Physics Simulation for Chip Design Used to Take Days. This AI Does It in Seconds.

Hardik Kabaria: Today, to get language reasoning capability, you don't have to be an AI researcher. You just need a phone, open chat GPD, call Anthropic, and ask a question to get an answer. We are building the same for Physics.Jordan Metzner: Built this week, breaking it down. Built this week, we show you how. A fresh idea, a clever tweak, you locked in.Sam Nadler: Built this week. Hey, everyone, and welcome to Built This Week, the podcast where we share what we're building, how we're building it, and what it means for the world of AI and startups. I'm Sam Nadler, cofounder here at Ryze Labs. And each and every week, I'm joined by my friend, business partner, and cohost, Jordan Nutsner. How are doing today, Jordan?Jordan Metzner: Hey, Sam. How's it going? Exciting week. Big week in AI. New announcements from OpenAI, Meta, Anthropix, seeing new live models coming out.Obviously, excited for 5.6, Fable going away and coming back. So it's been another crazy week. Obviously, there's there's a ton of headlines, and even a new Grock model potentially. So, yeah, just lots of stuff going on across the major laboratories and in the AI world in general, and, obviously, especially excited for, this week's guest.Sam Nadler: Yes. Lots to talk about. Every week is full of news and exciting developments. And I would love to introduce our guests, but quickly before I do, don't forget to like and subscribe. We've hit 30,000 subscribers on YouTube.So, please like new episodes out every Friday. And with that, I'd love to introduce our guest today, the cofounder and CEO of VINCI Hardik. Hardik, if you don't mind, give us a quick intro of who you are and what you're building.Hardik Kabaria: Well, first of all, thanks for having me. My name is Hardik. My technical background is in physics and geometry software. I've been in this space now twelve plus years. Did my PhD in the space and worked at other startups, but then realized that this is a great time to be in the building mode.So jumped off the boat two and a half years ago started Winchy. And our mission is to make physics accessible for everybody building hardware. Whether you are designing bottle caps or semiconductors or electronics, racks in the data center, you care about its physical performance. The moment you start designing, you care about how is this part going to behave and that basically guides your journey from your conceptualization to production. And that's the place we are deploying our software.We make physics analysis accessible to anybody who's building hardware at compute capacity. Sort of the same thing that language models have done and created language reasoning capabilities for everybody. Today to get language reasoning capability you don't have to be an AI researcher, you just need a phone, open to RGBD or Anthropic and ask a question to get an answer. We are building the same for physics, and we started deploying for semiconductor electronics customers, and we'll tell you more about it. But that's who we are, and that's what we are building.Sam Nadler: Very cool. So I I think we have a little video demo that I'll queue up right now, and then we'll get into some of the details. So let me get that going.Speaker 4: In semiconductor packaging, warpage determines whether products ship or don't at every level of assembly. Bonds form only when warpage stays within tolerance. When it doesn't, there is no electrical connection. The tolerance is measured in microns. The problem is becoming impossible to ignore.AI systems demand larger packages, more dies, denser routing, and increasingly complex structures. As complexity grows, so does warpage. Some of the industry's most advanced products have already been redesigned, not because compute was impossible, because warpage couldn't be predicted early enough. This isn't a problem the industry ignored. It's a problem the industry learned to live with.Physical experiments are accurate, but each design iteration can take months. Calibrated models work until the design changes. Simplified simulations produce results, but only after discarding the geometry that determines real warpage. Different approaches, the same compromise, because no practical alternative existed. For decades, engineers adapted their workflows to the limits of computation.What you're seeing changes that. This isn't simply faster simulation. It's the first time manufacturing resolution warpage analysis becomes computationally practical. The native design, every layer, every metal line, every via before engineering decisions are committed. Three micron resolution, 225 degrees Celsius bonding temperature.Run, 392 microns. But the magnitude isn't the breakthrough, the shape is. The warpage follows the actual layout, the metal distribution, the residual stress, the real geometry. Simplified models can approximate the answer. They cannot reproduce this, and this is the information engineers need before products are built.1,200,000,000 degrees of freedom. Manufacturing resolution, native design, under four minutes. Not an approximation, not a surrogate, a fundamentally new engineering