NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Hello, everyone, and thank you for joining us. We've got a very exciting webinar today titled Accelerating Plastics Recycling using Virtual Reactor Technology. Now if you're joining this webinar, then you understand that advanced plastic recycling is not an easy problem to solve. And it's one that's going to require multiple parties to come to the table and work collaboratively to solve this issue for the world. That aim, we've assembled a team of industry players that include myself from NVIDIA. We've got a professor from the University of Birmingham, we've got CPFD, Encina and Microsoft Azure to come together and walk you through the various different components that each of us players is bringing to the table to solve advanced plastic recycling. My name is Reynaldo Gomez. I'm going to go first. I manage partnerships for the energy vertical at NVIDIA, where I've been for 4.5 years. And I've been working closely with CPFD almost since the beginning. Now I want to start with just 5 minutes giving an overview of who NVIDIA is as a company, because some of you may be wondering what a gaming company has to do with advanced plastics recycling. So I want to clear that out first, that NVIDIA is not just a gaming company. We are not just an AI company. NVIDIA is an accelerated computing platform company. You have everything you need for accelerated computing. Accelerated computing is a branch of computing that cannot be solved with regular computers, regular CPUs. You need an accelerator like a GPU to solve them. The GPU is NVIDIA's hardware answer to that. It stands for the graphical processor unit, and it is one major component of our offering. The second component to our platform is really the software stack. And that's where the secret sauce of NVIDIA comes in. We have more software engineers than we do hardware engineers. And their sole focus is building and optimizing a full software stack. So that way, you can come in either at the CUDA programming language or at the application layer with CPFD and take full advantage of that GPU. Now I can sit here and I could try to explain the difference between how a CPU processes data serially and how a GPU processes data in parallel, but I'll never do as good a job as the MythBusters guys did when they came and presented at our GTC conference a few years ago. I'm going to play this video for you where they're going to visually show you the difference between a CPU and a GPU. There's no audio coming through, but it's okay because I've seen this a million times and can dub over it for you. So this is Leonardo. He's going to paint a picture in the way that a CPU will paint a picture. It's a series of discrete events that happen serially, one after the other. So this is the CPU painting the picture, again, serially. And it's not very fast, it's not very pretty, but eventually it gets a job done. And then this is the GPU, and it is going to paint a picture in parallel. Three, two one. Boom. That is how you want to paint a picture, and that is also how you want to process your data. Now it is a little tongue in cheek, but not by much. Most CPU architectures today, the more powerful ones have anywhere from 28 to 96 cores. So you could think of that as a 28 lane highway to send your data through, our latest H100 GPU has almost 10,000 and to go from a 28 lane highway to a 10,000-lane highway, and that really changes the types of problems that you're able to solve from a hardware perspective. From a software perspective, you look at the different industries, different types of problems that need that parallel processing. And we've largely bucketed them into 3 categories. The first is HPC, high-performance computing. That's where applications like CPFD sit. We also have artificial intelligence. NVIDIA has really been the driving engine behind the recent AI wave since AlexNet in 2012. And then the third is our Omniverse platform for building and hosting virtual worlds. Zuckerberg made the metaverse famous, but Omniverse is the engine underneath applications like metaverse that make it possible, and that really blends some of our expertise from the gaming industry. High-performance computing was actually NVIDIA's first industry outside of gaming. When we move from gaming to HPC to the data center, HPC was where we first got our traction, accelerating large scientific simulation workloads, everything from molecular dynamics to quantum chromodynamics. We've got seismic processing for the oil and gas folks out there. And we also have CPFD, computational particle fluid dynamics. So simulation in the middle, that's really where we started, and that's still where our heart and soul is. We've also made a lot of work blending HPC and AI with this new tool called the physics-informed neural network. Those tools have blossomed into a whole other set of tools for building and creating digital twins. We've seen a huge shift of HPC from only being in the data center out to the edge. If you look at some of these new scientific instruments that are coming online like the Square Kilometer Array, they are collecting terabytes of data per day in these remote facilities that are not anywhere near a data center. And so we've seen the rise of the need to process HPC data nearly in real-time at the edge before you move it to a data center for further processing. And then NVIDIA has also come out with a whole new slew of tools for quantum computing. We started in the quantum computing simulation space, accelerating quantum computing simulation. So that way, scientists can start developing algorithms today while we wait for the quantum computers of tomorrow. And we've also developed a whole new set of exciting tools called [ CUDA ], which is really going to be your bridge between your classical computers and your quantum computers. It's a set of compiler tools that