0:00 Perfect. Thank you. Thanks very much. So I’d almost say the presentations yesterday were almost too good cuz I got back to my hotel last night and looked at my slides and I was like not quite as pretty as some of the ones that were presented yesterday. Particularly the the Puma and the Hardmon ones were looked awesome. So yeah, great job for those who worked on those. So my name is Ian Pegler.

0:27 I’m part of the computer engineering team at NVIDIA. And specifically I’m in the developer relations group. So my role really is to kind of help people that are building engineering applications, simulation applications use and leverage the the NVIDIA platform. I’m originally from UK, live in Chicago now. And I want to try and give you kind of an overview of like what Nvidia is doing in to accelerate product development and simulation.

0:59 So I’ll probably like spend some time talking about like what our vision is and some of the underlying technologies that hopefully will enable that vision and then I’ll spend show some examples of some of the developers we work with and some of the cool things they’ve done. And it was really great to see quite a few of those developers actually going to present today. You know, you’ve got NTOP, Neural Concept, Physics X, companies like that that we we regularly work with.

So, yeah, let’s get going and hopefully this will be entertaining and educational for you all. So, pretty much every NVIDIA presentation starts with a slide like this. And the reason we start with this slide is really, you know, Nvidia is a platform company. We we don’t make any tools like you can’t get an NVIDIA tool to do anything and you know if we ever do any develop any tools then you’re probably going to end up having a pretty uncomfortable conversation with Jensen at some point about that.

1:57 So the idea is that we develop this platform that people build on and they build their tools on top of that platform. Right? So we’re we’re really aiming our platform for for developers. And what is that platform? Well I think most people know right Nvidia make chips and you can see those along the bottom. So GPUs, CPUs, chips that go in our networking products. I think some people don’t realize we make CPUs.

2:22 So CPUs are are important just as GPUs. There’s certain things that CPUs are very good for. And then obviously we take those CPUs and GPUs and we assemble them into larger scale systems. You see the GB Nvidia naming is quite something. So you see the GB200 NVL72 super pod which is basically our like AI supercomputers. So it’s basically acts as 72 GPU 72 GPUs acting as as one big GPU which is super important for these you know larger and larger more complex AI models that people are developing.

2:56 And it turns out it’s also really useful for doing you know simulation and engineering workflows as well. And then on top of the kind of hardware is the software layer. You know, one interesting thing about Nvidia is we have more software engineers than we do hardware engineers. So they make things like the underlying drivers for the hardware software libraries which I’ll talk a little bit more about.

3:15 Then also like application development frameworks as well which I’ll I’ll also talk about. And this is, you know, the part that hopefully makes it really easy for developers to actually develop on our platform and make really really cool tools. Right. So, this is probably the most I think the most important slide in the presentation in some ways cuz it kind of talks about where I think we’re going with with simulation and where our vision is.

3:44 So, so my background is I think I said I you know I spent most of my career as a CFD engineer before joining Nvidia. So I’m kind of a simulation guy with a simulation background and you know for a long time I think you know we people have talked about this idea of being able to explore a design space right like instead of just running like four or five different designs can we like run thousands and thousand designs and really explore a design space and hopefully find some good ideas and optimum places for us to go with our product design but that’s been somewhat limited right because you know if you’re designing aeroplane right there’s only so many, you know, complex simulations that you can run in a given time.

4:24 You know, on this chart, you know, we’ve said got several per day for for a CFD simulation. Which is great, but you’re never going to explore the design space with that. But the next step up is, okay, let’s take what you’re doing today in simulation and just run it a lot faster. And that’s what we, you know, big part of my job really is helping people move their simulation code from just being CPU to running on GPU.

4:49 To get these kind of orders of magnitude speed up and hopefully we’re now to run these codes a lot faster will also help us generate more data. So if we’re going to do any kind of training of AI models, we can actually leverage this accelerated compute to run these models a lot faster. And then hopefully then we can do thousands of of different combinations per day, you know, and the dream’s always been to do these like design space, you know, optimizations, multiddisciplinary optimizations where we can have all of these disciplines and and kind of really understand where to go with design and hopefully using some of these surrogates will get us there.