capability. For decades, manufacturing resolution physics was something engineers validated after design decisions had already been made. Today, it becomes something they can reason with while the design is still evolving, continuous physics reasoning.Jordan Metzner: Well, awesome company, and great to chat with you today. Obviously, a lot to learn, and I've I've worked in some simulation environments, especially around robotics and things like that, which were, you know, simulation has probably been around for probably a pretty long time, especially, like, using video game engines and things like that. So just tell us a little bit more about how, you know, simulation is entering the hardware space and, you know, kind of what impact it's having there.Hardik Kabaria: Yeah. So let's start with some basics. Like, the moment you wanna design and develop a hardware, you care about its physical performance. And for several decades, how has the industry attacked that problem? There are physics simulation software, there are physics simulation methods that have been around since the sixties, and there are tools that have been developed by Cadence, Siemens, and Synopsys that has been used by industry.But ultimately, the industry performs physical experiments. Whether it's robotics, whether it's consumer electronics or semiconductor, you design a part, you prototype, you take it to the lab, and you do perform experiments and figure out whether your hypotheses are correct or not, and if not, you kind of go back to the basics. Right? That's the physical world of experiments and then there is the digital world of experiments. There are tools across the board to do both the things.So the need of understanding physics in the digital world as well as physical world is absolutely there and you care about it across the life cycle of the product. And what we bring to the table is to do that faster, at a higher capacity, so it should be limited by how many computers you are deploying the software to, not limited by the number of people who can use those complex software or number of people who can perform those expensive experiments, right? That's a real choke hold. Like, many prototypes can you do? A tape out in semiconductor industry could cost millions of dollars and months.So you want to do these iterations really fast, accurate, so you can make design decisions, decisions about materials, your layout and the same goes for robotics like you want to make choices about actuators, the loads, is it going to be able to do the cooling itself or do you have to design some additional cooling mechanism for the electronics and the edge inference you are putting in there. So that is a different way of saying the faster you can have a digital physical physics infrastructure in your design loop, the better decisions the engineering organizations will make. And that's nothing new, we have needed it. Now that is the place where we are deploying our software. So the users and customers who are using our software, they are getting answers on their physics questions, specifically thermal questions like how is my semiconductor part going to heat up?Or if I make small changes here, will it will I be able to increase the the data rate transfer? Will I be able to increase the power load in matter of minutes, matter of seconds versus waiting a day. Or most importantly, not needing an expert who can perform those analysis versus just you ask your question, get your answer in a manner that you can trust the answers.Jordan Metzner: Okay. And and tell us a little bit kind of how where where it works in the workflow. So, you know, you're developing a chipset. Obviously, well before it's it's gonna get to fab, you're, you know, trying to build, I guess, like a digital twin of this of this chipset, right, inside your software so then you can then run these simulations. Right?I guess that's like the step forward, and then you run these simulations and have kind of a concurrent loop and then kind of go back and reiterate, etcetera.Hardik Kabaria: So the existing workflow pretty much goes like you described. There's somebody who designs the part, then somebody who performs physical performance evaluation, then you check for manufacturability, and then you sort of circulate. Right? And quite often, these are not only not the same people, they might not sit in the same team. In some cases, they might not even work for the same company.So the innovation cycle is never in days, it's often in weeks. Right? But you care about the performance across the board. So even as the designer who is making changes, they want to know if the changes would have effect on the physical performance or manufacture. So we are making this available across the loop.So let's make it more concrete. Of course, today, there is a loop, there is a thermal engineer who works for a semiconductor company. Now instead that engineer, she might be performing 10 to 20 analysis per day, exploring different configurations that came to her from design. With us, she's able to do thousands. So that is your productivity and efficiency gain.That's the story of AI. You make your engineer a super duper duper engineer. But there