allow you to communicate between a [indiscernible] computer in [ GPU ] and [indiscernible] really exciting new phase for us. I just want to take one more look at NVIDIA software stack and really drive home the point of why NVIDIA is so successful in this space. It's because we take a full stack innovative approach. We innovate down at the chip layer with our GPUs, our A100, our H100s. We packed a lot of science into innovating at the chip level. We also work with partners like our OEMs and our cloud providers like Microsoft Azure to take our GPUs and architect a system that can scale to tens of thousands of GPUs and have them work together, scale out workloads across these thousands of GPUs and really get the maximum performance possible. We innovate at the library and SDK engine with our CUDA tools, our HPC tools, our AI tools. We innovate at the application layer, working with partners like CPFD to take the tools that we've built and construct a full software solution that can solve a business need. And all of this has really led to us developing an ecosystem that is the strongest in the accelerated computing market. We've got over 3 million developers developing on our platform, over 30 million CUDA downloads. We've got over 2,700 GPU accelerated applications. So if you've got a GPU powered data center, there are over 2,700 applications that you can leverage. And we've got over 13,000 AI start-ups and growing. So this at a very high level in 5 minutes is an overview of NVIDIA, our accelerated computing platform and why we are such a central component in high-performance computing, artificial intelligence, and Omniverse. If you have any questions about that platform, please don't hesitate to ask in the chat, and we can get to it at the end. At this point, I want to hand it over to my good friend, Kit, from the University of Birmingham to talk about the research that he and his team have been doing. Kit?

Great. So Rey has introduced to us the solution. I'm going to talk to you a little bit about the problem, and it is a very big problem, the plastic waste crisis. So the first thing I want to do is give you an indication of just how big a problem the plastic waste crisis is. We produce 400 million tonnes of plastic a year. And for me, this is a very -- numbers this size, it's almost impossible to get your head around them. So to put that into context a little bit, every 1.5 years, we are producing a massive plastic equivalent to the entire mass of the human race. To put it differently, again, since the start of this slide, we've produced about 100,000 kilograms of plastic and introduced that into the environment. So this is a massive problem, made more massive still by the fact that 91% of this plastic that we produce is not recycled. And even of the 9% that is recycled hypothetically, most of this goes through a process called mechanical recycling, where it gets heated up, it gets squished down. And the second generation of products are not as good as the first. It's something called downcycling and I'll talk to you a bit more about that in a couple of slides' time. Now the question that I want to pose to you and I talk about the plastic waste crisis, it's a bit of an odd question, but bear with me, is -- whether plastic actually is the enemy that it's portrayed to be because most of you who've read the news recently would have come across the idea of eliminating plastics. So let's get rid of straws, let's get rid of plastic bags. But we have to ask ourselves 2 crucial questions about this. Firstly, is this an achievable goal? Can we really replace plastics everywhere we find them? And secondly, even if it is achievable, is it a sensible goal? Is this actually going to help us fix the environment in the long term? And the reason I asked this is because there's a reason that plastic's everywhere. It's because it's a very, very rare material in that it's simultaneously very lightweight but very strong, which means because not many other materials are like this, it is, in essence, irreplaceable without some kind of a trade-off. So what I'm going to ask is, is the trade-off worth it? And in many cases, the answer is unfortunately, no. Because if we replace, for example, plastic with steel or glass, steel and glass are heavier in plastics, which means it costs more to transport them. You have to pump out more diesel fumes as you take them across the country. As I said before, if they're the same weight, if there's something -- they're as a similar way to paper, paper is nowhere near as strong as plastic. So we end up losing material. We end up with less functional stuff that ends up just being thrown away sooner than the plastic would have been. If we don't wrap our fruit and veg up in plastic, even though some people have a real hatred of that, we are going to waste more food. And if we're building planes and cars purely out of steel instead of out of plastic, then we're going to lose efficiency. To take one of these examples and spin it out a little bit more, let's say, we stop using plastic bottles, and we go instead with glass. And Coca-Cola actually tried to do this once recently. So firstly, when you're creating new glass bottles, the cost of the raw material is higher. The energy usage is much higher. Then when we transport them from the factory to the shop, we use more fuel because glass is heavier than plastic. We also lose more of the product. We lose more of the glass because it's more brittle than plastic. We lose even more when we get it into shops and into our homes. And then when we want to recycle it, again, we've got to use more fuel transporting it to the recycling center, and we've got to use more energy processing it back into a usable product. And the quite shocking thing about this is, in actual fact, if we didn't replace our plastics with glass, if we just kept using plastic, it's actually, in terms of