5:24 So I’ll come back to this in a little bit more detail. I’ll talk a little bit more detail about the accelerated comput side and the AI physics side a little bit later as well. So where where are people today? And I mean I think that probably some of the people in this room are probably a little further ahead than than I maybe give people credit for in this slide.

5:44 But often I work with a lot of large companies and and the truth is a lot of them use simulation as what I describe as being like a virtual test. So what does that mean? But basically they’re doing virtually what they did in a test environment but just virtually which is good right you’re saving time money compared to doing physical test but it’s almost like you know somebody in the company will come to me and go like I’ll run a simulation on this thing give us the results and they’ll go away with those results just the same way as you would a test but the problem is you’re not really using the simulation to actually find a better design or really you know it’s not the simulation engineer is not going actually hey let’s tweak this and go in this direction they’re just giving the results back And that’s where honestly a lot of companies are today.

6:26 That’s how their their design processes work. So what we really need to do to really make the most out of simulation is move people to actually you know have simulationdriven design like you know use things like a joint which I think some people talked about yesterday or topology optimization to actually let simulation drive the design to a better place. And I’ll eventually get to this place I was talking about the multid-disiplinary design exploration where you bring all of the different fields together to actually hopefully you know get a truly optimized design.

6:59 But obviously there’s a lot of barriers to to getting there right to to move to this thing. Some of them are obviously cultural and some of them are just being able to run these simulations in a cost-effective way to enable you know the massive increase in simulations you need to do right more HPC resource and also to run these things faster that’s a big part of you know where we’re going with accelerated compute right you know if it takes 3 weeks to run a simulation and it does for some of these very complex cases then and your design cycles four weeks then really that simulation is not really going to affect the design too So hopefully that gives you an idea of like where my head’s at about, you know, where we can help, you know, accelerate this move to more simulation based design.

7:39 So what I thought I’d do now is talk a little bit about some of the underlying technologies that Nvidia we offer. Again, these are mainly aimed at developers rather than, you know, end users. So like nothing we develop, you know, you’re going to get a little gooey that you can like click on and stuff like that. It’s really like, you know, you’re going to GitHub to find the code and, you know, use it like that.

So, for a lot of people, you know, our customers are developers that are going to take these libraries and integrate them into their own software. But if you’re a keen hobbyist and you want to have a play around, then you’re more than welcome to obviously go. A lot of this stuff is open source and freely available. Go and have a play around with some of this stuff.

So, I’ll probably kind of go from the top. Warp QX, I probably won’t talk about NIM. I’ll talk about Physics Nemo. I’ll talk about Omniverse. So let’s start with warp. Let’s talk a little bit about what warp is because it’s kind of interesting and some people earlier in the in the sessions were talking a bit about using Python. So I thought it’ be a cool thing to to talk about.

8:43 So fundamentally warp is basically a way to write CUDA kernels but basically in a Python framework that’s very much optimized for simulation and you can kind of see some of the simulations that have been generated by warp. So, you know, sometimes to write kind of these kind of types of codes directly in C++ in in CUDA is is kind of a heavy lift for someone that’s not maybe got a computer science background.

9:06 But Warp is quite nice because you can use Python. A lot of the data structures that you would need are already built in. So, you can access, you know, mesh type functionality. There’s access to like we a lot of it’s backed by VDB, which we’ll talk about a little bit in a minute. So it’s a great way to to kind of if you’re interested in writing sort of simulation tools to to use the warp framework and we actually have had people kind of develop and autodesk research have done some work to actually take it and actually build you know simulations on a large scale.

9:43 So this is a CFD simulation of of New York. And it’s you know very large scale so 26 billion cells. So leveraging again the the kind of power of what we can offer as far as G G G G G G G G G G G G G G G G G G G G GPU compute to do these extremely detailed extremely large scale simulations and actually brings me something I you know probably forgot to mention at the start is one of the challenges pretty much we have sometimes is like with the scale of compute and what we can do today is actually sometimes just changing people’s minds about the scale of simulation they can do so many people are used to doing their like small scale simulations you know a couple of million cells and and really compute you can do today, you can massively increase the fidelity in a number of cases you can do.

