NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Hello, and welcome one and all to the Altair HPC Summit 2021. And a warm welcome to all the esteemed panelists to discuss on the topic of AI takes to the cloud. Let me introduce myself. I'm Sam Mahalingam, CTO, Enterprise Solutions, the host for today's panel discussion. I joined Altair in 1994 and contributed to Altair in various capacities right from being a developer to an architect to being a CTO now. Through my journey at Altair, I've helped Altair in shaping the enterprise vision. The technologies that I closely follow are simulation life cycle management, high-performance computing, artificial intelligence and data analytics, Internet of Things and cloud. So today, we have key leaders from Google, Intel, Microsoft, Nvidia and Oracle to share their thoughts on how AI and cloud are the enablers in digital transformation. I would request each one of you to take a minute to introduce yourselves. So Bill, can you go first?

Sure. I'm Bill Magro. I'm with Google. At Google, I'm the Chief Technologist for high-performance computing. I actually recently joined Google about 8 months ago. Prior to that, I was at Intel for a little over 20 years. And at Intel, I served as Intel Fellow and Chief Technologist for HPC as well. My own background is as a computational physicist initially, which is how I got started in HPC in the late 1980s. I'm going to spend my career in HPC. So now I'm at Google, helping drive HPC in the cloud.

Thanks, Bill. Vikram, can you please introduce yourself?

Yes. This is Vikram Saletore. I'm a principal engineer at Intel in the HPC and AI team in the data center group. And I have been with Intel for about 20-plus years. And our team focuses on partnering with ISVs, CSPs and developing solutions for HPC and AI, optimized for Intel hardware and software and libraries. I've led and I still lead collaborations with the enterprise and government, HPC, OEMs and CSPs. I'm also currently a Principal Investigator for CERN and SURFsara engagements that we have with Intel. Prior, I've been with Intel, as I said, 23 years or something, and I got my masters from Berkeley, and then I worked for DEC and AMD, designing CPU and networking products. And then I taught parallel programming and distributed computing for a few years at Oregon State University as a tenure-track faculty. And now I've been with Intel and working on HPC and AI. I also serve as Program Committee Member for supercomputing and ISC conferences for HPC and ML workshops.

Thanks, Vikram, and welcome to the panel, again.

Thank you.

Nidhi, can you please introduce yourself?

Yes. Hi, everyone. My name is Nidhi Chappell. I'm the General Manager for the HPC and AI team within Azure and Microsoft. My team is responsible for defining the purpose-built infrastructure for HPC and for AI that Azure is able to offer for our customers. My team manages all of the H-series products and the platform, the cycle and the batch offerings that we have. Overall, I've been at Microsoft for slightly less than 2 years. And before that, I was also part of Intel, where I spent 12 years between design and then working on HPC and AI.

Thanks, Nidhi, and welcome. Shanker, can you please introduce yourself?

Hi, everybody. I'm Shanker Trivedi. I am at NVIDIA. I'm an SVP at NVIDIA. And I'm responsible for the Enterprise business. And among my responsibilities, I spend most of my time on industry business development. So thank you to Altair and look forward to the discussion.

Thanks, Shanker. Welcome. Karan, can you please introduce yourself.

Yes. Thanks, Sam, and nice to see everybody after so long. My name is Karan Batta. I am VP of Product at Oracle. I joined actually Oracle about 4 years ago when we didn't have a cloud, and so it was nice to be part of a new journey here. I today lead a couple of different teams. AI, HPC is part of them. I lead Data and AI, which is a separate group. I also lead the VMware product suite and then a couple of other cloud native services. Prior to this, I was actually part of Microsoft Azure as part of the HPC team there, spent several years there, building the HPC products. And then I was actually part of Microsoft as part of an acquisition for a couple of start-ups back in New Zealand. My background, mostly in HPC, distributed computing. Fun fact, I used to run the New Zealand Supercomputing Center with a couple of star buddies back in New Zealand. So if you've ever seen Lord of the Rings movies, we had a hand in that.