carbon emissions, more environmentally friendly just to use plastics and throw them away than to try and recycle something like glass with all its comparative disadvantages. So what I'm trying to say here is, realistically speaking, removing plastics from our society is not the answer. The answer is actually recycling these plastics. So coming back to that phrase I mentioned earlier, downcycling, I talked about downcycling rather than recycling. And the majority of recycling we do at the moment for plastics is unfortunately downcycling. Because when I say the word recycling, what most people picture in their heads is you take a plastic bottle, a plastic bottle of drinking water, let's say, we recycle it. And we get back another drinking bottle, another plastic bottle that we can use to put water or soda, whatever else we want into it. In reality, this is not what's happening. What's actually happening is we take our plastic bottle. We send it off to our recycling center. And because a plastic drinking bottle has to be sterile, it has to be clean so it doesn't end up giving us some godawful disease, we need to make sure it's sterile. But plastic recycling, when you just -- mechanical recycling, when you just heat up the bottle, squish it down and make something else out of it, that doesn't make a sterile new product. So we can't legally use it to make another drinks bottle. So what we have to do instead is make it into something like a bin bag or a carpet. And once we are done with our carpet, we can't downcycle it any further. So what we end up doing is just throwing it away. So when it comes down to it, downcycling, conventional mechanical recycling isn't really recycling at all. In essence, it's just delaying landfill. It's just kicking the can down the road by a few years. Now of course, I wouldn't be giving this talk today if there wasn't a better way. I wouldn't just come on here to depress everybody. There is another way to do it. And that is something called chemical recycling, where we take our plastics. We break them down at the molecular level, at the chemical level. We turn them back into something like the crude oil from whence they came, and we use this crude oil to produce new virgin quality, sterile plastics then can become bottles or medical instruments, whatever else we need rather than carpets and bin bags. So the final question I'll post to you is how can we practically achieve this. Well, there are a number of ways that we're exploring in society at the moment. My personal favorite or at least the one that I've worked most closely with is being proposed by a company called Recycling Technologies. And the idea, which I won't go into too much detail of because of time, is to use these great, big fluidized bed pyrolysis reactors. You put your plastic into a bed of heated sand. It gets thrown around a bit. The molecules get ripped apart, and then out the top come these short-chain hydrocarbons, which we can then extract and turn into chemical feedstocks and fuels. We can then take these feedstocks and turn them into really high-quality new plastic products, and we can do this with approximately an 80% efficiency. So this sounds fantastic, doesn't it? So have we solved the plastic waste crisis? Not quite yet. Because the concept is proven. I've been to their site. I've seen this fake crude oil. I've seen it be turned into fuels. But though it's a proven concept, it's not yet a commercial reality. And the reason it's not a commercial reality is because it's not yet quite as reliable as it should be. And the reason for that is we don't fully understand what's going on inside these pyrolysis reactors inside these fluidized beds. And if we don't know what's going on, how can we improve them, how can we optimize them, how can we make them more reliable? And of course, the reason we don't know what's going on is because this is what they look like. These huge steel wall systems. So naturally, we can't see through the size of them if we try to put a camera inside of the cameras just going to melt because it's 600 degrees. So what can we do to understand these dynamics? And the answer is use simulations. The -- as Rey has just told us, computation is finally getting to the point now, we can actually simulate millions and millions of particles that we get inside our pyrolysis rector. And if we can simulate our reactor, then we can troubleshoot, which means we can find out problems that we see in our real system, and we can recreate them in our simulated system and fix them. We can also optimize. We can perform whatever kind of tweaks and twiddles and new geometries we want inside our system. We can trial it for a very, very low cost in simulation and then prototype it in real life, and we can even design entirely novel systems. So I've got a dozen examples to reach one of those, but I'll use the one that I personally think is the coolest because if you've got Barracuda and you've got a 3D printer, you can do something called advanced rapid prototyping. So you take some kind of CAD software and you design and you pop. So this part here is a nozzle and air distributor, which pumps gas into the fluidized bed like you should be able to see in a second here. We can model these in Barracuda. We can try dozens or hundreds or thousands of different designs, and we can then 3D print them so that we have a real-life version and use advanced imaging techniques like positron emission particle tracking to see if they work as well in reality as they did in our simulation. And with a software like Barracuda, you really can simulate things that are extraordinarily close to reality as a large amount of my research is currently showing. So this is just one of the many things you can do with the simulation model of an industrial system. And I'll let -- hand over to Peter now. He will tell you a few more things that you can do with it.