is a second thing that happens is now you can take that capability and make it available to the designer. And you see that playing out in front of us with large language models driven product for software engineer. Your UX engineer can make some changes.Everybody can build some level of software engineering features. So why can't we make that physics intelligence available to the design engineer? Now the productivity and efficiency again is not going to be from seconds to hours, it's going to be weeks. Concretely, your power engineer can make changes in the power capability of the chipset and start assessing, is this going to be air cooled? If it if this CPU cannot be air cooled, I have to think about other reasons, other ways to achieving this power cycle.Otherwise, the downstream customer who are gonna take my CPU and gonna put it in the board, they aren't gonna be happy because they don't want to deploy, say, robotics, a liquid cool system. So physics is once if you make it available as a digital physics infrastructure across the life cycle, the thing that I didn't foresee when we started the company is that you will make this intelligence accessible to broader set of roles and personas who all work together to deploy a product.Speaker 5: Planning a team off-site sounds fun until you actually have to plan it. Flights, hotels, schedules, coordination, Offsatio handles all of it.Jordan Metzner: Yeah. So the analogy here is just how cloud design has helped, you know, non designing engineers build front ends. This is helping non testing engineers improve the simulation and testing process of the develop.Hardik Kabaria: Yes. And then that means that insight is a lot more usable in action right away. Right? And one is like, analysis came by, oh, the hotspot is higher than I thought. Then you have to think, a different person thinks through what decisions or changes I want to make.Instead, if you make that analysis available to the person making the design, then the iterations cycle is much faster. Not to say the other roles are not important. You're just making the capability available to different level of people in the system.Jordan Metzner: And and let's talk about the downstream impact of that. So, you know, you talk about the scientists who can do, you know, 30 re 30 tests a day, now doing thousands. So that has obviously a downstream impact, and then even more so is bringing that upstream to the, you know, maybe the the board design PCB designer or something like that. And then that obviously causes those iterations to move upstream, which means downstream, those iterations are a derivative of those then iterations. Right?And so okay. But all of this adds up to what? What is the final output? Is it more efficient chips? Is it cheaper chips?Is it better design chips? Is it less, you know, failures in fab? Like, you know, what what is the now output that these laboratories get once they implement this?Hardik Kabaria: I mean, our question to ourself first is, what is it that the customer want? And our customer, in my mind, is, say, a leader of a hardware engineering company or hardware engineering organization. We often think of them as VP of hardware. And this person's motivation is to ship better product faster. That's it.They want to deploy a better chip. They want a time to market, they want to have a faster data transfer on their memory, a better physical and performance time at a faster time to market. And our goal is to become a tool in their toolkit. So we ensure that we get part of their workflow. We deploy behind the firewall.Their IP stays behind the firewall. They don't have to worry about their data leaking from their firewall and we get integrated into how their engineers work with each other. So you can drag and drop your design file, ask a question. It should be fault tolerant system. There are engineers who have performed queries on our software that took seven days to perform.So that means whatever system we deploy has to be fault tolerant. Like, of course, we all know compute goes up and down seven days, so that means we should be able to recover and finish the calculation. And of course, now that you are producing gigabytes of data, humans don't make decisions on gigabytes of data, we make decisions on two by two charts. So then you have to synthesize. Okay, can we make the intelligence simple enough for you to make decisions and explore?So all of that has to be part of the product, not just the raw intelligence. And the reason why we care about it because ultimately we keep asking our questions. Our customers would be happy if we help them deploy better product faster.Jordan Metzner: Okay. Yeah. Speed is the the net end result, which thenHardik Kabaria: Speed is the net end results with a better performance. Because imagine somebody developed a 7% better performing memory, they would have huge gains in the market share. The same can be said about anything that we see around ourselves. A few percent better performing inference chip or a few percent better performing training chip or a data rate transfer all the way to everything that makes up a data center or an edge device.Jordan Metzner: Yeah. And