10:28 So often it’s kind of changing like people’s expectations of what you can do with simulation. I kind of put this this in yesterday just cuz someone mentioned that they used VDB for some of their projects. So we we actually offer what something called nano VDB which is actually this kind of framework for dealing with like sparse volutric data. So BDB’s background, it comes from the visual effects world.

10:53 And we’ve kind of used it now for actually doing simulation as well because it’s a really nice way to get data like simulation, this sparse data onto a GPU. And the nano VDB is a smaller version that’s optimized to run on the GPU. And as you can see some examples there and I’ve it’s kind of embarrassing actually. You see, so you see Ken Mues, he kind of developed it to the start.

11:19 I didn’t actually realize he was a guy that did it for ages. I worked with him on the project for ages. I was like, “This guy knows a lot about this.” And I I it wasn’t until like a year later that I went on the website and I saw he was like the chair of the whole like thing. Okay, so that’s a little bit about warp. Next thing I want to talk about is CUDA.

11:39 So CUDA is probably one of the key reasons for, you know, Nvidia’s success in the last few years, right? Is it’s this way to take the GPU and actually write general purpose you know programs that leverage the GPU. So that’s what basically CUDA did right was this language for writing programs to run on the GPU. So obviously we still massively support that but one of the key things we offer with CUDA is is a whole bunch of libraries right so that’s why we call it CUDA X because CUDA plus all the libraries that go along with CUDA and again these are like optimized libraries with the idea to make people’s life very easy that we want to develop applications on on our on our platform I won’t go through all of these but there’s some kind of interesting ones QDNN the first one is probably one of the key ones for us Q is a way to accelerate deep neural networks so obviously very important for all of the AI applications.

12:25 Physics Nemo, I’ll talk more detail about that. That’s our framework for de developing AI surrogate models. KOP is an optimization library. So a lot of people are using it like for path planning like you see a little robot going around the warehouse. You know what’s the optimum way to kind of you know like the traveling salesman type problems, right? What’s the optimum way to do a certain number of tasks on the path?

12:53 One other one is coup numeric. So again, someone mentioned yesterday they were using numpy. So coup numeric is kind of interesting because it’s it’s not really a CUDA library as such, but it’s actually a drop in library. So if you’re writing Python and you want to actually have that run on the GPU, you can and you’re using numpy, you can just go in and replace numpy with coup numeric and then all that all of that will then run on the GPU.

13:15 So it’s kind of a a nice drop in way to actually kind of accelerate if if you’re writing Python applications. What else have we got here? We got QDSS which I will talk a little bit later. So there’s actually a whole ton of libraries that are super useful for for simulation. Some of them are here. So I won’t go too much depth to all of these libraries.

13:37 But when you’re writing simulation tools, there’s certain things you do a lot of, right? And often it involves solving systems of matrices for both, you know, CFD, you know, finite element analysis, electromagnetics. I mean fundamentally there’s a lot of matrices that you’ve got to solve either they’re you know you’re going to solve them in a direct way or you’re going to use an iterative approach. So we do have libraries to do both.

14:01 So QDSS direct spar solver Q sparse and AMG X as well. So I’ll talk a little bit about those. So most people maybe have from the CFD world are familiar with am the AMG approach. So this is again solving systems of matrices. So we have a library in Nvidia to do that and do it efficiently on GPU. So one thing I would say about our libraries is the idea is that you know you can use this library and don’t have to worry generation to generation that you’re optimizing it you know for the next generation of GPUs.

14:31 You can just use use a library and know that this is going to run as fast as it can. QDS so this is the library to solve like direct sparse solvers. So this is commonly used for like structural type problems that you want to run on the GPU. And what you can see is the whole range of different problems that we’ve we’ve tested. And you know at the high end the biggest problems we can get like 60x faster.

15:00 And and that was really a whole story about GPU that I’ll talk about later is like GPUs like big problems. Like if you’ve got a tiny little simulation like GPU is probably not going to like help you that much. Really really where it really shines is these big, you know, highfidelity problems because it’s really gives the GPU enough work to really show its benefit. And another cool thing like QDSS does for example, I forgot to mention is like underneath it’s actually got a ton of different methods for actually solving these sparse matrix problems and it’s actually got an AI algorithm that will look at what you give it and pick the best algorithms.