Thanks, Karan, and welcome again to all of you, and thank you for taking the time to participate in this panel. So what I'm going to do is let me set the context for today's discussion, and then we'll start the panel. So in the recent times, in one of the reports, Gartner predicts that AI will become more and more operationalized by 2024. And as organizations are increasingly embarking on their digital transformation journey, AI, data-driven analytics and the convergence of the cloud are identified as centers of this transformation. So AI is just the technology to revolutionize cloud computing and in making intuitive connected experiences possible. So what we are seeing nowadays is, as companies find more and more ways to deploy AI for real-time predicting and prescribing, the resource requirement for training data-heavy models is yielding to increased demand for high-performance computing infrastructure on the cloud. And then additionally, organizations are exploring the application of AI to HPC using HPC to augment HPC optimization and even automate cloud migration. So what we want to do today is in this panel, we speak with the chip industry leaders and the cloud service providers to get their take on the impact and trends that are seeing from -- that we are seeing from companies who are taking the AI-first approach and how it is driving the move to the cloud. What I would like to do now is get into the discussion with that context being set. And the first question that we want to ask ourselves is, there's a lot of discussion around the blending of AI and cloud and including AI and cloud as being the key technology enablers for the digital transformation. Now what I want to ask, Shanker, you is, can you provide your vision and your organization's vision for AI in the cloud as you and your organization sees it? And what are some of the successes that you have achieved so far as part of the vision that you have drawn?

Yes. Okay. So NVIDIA is an accelerated computing company, a full stack computing company, not just a semiconductor and a systems company. And so our vision is to democratize AI, democratize HPC, make it easy and affordable and available to everybody so that every one of you can do your life's work in your lifetime. And as part of this democratization, we take all of the software that NVIDIA produces and all of the relevant open source frameworks and so on and create what we call the NVIDIA cloud repository, NGC. And NGC is a catalog of containers, of models, of model training scripts, of vertical workloads, of deployment tools, all helm charts, all packaged together in one cloud repository so that you can deploy all of this software really easily in a performance optimized way on every single cloud and every single OEM system. So we take all of the NGC content, and we certify it on Oracle Cloud Infrastructure, on Microsoft Azure, on all of the public -- on Google GCP and all of the public cloud providers. And we also certify the same exact software, the same exact architecture on all of the OEM systems so that they can provide private cloud deployments. All of the software is cloud native. And the successes that we've had -- I'll give you -- by the way, this is a golden age for high-performance computing and data analytics for all the professionals out there. I'll give you just 3 examples of success. We've had a lot of success as many people will know, but let's take computational drug discovery. Both Schrödinger and Recursion Pharmaceuticals are using cloud computing for drug discovery, both in the private cloud and in the public cloud for all of their customers and for internal use. Recursion Pharmaceuticals is doing 1.5 million experiments a week, simply impossible to do in an in vitro environment, which is what you traditionally do with drug discovery. Likewise, for natural language processing, NAVER, which, of course, also owns LINE, the messaging application, have Clova for natural language understanding and natural language processing. So all of the things that you see in a search in Korea or in a live chat are AI-enhanced with natural language processing. And then thirdly is the digital twin, the use of AI and creating and using a digital twin. So BMW is taking what we call Omniverse, a completely virtual environment, where you can create a factory of the future without building the factory, with all the materials, all of the robots, all of the physics in this digital twin of the factory that you're going to build, thereby saving millions, maybe even of the order of $1 billion in terms of creating their new factories. So those are 3 great examples of the use of AI in a public cloud as well as in a private cloud environment.

Thanks, Shanker. I think the successes that you shared in terms of AI being augmented for drug discovery, for natural language processing and even for digital trends is pretty impressive. Now Karan, I would like to ask you the same question. We understand that AI and cloud are the true enablers for digital transformation. What's been Oracle's vision in terms of taking AI to the cloud? And what are some of the successes that you have seen as part of realizing the vision that you have put forth in place?