Thanks so much, Ken. Hello, everyone. I'm Peter Blaser with CPFD Software. I'm sharing my screen here in just a moment. And let me tell you a little bit more about Barracuda Virtual Reactors. So here at CPFD Software, we're advancing multiphase simulation and technology, specifically for things like these fluidized bed reactors that are central to many advanced recycling systems. So Barracuda Virtual Reactor, that's a product. It models the 3-dimensional transient hydrodynamics. That's the gas and particles moving through the system, the thermal behavior and the multiphase chemical reactions in these industrial units. And really, it's used to accelerate the commercialization and scale-up for this technology. As Kit mentioned, researchers and companies are able to more broadly explore the realm of possibilities during the research phase. It reduces physical testing costs because only the best ones then go to physical testing and prototyping. It speeds the commercialization and minimize that scale-up risk both in going to commercial and then making larger and larger systems to handle this vast quantity of plastic waste that Kit described. And really, virtual reactors used to communicate the technical -- technological merits of each company's unique IP to all parties involved. Virtual reactor, it is physics-based engineering software to model all those particles and gases and chemical reactions. It's fairly complex. So we were an early adopter on parallelization on NVIDIA's hardware and software stack. We've seen single GPU acceleration up to 100 times and showed as soon as we made these slides last night, I just got some benchmarks back on the latest test and saw closer to 200 times. So all this take with a grain of salt. By the time you watch the playback, it will change again. Multi-GPU, we've seen -- that's using multiple GPU cards to stay the same problem, seeing some other benchmarks with 1,000x. And we're continuing to work with partners like NVIDIA, in fact, we're a preferred ISV partner with NVIDIA to continue to get these speed ups and benefits to our customers. Barracuda Virtual Reactor comes with flexible deployment options, whether that's Windows or Linux, whether you're running things like a DGX or OEM-based systems, whether you run on-premise, on the cloud or seamlessly transition between the 2 working directly with us or through some partners. We've been honored with a lot of recent recognition. I won't go through the whole slide, but just note the transition. If you come down in 2019, we're highly commended at the IChemE Global Awards on a project with Viva Energy Australia. This is a refining technology. One year later, we are again commended but for work with a company called TRI, and this was on gasification and municipal solid waste. That's a mouthful. Basically, they've developed a process that is now running in Nevada that takes municipal solid waste, rubbish, trash and makes sustainable aviation fuel. Super cool. We were recognize as the best software technology from hydrocarbon processing. And most recently, this shift towards sustainability has been recognized as finalists multiple sustainability awards in our work with Encina who will be our next speaker. And that's just the tip of the iceberg. There's been a lot of recent announcements in plastic recycling. If you asked me 3 years ago, I would not know what advanced recycling or chemical recycling was. And last year alone, we had presentations from the University of Birmingham, Recycling Technologies, the press release by [ Nelitec ], presentation by Encina. A lot of those resources are available on the website. You're just getting a teaser and taste are here today. We're a member of an industrial partner of the chemical upcycling waste plastics consortium. And this is something that's taking off fast. So we want to get the word out there to the world that's addressing these tough sustainability challenges. And with that, I'll go to the next speaker. There's some next steps here. I will mention these again at the end, but if you would ever like to try a Barracuda Virtual Reactor, some very tangible next things you could do. We have a live demo coming up on our next webinar on February 15. And then you're welcome to try it yourself, talk to us about how to do that at our next web-based training class coming up in early March. With that, I'm going to transition over and introduce Carlo Badiola from Encina Development Group. And Carlo, go ahead and tell the world what Encina is doing using virtual reactor.