15:36 So it’s not just one approach. It’s actually a range of different approaches that’s that’s used underneath. Next on the list, I’ll talk about physics Nemo more in a bit. I skipped NIM because it’s not really that interesting and we’ll talk about Omniverse now. So Omniverse is kind of this like framework again another framework for developers to kind of build digital twins and and do that within a photo realalistic environment.

16:05 And and quite a few of our kind of developers have already kind of integrated into their platform like anis are going to have it as part of their like next generation fluids tool. So you can see like a photo realistic version of the of the vehicle behind there and I’ll show a little bit more about that in a minute. Neural Concepts who are going to present later used it and integrated it for this SP80 racing yacht that they have been working on.

16:31 Again, I won’t say too much. They may talk about it later, but again, you kind of can generate these really cool looking photo realistic images. Also, Seaman’s integrated into their team center X package because I don’t know how many of people have to deal with PLM on a regular basis, but one of the challenges with some of these legacy PLM tools is when you’ve got huge builds of material like the size of that that ship there, being able to visualize that photo in a photo realalistic way is almost impossible.

16:57 So they’ve actually integrated Omniverse to allow them to kind of visualize like a whole ship. And I’ve got a video I can show later of that in a bit more detail. So so far I’ve I’ve kind of gone through some of the key technologies that like we developed and enable simulation. And what we did actually for the supercomputing conference last year is we actually put a demo together where we tried to actually just show all of them working together to give people an idea of you know what what kind of things you can create.

17:28 So, if you go to build.imia.com, you can actually go online and find this demo. And it’s kind of cool. We basically did a CFD analysis on a whole bunch of different vehicles. And then we trained a physics Nemo AI surrogate model. And then we connected up Omniverse to do this cool visualization. So it’s a nice interactive way. You can kind of play around here and change the vehicle.

17:49 You can interactively move around these streamlines that you see. And it just shows you the power of, you know, leveraging accelerate compute to generate the CFD, leveraging AI to be able to like infer this in real time as you change the vehicles and change the ride height. You can change the wheel types. I’ve got some other cool videos actually that look nice. So, and you can also do it in like a really nice photorealistic way, which looks really, really nice.

18:11 So, you know, if you’re collaborating with designers, you can do it in a a nice, you know, nice pretty way. And and it’s funny, these vehicles are actually supposed to be generic, but anybody knows anything about that looks dangerously like a a Ram truck. So I showed this video to guys from Stalantis like last week and I was like, “This is a generic vehicle.” And they’re like, “No, it’s not.

18:36 It’s a it’s a Ram truck.” And I was like, “Yeah, yeah, look exactly like a Ram truck.” so yeah, this is I mean this this has been a cool way to kind of just show people like what the power is if you use these latest technologies because sometimes simulation and these tools can get kind of stuck in the past, right? So, you know, trying to leverage these latest technologies can can really improve the user experience and how accessible they are.

18:58 And sorry, this is a lot on this slide, but it does actually show all of the different pieces that went into that demo. So you know you’ve got the physics Nemo part that’s connect physics Nemo part that’s connected to in the middle connected to the kit application that’s basically omniverse that the pretty window you see and then you can see the data delegate which actually converting a lot of the the CFD data and things like that into that nano VDB format so we can like work on the GPU in a very efficient way and and visualize things.

19:32 We’re actually using warp to do these like streamlines are actually warp based so that the warp’s actually doing the calculation. So again, that’s where we’re leveraging warp. So again, it just hopefully shows you combining of all these different different things. And like I say, you can go to build.via.com and have a play around with it. It’s also called a blueprint. We have this concept of blueprint. So you can actually download all of the code that underneath that.

19:56 So if you wanted to like use some of this stuff for your own application or use case, you could actually download it and then like build your own application off of it. And that’s always what we want people to do, right? Is take this stuff, integrate into their own application and build something really cool based on our platform. And the other thing I’d say is like obviously we develop all these tools, but really it’s up to the developer to pick what pieces they want to use.

20:18 Like we never insist that you must use every single thing from our portfolio. Like some developers just use one thing. They might be, oh, physics Nemo is really cool. We’re going to use pieces of that or we’re going to use Omniverse but we’re going to develop our own thing because that’s our own IP for doing one part of it. So you know we’re very flexible. We don’t we don’t force anybody to use everything basically right so we’re back to this picture again.