then, obviously, you get that compute downstreamHardik Kabaria: and You get the compute requirement. And then in the wages AI, it becomes so if I'm inferencing faster, maybe I can infer a larger model and bring the intelligence at the edge. Right? So that we are all hungry for intelligence means we are all gonna be hungry for compute.Jordan Metzner: Yeah. Exactly. Okay. Just as we wrap this up, tell us a little bit about, know, maybe the customers you can tell us about who you're working with. Maybe maybe you can't tell us some of your customers, but tell us a little bit about, like, you know, who's using this kind of stuff.Obviously, there's a lot of chip manufacturers from, you know, the big the big famous ones of the Nvidia's and the Intel's all the way down, you know. Obviously, we've we've got a ton of mix of of amazing, you know, CV chips and all different types of of of chip manufacturing in The United States and and across the globe, actually.Hardik Kabaria: Yeah. I mean, so we work with semiconductor companies, electronics companies, equipment companies, the companies who make equipments that go in the foundry. So all the way from memory to fabless to foundry to equipment, a lot of these users for today as our tier one semiconductor hardware companies, so they are larger global enterprises. That is the place where we have started our go to market action. We are starting to engage with even the smaller and medium sized businesses, but so far, our focus has been on the large enterprises.We are attached to several programs, like we are attached to their flagship programs on the memory side and, you know, fabless side. Of course, our customers are quite sensitive about the fact so that we cannot give you names, but that gives you an idea. The other way to think about us is if you have thermal problems, thermal problems can happen all the way the nanometer level features in a memory or a CPU or a GPU to the board level where now you integrate the ASICs or the CPU or the memory onto your board, that could be your consumer electronics to the racks, to now edge devices like you might have control system in your avionics, and now you care about putting more and more inference power so you have to think about how the thermal heat you are going to dissipate. You have thermal problems. We have a product powered by grounds of physics model that can enable you to have these type ofJordan Metzner: Yeah. I mean, think, no one's building chips who doesn't have thermal problems. I mean, heat dissipation is always a problem. And as you're mentioning, right, like, the edge devices, the less powerful devices are getting more compute capabilities, and then, therefore, you're gonna and, obviously, I mean, we have these problems of battery consumption and retention. You know, no one wants an iPhone if it only lasts twenty minutes or something like that.Right? So okay. Just, maybe one last question. I get where we are right now. It seems like we're really early in this life cycle because even implementing your software seems like some efficiencies are gained, but it still feels like the early days of of the potential of this as the models get better, as the chips actually that you run your software on get better, etcetera, etcetera, and then the downstream impact of that.So just paint us a little bit of vision of where's the future, where is this going, where where are people gonna see, you know, chip development in the next few years?Hardik Kabaria: Yeah. I mean, so just I would say we start with something basic. It's like language as a modality is well understood. A model can understand it, provide intelligence with creates a useful information and downstream actions for us. We are going after physics as the modality, something as basic as, okay, I want to predict temperature, I want to predict the heat flux, heat dissipation, the mechanical stress.But we built a ground up model to do that for a couple of physical phenomena, thermal, thermo mechanical, and the model is the same level of capability as GPT-two watts. So it's a half a billion parameter model, We have trained it ground up. So we are the pre training team and we are the product team. Where we go from here is to cover wider range of physical phenomena. So of course we are now focusing on scaling the model, adding more phenomenas in it which is what our industry partners are pulling for.The ultimate goal is make physics intelligence available to anybody who has a question at compute capacity or what we would call it token maxing. Basically you should see your GPU is completely utilized and somebody should be asking a question and they should be getting accurate deterministic answer. We are seeing that in some of our deployment. They have allocated some compute and it's 100% utilization. So there is enough in blonde of queries that's going to happen, right?And that means for us we have to do both the jobs, we want to make the inference faster, so on the given compute allocation we can solve more and more queries And we also have to do the job of expanding the model capacity so you can not only get answers for thermal questions, but electromagnetics questions, RF questions, and so on. So those