respond in English, right, to a human. But what if you could go the extra mile and actually do the settings of the tools? Because in the question and answer, it says, oh, you should use tetrahedral meshing. Well, it turns out that you can actually go into the settings, right, using Python script to actually do the setting of this thing to run the whole thing automatically. So we have actually shown, this is not a product yet, but in terms of some early work that we have demoed to our customers to say, hey, you give it a question in English and it's almost like, okay, Siri, turn on the alarm, right? You say, okay, Fluent, run this, and it will run Fluent for you based on your English language specification.
So you're taking prompt engineering to workflow now, right? You're extending that. That's really cool. Now, I know there's a lot of innovation you're doing. You're seeing a lot of innovation in the industry. Talk to me about what innovation you wish you had today that doesn't exist.
Well, the ultimate innovation is, I mean, designers want to design a car, right? And so they would like to take a specification of a car like, hey, I want to design an electric vehicle which will go from 0 to 60 in 1 second and have a range of 1,000 miles. I mean, if I could build that, right, I would be the next Tesla, right? But you're doing that. So what happens is an engineer will go in, okay, I need to have 0 to 60 in one second, so I need this huge motor, right? It's a permanent magnet motor or an AC motor, DC motor, whatever, right? And look at, do I need a 100 horsepower motor or a 1,000 horsepower motor and figure it out, the size of it, right? Then they say, okay, so because of the motor which weighs maybe a thousand kilograms, right? And then you'll have four passengers each of whom will weigh, say, 100 kilograms, right? Now you put it all together to have a range of this kind of miles, right? I need a huge battery, right? Well, so that's the size of the battery. So once you put that gigantic motor and the gigantic battery, your inertia becomes so large that it doesn't go from 0 to 60, right? So there is, oh, now I have to iterate. I don't think I can have that big a battery. Make the size of the battery wrong. So this is how designers actually design products, right? They'll take a specification, break it down into subsystems, right? What should the motor be? Based on the motor, what should the powertrain be? Based on that, what should the battery be? And based on the battery, what should the power electronics be? And they're just doing these different parts to do it independently. And then I start writing the software to drive the motor and this thing and so on. And it takes seven years, right? This is why it takes seven years for a new, completely new car to come in the automotive industry, right? So the innovation is, what if you could take a specification of this car that you want, right, and just push a button, everything will be done for you, right? And which is sort of what we call the field of digital engineering, right? Sort of synthesis of what I call cyber-physical systems. It's a hardware-software combined system coming out, right, with a push of a button. It doesn't exist today, but that is something we are working on.
Very cool, very cool. Now, I'm curious, you have a fairly large team. How do you keep tabs on making sure projects are running smoothly, making sure everyone is being productive and the right technology is being built, and simple things like coding standards are being met, security is in place? Just maybe spend a minute on talking about what does your culture look like, what's your organization look like, and how do you keep things before they go out of control?
So as I mentioned, Ansys today has about 6,500 employees, right? We, as I say, a $2 billion company. Of those 6,500 employees, about 2,500 are R&D developers, right? They are the ones who are writing code, right? And in addition, we have about 1,500 application engineers. These are the people in the field that help customers with those things, right? So the 2,500 R&D developers and the 1,500 are the 4,000 engineers, many of the PhDs that I would call are the intellectual capital of Ansys. So in my role as CTO, I'm responsible for these. They don't directly report to me, but it's like I'm the cheerleader of this 4,000-person engineering team, right? And so what we have done is we have taken these individuals and we have approximately, broadly speaking, 10 sort of product groups. Like so there's a product group around structures, a product group around fluids, a product group around electromagnetics and semiconductors and so on, right? So these sort of 2,500 R&D developers across 10 product groups, that's how these R&D groups are organized, right? Then what we do is in order to make sure that these people have the right intellectual capability and the learnings and so on, we have broken down sort of our areas into five technology pillars, right? And those are things that are organized and orchestrated by the CTO office. So the first pillar is numerical methods, right? I mean, in our solvers, we have to use the latest numerical methods around finite element, finite volume, finite difference time domain, different methods, multi-domain, multi-scale, multiphysics simulations and so on, all the geometry and meshing. So we have a center of excellence in these areas led by a couple of the smartest Ansys fellows, and they do a regular meeting about, okay, making sure the different people working on numerical methods across these 10 different product groups are all aligned and they're using the latest technology. If there's some latest stuff that is happening on sparse matrix solvers, they immediately bring it in and make sure that everybody's aware of it. So the latest technologies out there, right, coming from academia or in conferences, are immediately being used in different tools. Second pillar is in high-performance computing, right? The latest technology in shared memory, message passing, this, that, GPU. Oh, there is this thing coming from Nvidia, the Blackwell machine, or the AMD's MI300, or here is this QDSS sparse solver coming in, right? So the HPC Center of Excellence is making sure that the latest HPC technology is leveraged across Ansys. Third pillar, AI, machine learning, right? The latest work about RAG models and large language models, right? Whatever you're doing, right, everybody's aware of it, and we are using AI across different products. Fourth one is around cloud and platform, right? I mean, what should you do in terms of working on the cloud for Azure cloud or AWS cloud, cloud marketplace, this thing, access and so on, software as a service. And the fifth one is digital engineering. This is what I kind of talked to you about, right? Taking system-level specifications and automatically able to generate different things called digital twins and so on. So the way we coordinate our technology strategy and the product strategy is making sure different products have product areas, right? Oh, we are going to have this feature in 25R1 or whatever. What kind of technologies do we do? So there's a very nice interlock between the product strategy, these are the features we want, and what kind of technologies we'll need in order to do this. And of course, we do the right things with respect to security and so on and so forth as we are building these products.