So I basically I wanted to talk a little bit more detail about these two last things because these are the newer areas I guess you know the accelerated comput and the AI physics. So I’ll talk about accelerated compute a little bit and then I’ll talk about AI physics and then then I’ll show some videos and that should hopefully about be about right. So so accelerated comput then so this is basically running a lot of simulations on GPU rather than CPU.

21:06 So, so when I joined Nvidia, I was obviously a CFD engineer and like people saying, “Oh, it can run like loads faster on GPU like you know these 40x you know orders of magnitude faster.” And I guess the question I never really understood is is why like why is it so much faster? Like what what is it about GPU that makes makes it particularly fast? And the analogy that’s often used when we do our like introduction to CUDA training courses is is this comparison between a car and a train with the with the car being the CPU and the train being the GPU.

21:39 And you know, so imagine like right now you want to go back into the city, right? Back over to Manhattan. You could either drive or you could take the train. And if it’s in the middle of the night when there’s nothing on the road, the car is probably going to be the quickest way, right? There’s no traffic. You just drive straight over. It’s it’s the quickest way.

21:57 But as more and more people obviously want to drive into the city, that road is going to get more and more clogged and eventually it’s going to come very very slow. So at that point, it’s actually probably going to be more efficient to move people into the city is actually to use a train. And as long as the train’s running very regularly and you can fill that train with people, it’s a very efficient way to move things.

22:18 And that’s kind of like what the GPU is. It’s it’s this kind of way to do it’s like a throughput machine. It’s got a ton of of of memory. It’s got a ton of cores. So if you’ve got to do a lot of the same thing at the same time, it’s extremely powerful. And you can imagine the architecture has always been built for that. Coming from the days of graphics, right, fundamentally a GPU, you’ve got to paint all of the pixels in the screen at the same time.

So yeah, this analogy actually is is pretty accurate. And sometimes the actual clock speed of a GPU is lower than the CPU. So the CPU, but it can just do so much in that clock cycle. So hopefully that that kind of makes some sense. But it it’s actually pretty good. But one of the hard parts of actually GPU programming sometimes is actually keeping the train full. You’ve got to keep especially with the newer GPUs, keeping enough work on the GPU to keep it busy is actually sometimes the the hard part.

23:12 So let’s just talk a little bit more about the hardware then. So on the left is like a classic kind of more classic PCI GPU, right? The thing that you you’d put in your desktop. And on the on the left side there is the kind of data center GPUs. Which are more commonly used for like the the AI side and and to be honest, the GPU is actually not really the correct really the right name for it, right?

23:41 Because you can’t actually use those for graphics. But the kind of GPU names kind of stuck. So you’ll see different generations of GPU. Hopper, there’s Blackwell and there’s before Hopper there was ampair. So there there’s always like a new generation of these data center GPUs every year. And the key thing is just looking at the memory bandwidth increasing and also the memory. So every generation we put more memory more memory bandwidth and that’s this basically means you can run bigger problems on the GPU faster.

24:07 Generally a lot of simulation is very bound by the memory bandwidth. How quickly can I get from the memory to the GPU? And we get this really high memory bandwidth with these data center GPUs because we’re using this this HBM memory. And if you see there’s that little picture of the GPU here. The key thing with these GPUs is the memory is like mounted right next to the GPU.

24:29 So this is the HBA memory. This is the GPU. And this is what gives us this really really high bandwidth. But it’s kind of a more expensive process, this co process to mount the memory onto the GPU. But that’s really a key part of why we can get this super high performance. And you don’t get it on these desktop GPUs. Generally, it’s more traditional a traditional like DDR type memory.

24:49 But the problem is like obviously for some people they’re like, “Oh, that’s great.” But these things are really expensive. And yes, they are like this this these data center GPUs, you know, these are hundreds of thousands of dollars to get systems. So, you know, sometimes it’s hard to get access to them. But with this RTX Pro 6000 Blackwell, again, a catchy name. I know. It’s it’s really cool cuz we just brought this out.