Sure. Good morning, good afternoon and good evening, everyone. Let me make sure I'm sharing my screen. I promise it's coming up. Yes. So the -- and for the portion of this discussion, I would certainly make an attempt to make everyone here fluid bed or hydrodynamic experts in 5 minutes. But part of my focus here is certainly introducing Encina and how we work with our partners and collaborators such as CPFD, NVIDIA and Microsoft. And a lot of our focus here is how do we deploy their technology in our business or our mission for manufacturing and drop in circular chemicals using our proprietary approach of integrated single stage and use of a catalytic conversion process. So what Encina does is we focus on end-of-life plastics, as Kit was describing. A lot of our focus here is focusing on the streams of what they call residue. So these are mixed plastic waste, all the different qualities. And part of our focus here is how do we make sure we can produce and support a supply chain that is supporting a chemical product and monomers as kind of our values and how do we make sure we can convey the [ IOCC+ ] aspects as well as making sure that we can meet the high specifications of chemical grade products. Part of our approach here, too, is that we do work across the industries, not only within, let's say, the advanced recycling space but also within the mechanical recycling space of how do we support and really build out the mission. But as much as I would like to go into every aspect of what we do at Encina, our focus here is how do we make sure we can describe to folks here how we deploy our competing technology and how we deploy cloud-based computing as part of our kit and moving forward with our development. So what Encina does in a nutshell is really broken down into 3 parts. We are developing assets where we envision a preprocessing step where we prepare our plastic materials. So we are in the space of sourcing and identifying and understanding how do we move away from the landfill and in the incineration-based streams and move that towards a higher value and making sure that we can reduce the -- those volumes going into those streams. So that's what we call our preprocessing or our front end. And then the focus for this discussion is really on our reactor tech, and I can give you some brief snippets of how we deploy CPFD virtual reactor within our system and then making sure that we can understand where this sits in our overall development and process. And then towards the back end is where we deploy a lot of our more or less conventional technologies when it comes to chemical separation, and that's usually done with our collaborators and partners as well. So if you allow me for a few minutes, I'll go into a little bit of how we -- give you some flavor of how we deploy virtual reactor. And again, the focus here is we want to certainly reflect kind of the underpinning of experimental reality, demonstration reality and how do we make sure we translate that into computational domains. So part of our emphasis and application within virtual reactor is do you get good concurrence when it comes to what you see versus what you simulate. Because I think one thing to kind of reference here is there's a lot of complexity as Kit was suggesting with respect to how we deploy and a lot of expertise that's required to really understand. And what I'm showing you here in the video is an example of a Cold-Flow model, a physical Cold-Flow model that is a large scale with our partner in Chicago. And a lot of our focus here is how do we make sure we get good concurrence and high fidelity when it comes to the experimental demonstration as well as our computational and execution domains. So maybe the high level here, kind of good snippets, you can kind of -- you get a good sense of physical phenomenon. You see a lot of good solid mixing. You see a lot of good publicized development, variations in density and physical state. I'll skip a lot of those details just to make the 3. But why this is important is because once we understand a lot of those aspects, what we do is in a one-for-one translation in the computational domain, so that we can be sure that we get good fidelity and good concurrence with respect to our simulation. So part of our emphasis here and part of our objective is making sure that as we move forward with, let's say, a specific project or with, let's say, with specific design iterations, we want to make sure we're benchmarking ourselves and pinning ourselves to a good basis and then from there moving forward and then layering in additional complexities and the way we understand how we would, say, deploy our reactor. So in here, kind of a replay of this is just another example of our computational translation of what you saw in the previous slides, just to give you a sense. And all the processing details and, obviously, maybe there's more to come, especially if you're interested in a lot of the details. I certainly