Yes. Thanks, Sam. No, I think with Oracle, right, Oracle is a very different company than it used to be maybe 5, 10 years ago, right? We, obviously, at its core, are a database company, but we are turning ourselves to be a cloud company. So really, we're in 2 sets of businesses, which is one is applications that makes it like our ERP applications, our SaaS suite, our global business unit applications in different industries like financial services or perhaps like FLEXCUBE as an example, our PLM solutions like PLM Agile. So all these infrastructure applications, and then we're in the infrastructure business like a traditional cloud provider, the core fundamentals of like compute, storage, network, et cetera. So -- and that also includes our database services. So when we talk about AI specifically, what we mean is AI being sort of embedded and seamlessly embedded across all of these different stacks, whether it's providing a bunch of accelerated infrastructure for customers or it's embedding capabilities into our applications. So that's what it really needs. And in some cases, some of these customers are getting these capabilities without even knowing that AI is sort of deeply really integrated into these applications. So I'll give you some examples like analyzing sort of petabytes of data to determine buying behavior for our retail customers of travel industry to really sort of tailor product suite and services, doing lots of demand forecasting and big warehouse systems, et cetera. And then as Shanker said, like, we have a lot of manufacturing and specifically automotive customers that are using AI algorithms to track an item from the manufacturing stage until basically the moment it's sold. So they're monitoring hundreds of assembly lines in different countries across the globe. So we're seeing these kinds of capabilities really ramificate through all of our software services. And as I said, that's the history of sort of Oracle in the sense of all the different business applications in the different industries but then also enabling customers directly to run training, to run inferencing. Those are the end customers that are coming to us for our infrastructure. So it really is across the board. And really, from a success standpoint, I can sit here and kind of give you hundreds of names of customers that are doing these things. But if you look at some of our application customers like Dropbox and AirAsia and FedEx and Mazda and then recently, some of our wins in the HPC space and AI space like Nissan and others. So we have, obviously, hundreds and thousands of customers that are doing these kinds of things on our application side and then also on our infrastructure side.

Yes. Thanks, Karan. I think it is good to know, it's not just the successes on the customer or the application side of things, also how we are embedding AI into your own application so that these applications become more and more autonomous and more and more smarter and intelligent. And that leads me to the second question, which is what are some of the common AI and ML workloads that you folks feel are moving to the cloud? Because when we talk about AI, we talk about all the way from data discovery to training to inference, okay? So are we more and more seeing like is it the AI infrastructure? Is it the infrastructure that the cloud providers are providing for the training of these models? Or are these AI cloud services where these pre-created models are being served in the Software-as-a-Service fashion so that they can be embedded into other applications to make the application smarter? Or is it more and more the sifting of the data because the cloud providers now have all these data lakes? And how do you basically sift through this data in recognizing and in reinforcing patents. So what are some of the use cases that we see? Nidhi, what do you think upon it? What are the use cases that you are seeing in terms of the AI/ML workloads coming into the cloud?

Yes. There's a lot of use cases, again. You mentioned some of them. I would just say, if you look through our customer base, and this is where Microsoft, in general, has anywhere from Bing, our internal first parties like doing research -- search kind of things, to Uber, BBC, KPMG, where they are -- we have lots of customers that are basically trying to see how do I do predictive analytics on my data that I have. So they all look at their use cases and see how they can embed AI, intelligence, personalization like Karan was talking about on customer behavior. All of this comes in many, many different forms, whether it's in maintenance, in other use cases. I would just say -- give an example. We have a customer who's looking at wind simulation, a very classic HPC use case, and augmenting that with reinforced learning because there is a lot of wake effect that happens when you have wind, and it creates basically inefficiency in the system. Your -- so what they're trying to do now is they're basically using reinforcement learning to see how they can further harness more energy from wind. And so these are use cases that are very typical HPC use cases. And that's one end of the spectrum, right? And then Karan talked about the retail, banks, fraud detection kind of things. So it's really every industry across, whether it's manufacturing, autonomous driving, financial services, drug discovery. There's use cases all across. I feel like AI has become a common language that is being spoken across all of these industries.

Thanks, Nidhi. And I think you summed it up quite well saying AI is becoming the common language and more and more software is trying to create software, which we had never seen in the past to sort of make things smarter. But Karan, I want to go back to you, and you spoke about embedding AI into your own applications to make these applications smarter and for it to be more and more autonomous. Can you talk about other use cases? Like, for example, are you seeing high-performance computing clusters being delivered for training these ML models? Are those workloads coming in into the cloud? Or do you also see these large amounts of data, historical data that has already come into the cloud? Do you see a lot of AI models being applied on them in order to come out -- in order to automate data discovery? So are you also seeing such workloads coming into the Oracle cloud?