25:16 It’s got 96 gigs of of memory on it. So, you can run some really big problems on it. It’s got 1.6 TB of memory bandwidth. So, obviously not as big as these ones, but you can actually just in a workstation these day do do some really cool stuff with that GPU. So, that’s a lot more accessible, I think, for a lot of people to run that GPU.

25:30 So, you can do some really cool stuff on it. So, we try to make a range of hardware, right? Like from the really high-end down to like your, you know, developer on their workstation playing around with things. I mean, you know, we strongly believe that a lot of the innovation really comes from people on their desktop, you know, who can run CUDA on their local GPU, that’s really where a lot of innovation starts.

25:50 I mean, even when you look at AI, right, when they start with Alexet and the image recognition competition, that was just like some students playing around with some gaming GPUs where they could run CUDA. So, you know, it’s it’s a key, you know, part of where a lot of innovation starts for us. Just a little bit more about the hardware. I think I talked about the big rack systems earlier on.

26:16 These systems where they’re, you know, there 72 GPUs in a rack for these really large AI models. All Envy linked together. You can kind of see the back of the Envy link. Oh, you did give me a laser, didn’t you? There we go. You can see the back of it which is all envy linked together here and the memory bandwidth on this this connection actually is enough memory bandwidth to actually handle the whole of the internet traffic at once.

26:38 It’s it’s massive memory bandwidth on the back of these systems. Some other things that kind of cool though that we come out recently. I don’t think I put a slide in about it but the DGX Spark I don’t know if you’ve seen that. So it’s a really cool little desktop AI development system that’s kind of like a mini version of one of these big systems. So, you know, it’s got you can run like a 200 billion parameter model on that.

26:57 So, it’s a really nice way, you know, again to increase accessibility to this kind of technology. Okay, I’m going to skip this slide. And again, on a yearly rhythm, we’re always there’s always, much to our engineering team’s probably disbelief, we move to a one-year rhythm of of new GPU generations. So obviously at the moment we’re on Blackwell, then Reuben is coming. I think they’ve just got the first systems in the lab and then Fman’s coming as well.

27:26 So, always something coming. Cool. So, we’re still talking about accelerated computer side. So, I’ve just talked about a little bit of a background to the actual hardware that we use and now I’ll talk about some of the like outcomes of of running some of these tools on on on GPU. So, we did actually a lot of work recently with BMW. So, BMW traditionally ran their external aerodynamics simulations on on CPU.

27:54 So we work pretty close with them to test some of their models and we continue to actually work quite closely with them to run on GPU. So we showed that you know by leveraging GPU we can go from like you know hours to to kind of 20 minutes for some very large cases like the the case you see there is almost half a billion cells. It’s in a big external aerodynamics model of a race car and we can really get that down to like being tractable in like the the minutes time frame from potentially days for some of these cases.

28:19 And it’s kind of nice. They actually posted about it on their social media, which was nice. And that’s I think the picture that was on the on the program this year. So, it’s kind of cool story. And we continue to work closely with those guys to move more of their simulation workload onto onto GPU. We’re actually working with them actually at the moment on thermal management because thermal management is actually kind of a tough problem because you’ve got the external aerodynamics plus you have to model the interior of the engine bay as well.

28:48 So you have like loads of geometry inside the engine bay because you’re managing you’re modeling all of the thermal properties of the things heating up and cooling down. You’ve got to run it for long transient. So some of the simulations on to do this can actually if they do the full fidelity simulation can actually take 7 days to run this on a big on a pretty decent HPC system.

29:08 And like if they can get it down to 2 days they’ll actually do that in production. That will be part of their regular production work. So getting this kind of high fidelity simulation faster does does have a real effect. Similar slide actually from anisfluent. Another big CFD code. Don’t worry about all of these different graphs. They’re not really that important. The key one to think about is actually this one.

29:32 So that this is running this 250 million cell aerodynamics case on 12,000 CPUs here. Has 12,000 CPUs in it. And it’s 17 times faster than running on the 100 512 CPUs. But if we run it on just eight H200s, we’re 34x faster. And you can kind of see that kind of graph starts to level off. So you’ll probably actually never get to this 34x faster than 500 cores.