recommend going to CPFD's website as well as their user seminar that a lot of our folks would attend and present. I think another example as well is now if you go into -- once you get past the underpinning of, let's say, experimental or demonstration and good process development. You then can proceed on to the more engineering and value-add purposes and how we use CPFD Barracudas, making sure we can make advanced or informed decisions using computational needs to make sure that we can evaluate that. So one example here that you see is basically a video of showing how we progress through the different iterations. One concept on the left is showing no IP, meaning no internals or no details that we understand from our physical phenomenon hydrodynamics to one where you show a progression of the different states of development in terms of enhancing our approach or what we perceive to be the importance of the chemical aspect. And then in summary, a lot of our focus here is making sure that we integrate ourselves to technologies like CPFD's virtual reactor. Part of that is really how do we make sure we can translate that well between the 2 domains, computation and physical with high fidelity. Obviously, we want to execute validated and computational models, especially as we deploy technology. And I think from our experience, there is a higher precision, especially when you use hardware from NVIDIA and then software and cloud computing resources from Microsoft just because it does give you several orders of magnitude improvement versus computational means and then predictive results that basically can be accomplished in days versus weeks. So with that, I wanted to introduce Manish from Microsoft, just so he can describe to you a little more as far as what Microsoft can do with respect to deployments and platforms like these.

Thank you. This is Manish Chopra from Microsoft. I'll keep my introduction short. With Microsoft, my role is to help our customers accelerate their journey. On Azure, leveraging HPC and AI, obviously, GPU accelerated technologies, focus on energy vertical, and this topic on sustainability is very relevant to our energy and manufacturing customers. It's a pleasure to participate in this discussion today. So I'll just jump into this kind of provide you next 5, 10 minutes a glimpse of results from Barracuda Virtual Reactor simulation on Azure, which ran on A100s on Azure and some interesting takeaways. So as we look at the slide, just so that we are mentioning this, let me start with mentioning that Azure -- in Azure world, we have GPU optimized VMs that are specialized virtual machines available on single, multiple and fractional to GPUs. So here you see NVs, NCs and ND series available in Azure. These sizes are designed for AI, computer inventive, graphics intensive and visualization workloads. In short, NV, VMs are optimized and designed for VDI and remote visualization. But for today, we will focus on A100, which is the A100 accelerators. In short, a little background to get everyone on the same page on the A100s. These are the 2 Azure offerings on NVIDIA A100 Tensor Core GPU for AI inferencing and training, deep learning and machine learning type of workloads. NC is a mid-end offering comes in 1, 2 and 4 GPUs, and ND is a high-end offering 8 GPUs as which is capable of handling large workloads like what I'm going to talk about in a few minutes. So by and large workloads, we're talking about billions of parameters and maybe trillion. And in a Barracuda world, plastic recycling Encina model which we ran was pretty large. And we used ND V4 GPUs for the purpose of the simulation. So let's jump into the meat of the simulation model and some of the results. So as you see through this, a little background about plastic recycling model simulation. It's a real-life production model we ran on Azure. The detailed configuration, obviously, highly proprietary, but we can share the shape of the model here. To create a model with 96 million articles was not done before, and it's a real-life production model, not a benchmark, prototype or generic model. As per words of Encina, this is largest real model -- real-world virtual reactor simulation ever run, so much larger than simulations that most Barracuda users typically run primarily because of hardware constraint. We do not have hardware constraints on Azure. So we took advantage of that. We took advantage of advanced GPU systems. And we're challenging Barracuda users to bring a larger model so we can achieve same and better results. For the purpose of the simulation, a little bit about the 2 different GPUs we use for the -- running this model. One is V100, which you may hear it called NDv2s, and we used A100s, which you may hear called as NDv4s. The results show that A100 NDv4 outperformed A100 NDv2 by big margin. And I'll talk more about this in a minute. But here are the specs of the 2 GPUs, just some highlights. So specs to highlight V100s, average