are the two vectors that we are focused on, but quite a bit in sync with our industry partners. Like, they are the ones guiding our roadmap. Like, hey, you solve thermal problems, can you solve this? Then you do system level cooling problems and so on because at the end of the day, they don't care about just one perfomp.They care about all of it. Why? They wanna ship a better product faster. They don't wanna ship just thermally better product. They wanna ship a holistically better product.So they care about multiple phenomenas that they want to cover in the design process.Jordan Metzner: Yeah. So it it comes down to then, you know, even though you know that certain behavior might cause an increase in thermal, it might be a trade off for something else. And so inevitably, you're faced with an inevitable list of trade offs, right, all the time.Hardik Kabaria: Absolutely. And universe is multiphysics. Every part we create is gonna, you know, enact in a multiphysical world. So, yes, engineers care about it. These disciplines are complex, so there's no workflow today that ties everything.But, ultimately, that is the need.Jordan Metzner: Yeah. That's awesome. Okay. Really great demo. Awesome.Thank you so much. Alright, Sam. Shall we jump into the news?Sam Nadler: Yeah. We're gonna jump into the news. We're gonna we're gonna go a little bit more, consumer. So Meta just launched the a new AI generator, Muse Image. I've had an opportunity to play with it.And, you know, honestly, it feels pretty similar to Nano Banana or OpenAI. What I think is controversial about Muse, and I and I've I've done a little thing here, is if someone's Instagram profile is available, which yours is, Jordan, I said make mister Metz into a rock star. It you know, I think it's pretty good. I'm Jewish.Jordan Metzner: I'm not sure how the good the cross is.Sam Nadler: Yeah. I didn't see that. But, you know, I think and I did ask it as well, and we can get your expert opinion if we can if, creating a diagram of how VINCI AI works to accelerate the physics simulation process for chips. So I have no idea if this yeah. I have no idea if this is accurate.I asked the second image to be more technical so we can dive into that. But, you know, what I want your opinions on is is how controversial should this be that you could basically, you know, immediately, I guess, alter someone's public images. And is that gonna cause problems that, you know, they probably could do via another platform anyway?Hardik Kabaria: Yeah. I mean, I think with any capability, you will have this. Right? And my thought process is is, like, even in the early days of language models, we cared about, oh, it's not giving accurate answer, oh, we could not trust it. And then we developed certain level of trust.Then we developed certain level of guardrails. Then we started using in the production no, you know, human oversight, and now we have developed the trust. What I think this says is this image is a different modality. We are going to go through the same evolution on the raw intelligence, so the model architectures will get created, data pipelines will get created, then the application layer will start getting built, guardrails will come up and there will be all the types of usages we can think about images that's going to happen. So that's maybe my too much of a technical perspective, but I think yes the guardrails will get implemented and we will have all kind of usages, but I think it tells us that the images are going to be very different modality than language.So we will see the learnings from language being applied, but I think the different classes of product and models are gonna show up here.Jordan Metzner: Okay. Cool. I you know, first of all, obviously, coming from Facebook, you know, I think in general, there's a little hesitation on, you know, the meta laboratory. And I think in the case of image generation, where you have, you know, a company who's banking on social network and content generation that, you know, there's a little bit of a questionability just in general, you know, just, I guess, some hesitation in general. But, you know, I think what's interesting is that, you know, since we've seen these models produce high quality images and video I mean, we've seen video been around for a long time.We have not seen a TV show. We have not seen mean, I you know, we've seen people make some trailers. I've seen some TV commercials. You know, I saw the Monopoly Go game the other day said they used AI generation in the commercial. But, you know, we still haven't seen kind of like the, hey, you know, we used AI and made this movie on Netflix and, you know, blah blah blah blah blah.And, you know, we produced this whole thing. And, you know, the question is is why? Why haven't we seen that? And my guess is that we're just still early too early to produce high enough quality content that, you know, producers or directors or creative, you know, developers believe that is enough to keep people exciting. And I know, like, we internally have made a bunch of, like, AI videos and commercials and images and things like that.But, you know, it ends up like, you know, after the fact, it comes up ends