29:58 And this is why this is so cool because, you know, 34x faster is is significant, right? Like that that that’s like gamechanging for a lot of people. Similar we’ve worked with you know cadence have got a a code that’s works really well on GPU fidelity another use case that’s like kind of cool is in aerospace you know these high left lift configurations when you’re basically coming in to land or you know you’re normally in this high lift configuration where the landing is down the flaps are extended you know to really accurately capture this you really need to run these very high fidelity lees you know capturing the real time transient behavior of the flow to get the right numbers and again to run that is extremely expensive but again if we run it on GPU you can see there we can do it I mean GB 2000 was like 40 50 times faster than than a CPU node and then but then if you often the argument people come back with is oh yeah it’s super faster it’s a lot faster but it’s a lot more expensive and you’re like yeah it is like a a GPU node is a lot more expensive than a CPU node so that’s why the graph down the bottom we kind of normalize by the cost to say If you spent the same money to get that performance, we could do it like seven times cheaper basically with with GPU.

31:11 Similar similar story with this, but I just like this. I don’t know if you can see it, but it’s really cool. You can see like the flow pattern on the on on the fan blades there. And we continue to work on a lot of different use cases to accelerate them. And one of the things you’ll see is actually this is a really important area for Jensen. So often this was from the keynote he gave at DC two days ago and there’s often you know CE and simul engineering simulation examples in there cuz it’s like a it’s an area that really interests him.

31:43 Okay, let’s talk about the AI physics side. So this is the area where there’s definitely a lot of interest at the moment and it’s an area that’s rapidly changing. I think every I think every day I see someone’s come up with a new model architecture to try and use AI approaches to predict things. Some of the examples I’ve shown you predict flow fields predict structural problems, predict electromagnetics.

32:09 And by the way, like even though I’ve shown a lot of CFD examples cuz I’m CFD biased because that’s my background. You know, some of this stuff is also equally applicable to like electromagnetics, structural simulation. Structural is a little bit more fussy though. But yeah it it’s not just not just like CFD. So again similar to what we do generally at Nvidia right we we’ve built this framework called physics Nemo to allow people to kind of take that framework and build AI surrogate on top of it.

32:45 And on the right there you can kind of see the different approaches that you could take. I mean you can just use PyTorch right and just build your own thing from scratch you know at the other end you could just buy someone’s software for it so for most people that’s actually going to be the best approach for them you know everyday CA engineers they’re not going to develop their own software you know they’re going to work with one of our partners like Neural Concept Physics X Beyond Math you know companies like that to actually leverage the the software directly but hopefully for a developer it’s going to be easy to more it’s going to be easier to build on top of PyTorch with physics Nemo than that is going to be to build on PyTorch alone.

33:18 So that’s that’s what hopefully what we’re going for, right? Just to make the developers life easier and physics emote itself, we put together a lot of pieces to hopefully make it easier and part of that is obviously we a lot of the underlying underlying plumbing you need to actually bring in CFD data, get it into the right formats for training, be able to train on multiGPU, multi- node, a lot of that that stuff’s already done.

33:47 So a lot of people can take that and just integrate their own tools. We also actually build model architectures as well and I’ll talk about one of them in a minute. So Nvidia research we often task them with like looking at new things difficult problems. So they help develop new model architectures to like to actually come up with architectures that can efficiently be used for like more of these physics applications.

34:05 Some of the applications we’ve done you can see here there’s a lot of different ones. I probably point you to the top right hand corner. This is really where physics demo started was actually internally we used it for a lot of our own cooling analysis. So that’s actually a the gaming GPUs in a in a box. Also weather prediction. That’s a big area for us. There’s a separate project called Earth 2 where we’re working to actually come up with a a model to predict weather and upscale like weather data as well.

Also, one thing I would say about some of these approaches though is is sometimes people, they kind of see it as an andor approach, right? Like it’s like, oh, you know, this is going to replace, so we go back to this slide, right? You know, this is going to, you know, CFD or AI physics, but it it really really isn’t. They’re going to complement each other. I strongly believe that.