Yes. I mean I think it's definitely across the board. It depends on customer-to-customer and where their journey is into the cloud or even AI for that matter, right? So I kind of mentioned this before, right? Like, we're using accelerated computing ourselves and even HPC ourselves to embed these kinds of use cases into our services and products, whether it's ERP systems or our back-office applications for supply chain management or whether it's ticketing systems or billing systems in our travel industry GPU. So across the board, we're our own sort of customers from an AI perspective. But also, we have, as I said, customers that are coming in to the cloud, but they're end customers of ours, like, from a cloud perspective. And in our mind, they kind of fall into 2 buckets. They are existing sort of large enterprise customers, and they're looking for getting more inside of their data. They're getting -- they're sort of trying to make their data more intelligent and work through them, right? Because a lot of these customers still -- they still deploy in heaps on-prem, and they're trying to modernize their systems. They're trying to understand their data better. And so for them, really, they need a helping hand. And what that means is we're investing in places like building cognitive services, anomaly detection, speech, vision, text. But then we're also giving them the tool sets like Shanker mentioned, like, where we obviously allow for customers to directly go in, use their own data and then build models on top of NGC and our GPU infrastructure. So that absolutely is happening today across the board. And then we obviously work on areas like research and sort of the wellbeing in the world we are today and sort of the unfortunate situation. We've been adding a lot of AI infrastructure for COVID research, as an example. And then on the other end of the spectrum, like sort of Nidhi mentioned, right, we're also working with lots of retail and social media giants on things like social media analysis. And so really personalization is probably the commonality that I see across a lot of customers, whether it's looking at sort of buying behavior or even financial services that are using this data to build a sort of a financial model of you, right, like to do things like loan discovery or credit rating systems, et cetera. So at the end of the day, it's the same infrastructure, but it's being used in probably a million different ways, depending on what kind of industry you're in. And every segment of those customers are coming in and doing it a different way, whether it's self-managed, whether it's sort of partially managed or going all the way to the SaaS route, where you just want to use the application and get benefit.

So talking about the infrastructure, Karan, it's a good point that you make. So I want to direct this question to Vikram from Intel. So Vikram, how do you see this AI and ML helping the CIOs? Or how do you see the AI and ML optimizing the IT operations? So what are your thoughts on it, Vikram?

So the reason a lot of the IT operations are going to the cloud and using AI is because they are actually seeing that it's getting them efficiency, it's getting them -- they can get insights into how the IT ops go and essentially leverage that from using AI in the cloud. So IT ops is just one. There's definitely -- there are other domains that -- some of the workloads that we have talked about, Karan just talked about and Nidhi also said, is what we're seeing is -- it's not -- it's IT operations, of course, and then there are other domains such as -- in various training workloads and inference workloads. The reason for this -- what we're seeing is this convergence of AI and HPC is because the payoff actually, the benefit that user -- customers are getting are because combining the AI either to sort of replace or displace or augment HPC modeling and simulation, the payoff from the prediction is huge. And so there are multiple domains that -- in addition to IT ops is actually in aerospace industry, where I think Altair has done work with Rolls-Royce, for example. And in high energy physics, we have actually seen our own working with CERN that we can actually get over 100,000x improvement just on predictions. And then also in the health space, where there are large memory requirements and drug discovery, as Nidhi pointed out. And then this augmented simulation is something -- I think the physics-informed AI models is something -- is what's actually happening, and leading to, I think, Shanker talked about and Karan talked about is the digital twin. And I think that's where it's going. And our -- what we are seeing is that one is able to simulate that environment, whether it's smart cities or digital twin of a human body in the health space is something that's actually something very -- that's where people are actually going and looking to the cloud given the infrastructure that's available.

Thanks, Vikram. So moving on, since we are talking about AI is taking to the cloud, Bill, I would like to ask you, what are some of the main differentiators in terms of AI taking a cloud-native approach right from data discovery to training to inference over a hybrid approach where you start with on-prem and then you sort of fully move over for data synthesis or for training to the cloud and then come back and operationalize the model at source or where the data is immediately available. So what are your thoughts in terms of what are the main differentiators here?