up being, like, kinda suboptimal to to reality. Right? And I think we're seeing the same thing in music too. Right? You know, Suno is raising at, you know, high valuations, but how many AI songs are in the top 100?And it's probably zero. Right? So, you know, we we haven't really crossed that schism on to your point, kind of on LLMs we have, we now trust in, like, their belief and all that. But, you know, from a creative side, whether it's images, videos, music, and, you know, what else beyond that, it seems like we still haven't crossed the schism of high quality enough that people are, you know, willing to kind of consume it at a high level. So I I don't know about this one, but it does feel like in general that we're gonna move to a point where whether it's through regulation or not, you know, there's gonna be some type of watermarking or some type of easily detectable mechanism to say that, this image has been fabricated, basically.Yeah.Hardik Kabaria: Absolutely. And I think this we see the physical world very similar because if we are generating physics, then you can say we are generating design. How do we know that this is, say, a design generated by company x y z? Right? Because we care about now it even more.Those designs are gonna show up and actually drive the robot. Some of those robots might be doing very critical tasks at the level of inference and now you you can see how these things are going to get very complex and could have drastic actions taken in the real world. That's what happens when the AI gets into hardware. And that means as the industry walls around different modalities just like image, videos, I think the regulations and guardrails will show up. And that's what will make the real application layer more robust.Right now, I do think we are in the early days of can the raw intelligence exist? I think we are starting to rate the answer is yes, it does. Then it has to become a useful product. That era I think now we are all going to discover for the modality that is not language. We think about physics the same way.There are transient phenomena. So I tell my team that's sort of like video. You know, you are seeing temperature as a movie every few frames per second. But the user who is using it cares about really high fidelity answers. So you have to treat them with a different level of attention, accuracy requirements on the product versus a creative user.Both the users are important, but the product layer, I think, is going to be significantly different and that will drive innovation up and down the stack.Jordan Metzner: Yeah. That's awesome. Yeah. I mean, you could take a picture of an inside of an iPhone, and even if you block all the, you know, labels of the parts, like any LLM or even most hardware engineers will tell you that's an iPhone, and they will know that it's not a Samsung phone or an LG phone or something else. Right?Because, you know, we've seen enough inside of iPhones and how Apple designs their hardware that, you know, you can recognize that. And so I I can imagine that's gonna get all the way down to the PCB level andHardik Kabaria: All the way down to somebody who's designing a memory because that is that's the secret sauce of the companies who are designing. So they will also care about, hey, we have a way to do these things. And it's starting to matter and we want to develop more on it. We want to take advantage of the fact that we have been designing memory for twenty five years. Right?So how will those companies stay ahead? Because they do have something and they wanna take advantage of it. And the startups might say, hey. We have we don't have the institutional knowledge for twenty five years, but we wanna explore the technology stack ground up. So I think we will see innovation on both the sides just like we have seen in software.Jordan Metzner: Yeah. That's awesome. Yeah. I mean, they just look at, like, NVIDIA's chips. You know, this Rubin chip has hardly even been implemented in a single data center yet.I can't even imagine what, you know, power it does when it's implemented at scale to even just NVIDIA themselves and being able to leverage it for future development of other tools. So, know, you it's this vicious cycle of, like, you know, yeah, use your old technology, develop something new. Now that something new has to get used to get to then develop something new again and again and again. So this was an awesome episode. It was really great to hear about your product.And, obviously, again, early days in showing how kind of AI is implemented in in a different part of the space, which is, you know, hardware development. So I thought it was super cool and really awesome demo, and and thanks for joining us.Sam Nadler: Yeah. Hardik, where can people find you? On LinkedIn? On X? Where where can people follow-up?Hardik Kabaria: I mean, of course, our website is easy, winchie40.ai. You can reach out to us, winchie physics on Twitter. But, yeah, we'll be happy to show you guys what we have. And if, technologies of, of assistance to your ambitious hardware program, we would love to help you out.Sam Nadler: Perfect. Thanks, everyone. And we're here every week, and thanks for joining.