35:10 Like I feel like you know we’ve already seen some companies for example actually Ford did a presentation on this at GTC a couple of years ago where they’re actually using the AI surrogate models to initialize the flow field so they they can actually run their traditional CFD solvers a lot quicker because they’re starting from a really good like first guess you know there’s some often there’s you know subm models that are used in simulation can we use an AI surrogate for that submodel and integrate with a traditional solver and then obviously there’s just using the AI surrogate from scratch to do sort of the design optimization.

35:45 So I feel that it’s not quite a clearcut like we’ll use you know this is going to replace that that won’t be the case in the same way I think about the like testing right like the CFD hasn’t replaced the wind tunnel and and that’s because you still there’s still a lot of value in what the wind tunnel can give you right and you might use it a little bit less for more of the final validation stage but there’s still value in that and I expect a similar pattern to emerge with with these technologies so I’ve got five minutes left So I will which actually is pretty good actually.

36:14 I’m actually relatively on time so that’s good. So I’m going to flip to this slide. So this is a slide of someone that’s actually integrated physics Nemo. So Simscale actually integrated it into a tool to actually do pump performance. So they have a lot of customers that use their platform to to to simulate pumps. So they have a lot of pump data and they they’ve developed a way that you can quickly get a pump curve or at least an estimate of a pump curve by integrating physics nemo under the under the hood to generate these surgo models and this is exactly what we want right you know and in the application under under the hood’s powered by one of our frameworks lumary cloud they actually use physics nemo and they’ve actually done this where they’ve kind of integrated it into I think this is blender and they you know you can within Blender you can you know put a design in and then you can just infer and get like a rough you know idea of what the drag’s going to be.

37:14 So again giving that early quick feedback and then we obviously work quite closely with nTop who a lot of you are already familiar with and they’ve done some cool stuff you know integrating physics demo as well. Sorry I skipped the video there. It’s always tough when it’s YouTube. Anyway, you can kind of see the idea, right? You could kind of change like quickly change the geometry in a robust way which is what they’re really good at and then quickly infer get an answer and then move on to the next design direction.

37:52 Another big thing we’re working on is actually to generate more data. I mean one of the issues with do using these these approaches is you need training data and the training data is expensive, right? So we do work to produce different training sets that people can use to train the models and we’ll continue to kind of support those efforts to and open source these these training sets for people to to go and use.

38:14 Okay. So just that was kind of coming to the end a little bit of what I had to present just to talk a bit a little about exactly what I do in developer relations. You know, I work with I particularly focus on working with a lot of the small and startup companies in this space, which I really enjoy because often a lot of the innovation comes from those companies.

38:40 So dayto-day I do a lot of work with the companies to help you know, get access to some hardware to do testing of their of their the software developing on the latest hardware if they want to benchmark. Giving access to some of our experts internally. A big part of what I do is work with some of our devte people. So if you’re developing something and you want to make sure it’s running the fastest it can on on GPU then we have specialists that can take a look at profiles which is there that’s actually from NSI systems really dig into the details about how your code’s working on GPUs and make suggestions about the best way to optimize it.

39:14 We also work closely with our inception team which is our program for startups. So if anybody’s in a small company, they’re not part of Inception, I’d recommend joining. You get some discounts for for GPUs. Also, it actually gives access to N ventures. So I work closely with our venture capital arm and ventures. So we often fund a lot of companies in this space if if they’re using our platform and we think it’s a good fit.

39:36 And obviously we also like to celebrate success, right? That BMW story was a great one, right? You know, leveraging G G G G G G G G G G G G G G G G G G G G GPUs Seaman’s you know integrated into their software and then BMW are getting some real benefit and real value out of running this these things on GPU so always keen to promote these kind of cases and you know I often help with webinars and things like that this with with flex compute who you heard from from yesterday cool I’m I’m probably exactly at time so just I’ll I’ll end with this slide again so I think the main messages to take away are you know we’re developing this accelerate compute platform to make everything go faster from just your traditional simulation tools also to leveraging AI approaches and really you know hopefully our platform makes it really easy for developers to to make these kind of next generation tools and like I say my role is to to help developers in that process by connecting them with the right people in Nvidia so you know please come and have a chat you know if you’re in this area and happy to help where I And so with that, I’ll I’ll leave it there. Thanks a lot. To see the full recording of this and previous presentations, as well as information about future CDFAM events, visit CDFAM.com.