That's a great question, Sam. I think everybody knows that one of the big challenges is data movement, right? Data movement is a big issue. And one of the challenges the folks face is that AI really is a high-performance computing workload in the sense that it stresses the system in the same way a lot of simulation modeling workloads do and also it benefits from the innovations that have been driven through the industry for the last decades around high-performance fabrics, around databases -- not databases, data storage and storage systems and so on and even, obviously, vector and dense linear algebra support in processes, whether it be GPUs or CPUs. So on-premise systems tend to be very full. And when they're full, they have a good value. At the same time, the cloud offers flexibility, right? And so the challenge is how do you move things back and forth. And in a hybrid environment, there's a lot of economic advantages of taking the hybrid approach to HPC or even ML and HPC. But you do get into a situation where you have to think about where is, say, the source of truth for that data. And what we're seeing is more and more -- as people start thinking about smart factories, they start thinking about instrumented fleets. Rolls-Royce was mentioned earlier. We've heard about car companies. These connected fleets are generating massive amounts of data. That data is being generated around the world, and it has to be ingested. It's coming in through the networks of our network and other cloud provider's networks. So the data is coming first into the cloud. And that's where the tooling and the specialized hardware exists to really do the best job in terms of training and inference and tying it into a larger workflow. So we're actually seeing that the benefits and the capabilities of the cloud for AI is starting to serve as an attractant to pull the simulation and modeling into the cloud, not necessarily because it's faster or it's cheaper, but because it's close to the data, and it helps close the loop of design. So for example, if you're a manufacturer of industrial equipment or engines or cars, you, in the past, would make assumptions about the operating conditions of those vehicles and feed that into your R&D and potentially make some trade-offs. Now you can actually take real-time data coming in from an instrumented fleet and look at simulations of your product with those environmental conditions. Being able to put that close to the data is really powerful for closing the design loop.

Thanks, Bill. And I want to further expand whatever Bill said, and Shanker I want to ask you. Okay, when we talk about the hybrid cloud approach as more and more of these things are becoming autonomous and AI is being augmented into them for them to truly be autonomous, so which means when you operationalize these AI models, now edge also comes into the play, whether it is the near edge or whether it's a far edge. So would you also now consider that to be a hybrid cloud in terms of not just where the data resides and how you take the initial data, you train your models. And then once you operationalize it, it's going to the far edge and the near edge, now would you redefine hybrid cloud to be cloud, edge, near edge and far edge. So what are your thoughts here in terms of how people want to operationalize AI solutions and AI models?

It's not my job to define things, but you'd start with a cloud native software, which is what we do. And then you want to make it really easy. And of course, public cloud is super easy and it's super easy to get started, right? But many people, as Bill pointed out, have concerns about data. And sometimes, people don't want to have the data. And indeed, in other places, like, imagine if you're in an oil rig or you have a telescope in the middle of a desert, right, where there's no Internet connectivity and the cost of moving the data are very high, local computing is super important, right? In many -- no -- that's kind of the HPC example. In many cases, people are also going to be installing private 5G networks for their campuses, for their factories, for their hospitals. And the data will be local in those cases. And so we see this as a seamless integration, edge -- device to edge to data center, one architecture, one way of doing things. And you choose based on your budget, based on your time scale, based on your security, privacy issues, where best to deploy. And what we've done is we've done 2 -- 3 really cool things, actually. The first is we've built this thing called a DGX SuperPOD, which is an AI supercomputer that's cloud native. It is the world's bestest training and inferencing engine. And with the new A100, you can slice each processor into A7 and then depending on how many processes, you have explosion of computing power, all cloud native AI supercomputer. And then for the small workgroups of engineers and data scientists and simulators and creators, we have DGX A100 Station, which is a workgroup server. And by the way, that workgroup server has 2.5 petaflops of AI computing. One little workgroup. You just plug it into your wall socket, 1,500 watts. And then for the instrument designers of the devices, we put the exact same technology into something which we call AGX, which goes inside the car, inside the medical instrument, inside the robotic surgery -- the robot -- the surgical robot, right? And the exact same architecture with the same cloud native software that you can deploy easily with trained models and what we call TAO, train, adapt and observe -- and optimize. Train, adapt and optimize, NVIDIA TAO, as a workflow for making things easy to deploy in this, let's call it, hybrid cloud seamless way.