GPU is powered by 8 cores, each one 32-gigabit of memory inside. And then for A100s, let me spend more time on highlighting NDv4. These A100s are scaling powerhouse. They are known for fast processing with higher memory. Obviously, powered with 8x NVIDIA -- A100, 80 gigabyte, Tensor Core GPUs. You'll see that they are designed for high-end, deep learning training and tightly coupled, scale-up and scale-out HPC workloads. Obviously, each GPU is with own dedicated topology agnostic 200 gigabyte per second NVIDIA InfiniBand connection, and it supports all kinds of standardized analytics loads for GPU acceleration out of the box. Tensor Flow, CAFE, Rapids, et cetera. So those are the 2 GPUs that we focused on. So here is -- here's what I'd like to share the meat of the results when we use the V100 and A100s and compared with the CPU serial as we ran through the plastic recycling, it's in production model. Comparison between V100s and A100s is pretty interesting as well. Every reading as we started increasing the number of GPUs. In blue, you see V100s and A100s is -- all the readings are a comparison to the baseline serial CPU serial. The y axis is how many seconds of simulations per day. So that model is a pretty large model. You will say the x-axis shows the results when we added GPUs for V100 and A100s. The bottom line, 2 key observations. So with 4 GPUs, run comparison, A100 GPUs achieved a simulation of 20.22 wall-clock seconds per day, outperforming V100 GPUs, which achieved 6.61 clock seconds per day. Also, if we divide 20.2 by 0.04 for serial -- CPU serial, we get 506x better simulation speed with A100s. And this is not absolute value. This is relative value compared to the CPU serial and just with 4 CPUs -- 4 GPUs rather. Just another look, which is comparing the simulation speed if you see in the y-axis speed up compared with CPU serial. So as we see the GPU simulation speed [indiscernible] CPU simulation speed, A100s speed up to 505.6x better than CPU cereal. Even within the ND series, we see the speed of factor to be 3.06x in favor of A100s given 4 GPUs each running this model. So this was fascinating. Just another look at cost effectiveness when we start focusing on dollar per simulation second. If you do the math, A100 VM pricing is more expensive, even when it comes to reservation compared to V100s. But if you combine both speed factor and VM pricing, the cost of running 800 per simulation second is less expensive, that is 19 as opposed to 48. So essentially, if you see the bottom table, the key takeaways, we can emphatically say that the fastest hardware, which is A100, is also the best value. When considering price per simulation speed -- price per simulation second, A100 systems is better value. And given its higher performance to price ratio for Barracuda simulations, compared to V100s, customers can run at maximum speed and save money by using A100 systems. So that's one of the key takeaways. So to kind of summarize, we run the largest industry-grade simulation with Barracuda to get some really fabulous results. There's one key takeaway I would take and highlight is with the simulation results and other benchmarks we have, we can challenge that if there's anyone from Barracuda users who is able to create a larger model than proven here, bigger than 96 million particles, we believe A100s can scale up beyond 500x compared to CPU serial. We can make this statement very confidently, and we are challenging anyone to create a larger model, and we are sure A100s can still serve them with the same or better speed. Peter can talk more about it, but I'll say we can take this plastic recycling simulation use case. And we've seen how much it scales up. We can prove the same testimonials for anyone with A100s. We saw with 1, 2, 3, 4 GPU use case that fastest results was actually the lowest cost using A100s, and there's absolutely no barriers to deployment on Azure. We've seen -- Peter just mentioned, we've seen some recent benchmarks from CPFD, and they can share with 8 GPUs on Azure, all networks, whether we saw about 100,000x speed up for large models, and this is fascinating. And we're talking about A100s, and we expect to see even better efficiencies never seen before. So to start with, I'm leaving some links here, but to start you using A100s and CPFD, we have some reference architecture, HPC cookbooks available, so you have the right recipes to achieve the desired results for your models. And this team is here to kind of help you with any follow-ups or our interest from here. I will also say to close, at Microsoft, we are always looking at latest innovation and always bring it to the cloud first, closely working with NVIDIA. So people are always up to date and get them best results available in HPC and AI on Azure with the best cost-effectiveness. And please do reach out to me, if you have any detailed questions, we are able to answer here or otherwise. Thank you for the time. Peter, back to you.