Thanks, Shanker. I think it's a good analogy in terms of train, adapt and optimize, and how the DGX Cloud, the DGX Station and the AGX are really having the same architecture deployed on them for the operationalization of AI and for the training of AI as well. So going back, Bill, to you, you touched on this a little bit in terms of AI being augmented into the world of scientific and engineering computing. Can you expand a little bit of -- a little bit more on how AI is impacting scientific and engineering computing?

Sure. Well, I mean, as we talked about the convergence or the confluence of HPC and AI, going back to Nidhi's point, AI is everywhere and it's being used in every industry, but some really interesting and unique things are happening in the world of HPC. First thing is, obviously, the infrastructure affinity, right? The machine learning training models, in particular, look very much like traditional dense linear algebra all-to-all communication patterns. And so the things that we've invested in there make it good for running AI workloads. And that means that there's this coexistence of workloads. The second thing is actually getting to workflows that are integrated. So thinking in terms of an engineer who maybe is doing airplane or automotive design, having to keep track of all these experiments, trying to explore different paths of design optimization, and AI can now come into that workflow and actually do some of the prediction like we're hearing before and actually call out ones that are not going to be promising or identify ones that are promising and help prioritize and drive the focus of that engineer or scientist so she's looking at the right things. And that comes into drug design, it comes into automotive, it comes into aerospace, all kinds of fields. And then finally, how do you actually make that deep discovery and analysis when you really need to get it right, whether it's a car crash simulation or you're predicting the weather or climate forecasting. We're already seeing that the mathematics behind the models that were created really for image identification and other AI things can be repurposed and actually start delivering order of magnitude improvement in simulation. And I think it's going to be a challenge to bring this in, but there are a lot of people working on it. We're already seeing promising results there. So just accelerating the discovery process by essentially revolutionizing simulation and modeling through AI, it's not going to completely displace the traditional mathematical methods, but it certainly can augment. And even if we can't replace that final calculation, we perhaps can get to a point where the prediction is so good that someone very, very quickly is led to a promising solution and the final verification of the car crash or the aerodynamics of the airplane or weather forecast is what you do, you run one time as opposed to having to do thousands upon thousands of simulations to discover a solution. So AI is really going to impact high-performance computing and the way we work end-to-end.

Thanks for those insights, Bill. Nidhi, going back to you. In fact, Karan spoke about this as well. Going with the philosophy of drink your own champagne or whatever you call it, dogfooding, how many Microsoft business units run large-scale AI and HPC workloads as of today? And how does that help drinking your own champagne in shaping Microsoft's AI and HPC strategy and road map?

Yes. No. So within Microsoft, all the teams, whether it's Bing, Office, they're all using at large scale all of their training and inferences happening through Azure. So we do dogfood ourselves, right? The infrastructure that we provide to what we call the third-party customers, we actually make sure it's based on and meets the requirements and learnings from the Bings and the Offices and the Dynamics' internal teams because those are also our first closest customers, if you will, right? And Bing does it at much larger scales than some of our enterprise customers do. So what we end up bringing is a lot of the learning from what Bing is doing and not only just a learning and from a pure infrastructure point of view, right? You can learn from the infrastructure point of view, but they also -- Microsoft internally has developed a lot of models, pretrained models for speech, vision, language that we are using for these components, and we're actually making that available free for trial for our third-party customers because we want customers to be able to plug and play with these models, these cognitive services models to go and say, okay, you have a 0 code, no-code approach to integrating AI, we will make some pretrained models available to you so you can plug and play with this. So back to your answering your question, yes, a lot of our teams -- internal teams at large scale do use our Azure infrastructure. The learnings that we get from them, whether it's for infrastructure, whether it's for the trained models that we have, we pass that learnings to our third-party customers. And there's also another element to this, it's not just internal, but like we work very closely with OpenAI, for example, that has some of the larger language models and is making huge progress there. So we bring a lot of this learning from 1P, our internal teams, and our close partners outside and make it available to everyone.

Thanks, Nidhi. I think that really helps like Karan -- like Oracle, like what Karan said in terms of how they are leveraging AI within their own applications. I think your business units leveraging your cloud for a lot of these AI workloads and then sort of coming out with solutions and exposing them to your customers is probably the only and the right way to do it. So moving on. So Shanker spoke about the DGX for the cloud and the DGX Station and then the AGX. So Vikram, I have a question for you, which is, can you tell us what Intel is doing for your customers and the cloud vendors looking to leverage the convergent solution or a joint solution of AI, ML and HPC? So what are your thoughts and what's Intel doing here?

Yes. Sure. So see, the customers understand that their workloads benefit greatly from HPC infrastructure in the cloud. However, what they're not sure about is how they're going to perform, whether it's going to be cost-effective and how much effort there actually is required for them to move their workloads to the cloud. They're also not -- they have a huge amount of data that's actually in their -- which, as Bill pointed out, moving data is, of course, expensive. So what we have done in the past is we actually are -- we have developed and are continuing to develop working with our partners, what we are calling Intel Select Solutions and working with our OEMs as well as ISVs. These are predefined optimized solutions based on our CPUs, Xeon product line and networking and, of course, storage with software, including our cluster checkers supporting by OpenHPC, OpenVINO, Apache Spark and, of course, the more -- and definitely the deep learning frameworks, TensorFlow and PyTorch. So what we do is we actually minimize the challenges that customers are facing to -- for infrastructure evaluation and deployment. We actually validate these -- the complete stack with -- working with OEMs and ODM partners. We are -- we verify them with -- via benchmarking and also their own workloads. And one of the examples that I would like to give you is the work that we have done in the past -- actually last year is through working with Broad Institute in Boston is that we actually delivered a full genomic analysis toolkit stack or, let's call, genomic stack to accelerate the GATK 4.0, which is open source now. So what we've done is, we have actually provided a full integrated solution to help set up and run GATK workflows on an HPC cloud infrastructure. And the key software component, some of them, if I just were to -- I don't have a picture here, but I can tell you what -- so at the top level is the application with GATK 4 and the Burrows-Wheeler Aligner, which sits on top of Cromwell workflows from Broad. And then it's optimized for GenomicsDB as well as for Genomics Library. Underneath that is the condor, scheduler, the Docker, Spark and Lustre filesystem, and underneath that is the hardware that we provide with Xeon, which is CPUs, Intel SSDs and Omni-Path Architecture and FPGAs. So this complete stack is something that we actually -- is one example of a solution that we provided to Broad. And we are actually working with our OEMs and CSPs to develop -- and ISVs, actually, to develop more solutions in the -- in this coming year, and we hope to do some or more of those.

Thanks. Thanks, Vikram. So thank you all. I think it's about time that we sum it up. So thank you for all your participation. And based on the discussion, what's clear is that AI is pushing the edge to the horizons for sure. It is also becoming a dominant weapon in the arsenal to battle and dominate cloud computing. I feel that's also happening based on everything that you folks answered. And what I also feel is HPC is essential as we are moving more and more to the autonomous things because artificial intelligence, along with automation, is leading the maturity of autonomy and things. When I say things, it's not just devices, it is software as well. And HPC in the cloud, I feel, is an enabler for more and more autonomy and things. So that's the way how I feel based on the discussion that I've had with all of you. And to sum it up, ISVs like Altair are building and providing the cloud native tools, solutions and platforms to truly address the convergence of simulation, HPC and AI. And I would urge some of you to attend the HPC Summit to further understand how PBS Works, our HPC solution, can provide HPC for AI and -- as well as with our DesignAI solution, like, what Bill was talking about, how we can build and operationalize AI for the engineering workloads and into the engineering tools. And finally, with our SmartWorks solution, how our cloud native AI platform is truly low code, no code like what Shanker and Nidhi were talking about is the way -- is going to disrupt the way AI is going to be built and operationalized by leveraging all of the infrastructure that are being provided by all of the cloud providers as well as by the chip manufacturers. So finally, I would say thank you again, and thanks Shanker, Vikram, Nidhi, Karan and Bill for all your insights. Thank you.

Thanks for the opportunity.

Thank you.

Thank you.

Bye-bye.