Okay. So thank you, Manish. Thank you to all the speakers. I have been nominated to moderate the Q&A. Look, we have intentionally left time here to take your questions. Thank you to everybody who has sent in questions. I see someone -- 15 of them here already. So we're going to start going through those but please keep your questions coming in and we're going to get through as many of these as we can. So hopefully, all the speakers -- Manish had a chance to catch your breath, and let's just get started with these. This first one -- looks like it will be for you Carlo. How is Encina deploying their technology to solve the plastics problem?

Sure. With respect to how we're deploying right now and specifically the application here. So we use the application of CPFD Barracuda in our engineering as well as when we validate, let's say, our reactor systems. So for example, in our facility in Texas, we use Barracuda to design that reactor as well as troubleshoot and basically do a performance evaluation. And then as you've seen from the Microsoft example, we're also deploying this on a commercial scale for engineer for the larger facility elsewhere.

Okay. Great. Do I need an in-house specialist to run virtual reactor? Why don't I take the first pass at this and then I'll see if some of the users in the room there, Carlo or Kit, if you have anything to add. But virtual reactor is not designed necessarily for a CFD specialist. You don't need a PhD to run it. It is designed with the process engineer in mind. So it speaks the language of bringing in the cat of your geometry and specifying your particle size distributions and data and the information you know, the flow rates, the kinetics you want to use. And then we're available to partner together, either through services or through additional consulting to help people that are new, get started and get success quickly. So I would just add that. But Carlo or Kit, do either of you have anything to add?

Yes. So I think one of the nice things about Barracuda is that when you get the license, you start off with a week's intensive training. So by the time that's finished, you do have a specialist. So the people I put forward for the trading were decidedly not CFD specialists, one couldn't even program and they've picked it up extremely quickly. And on top of that, I think Barracuda have always been extremely responsive to us if we do have any struggles beyond the initial training. So decidedly you don't need a CFD specialist or even a modeling specialist in my experience.

Thanks, Kit. What's next after the A100 GPU? Rey, you might be the one for this one.

Yes, we have already announced the next generation. The A100 stands for Ampere, we name all of our GPU generation after famous scientists. And the next generation after Amperes what we call the Hopper generation, the H100. For HPC applications, it's going to be about 2 to 2.5x faster than the A100. And so if you think back to the -- to Manish's slide where he talked about the price per GPU and then the price performance, we expect the H100 to be even cheaper per simulation second than the A100. And the other exciting thing from the next generation of our GPUs is that NVIDIA is actually coming out with a CPU architecture as well, and we are going to tightly couple that CPU to the GPU to allow even faster communication between the CPU and the GPU. Our CPU is named Grace. And when you combine it with the GPU Hopper, we have the Grace-Hopper architecture, which is an homage to Grace Hopper, one of the early pioneers in computer programming. And I tell you what, I am very excited to see CPFD performance on a Grace Hopper architecture because the entire CPFD application doesn't run on the GPU. You pushed some of the heavier, more parallelizable workloads over to the GPU, but they still have workloads that run on the CPU. Once we've got an even faster CPU and the communication between the CPU and the GPU has increased with increased bandwidth between the 2, I expect CPFD performance to be off the charts.

Thanks, Rey. I can't wait. In fact, we -- I don't know if I should share this, but we've got an early access as through our ISV partner program through one of your [ launch fan sessions ] and tested the H100. Last night, I saw results, and it's significantly faster than they have here which has been significantly faster than the Volta. So this is the beauty of parallelizing on the NVIDIA software -- a hardware and software stack is that as they do their thing, Barracuda Virtual Reactor users immediately benefit. And if you don't have the hardware, you can immediately access it, assuming it's available on our partners like Azure.

And to your point, Peter, you add that -- in the next big H100 series, you would -- I'd call it big ND is forthcoming.

Sweet. Okay. So this question is, obviously for somebody who's run simulation before, simulation takes effort to -- and it sure does as you want to do it and do it right, is the overhead of incorporating simulation justified compared with traditional engineering and physical testing. Kit, Carlo, do you want to feel that one?