NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Hi, everyone. Thanks for joining us today for our webinar on Transforming Transportation with the Metaverse and AI. Before we begin, we wanted to cover a few housekeeping items. At the bottom of your screen, you can find various widgets to use during the webinar. Each widget is resizable and movable. If you have any questions during the webinar, you can submit them through the Q&A widget near the bottom of your screen. We will try to answer these towards the end of the event. Additionally, you'll notice a clipboard icon, which is a brief survey. Please take a moment to fill this out as it will help us tailor these webinars in the future. Here are some tips to help ensure you have a good viewing and listening experience. To maximize the quality of the audio, please close any open applications aside from your browser window. Also, a good old-fashioned browser refresh can cure many issues. So if your audio stops or the slides seem to be lagging, just give your browser a refresh. If you do encounter any other technical issues today, please let us know in the Q&A box, and we'll be happy to help. Now without further ado, we'll turn the event over to our speaker, Norm, to begin the presentation.

Hello, everyone. I'm Norm Marks from NVIDIA, Vice President in our Automotive business, and I'm excited to talk to you today about Transforming Transportation with the Metaverse and AI. Thanks for joining our webinar. We've got quite a bit of material to get through so I'll jump right in. From an agenda standpoint, I'll talk in the beginning about the breadth of things that we're seeing for automotive transformation. The -- we'll then jump in and talk briefly about software defined autonomous vehicles, and then get into the [ meed ] of the presentation around the Metaverse, what we're seeing with Digital Twins and Immersive Experiences, and then talk about how AI can apply in automotive to enable those Digital Twins and more broadly, and round things up with some concluding thoughts. And then we'll take questions live from you at the end of the webinar. So please feel free to submit questions throughout the webinar. And again, we'll take them live at the end. So let's talk about the transformation in automotive. There are a number of areas where software is transforming how we have an impact in automotive today, from AI to Omniverse to HPC. Let me talk about 5 areas, all of which will be touched on at some point in this discussion in how NVIDIA, as an open full stack computing platform, is bringing both hardware and software to bear for the industry. First and foremost is accelerating autonomous vehicle, training, development and testing, where DGX SuperPOD and software like NVIDIA AI Enterprise and Base Commander enabling customers to develop the DNNs for AV and get to market quickly. Simulation is critically important here, starting with synthetic data generation, to fill in rare and dangerous corner cases and even capture data in markets where it's difficult. And we provide both hardware platforms to create the synthetic world in OVX Server as well as hardware in the loop would drive constellation and drives them and modules like IX Sim and DRIVE Replicator are key enablers. I'll be talking quite a bit today about Omniverse is the enabler for the Metaverse or Digital Twins and Immersive Experiences, and again, OVX Servers, what creates those world-class renderings, and I'll say much more about this, but we're increasingly seeing opportunities to AI-enabled Digital Twins with NVIDIA AI enterprise. And broadly, enterprise AI and high-performance computing are being adopted in the industry. And our DGX platform is well built to handle these workloads. And finally, while we'll be focusing mostly on Omniverse here, of course, speeding up CAE and CFD workloads and enabling world-class visualization and remote collaboration in this post COVID-impacted world remains critically important. So let's talk about Software Defined Autonomous Vehicles. The future car is software defined. What a transformation that we're going through. It used to be that when you would buy a vehicle and drive it off the lot, it immediately depreciates. But today, the opposite is true. We'll purchase a vehicle, and over the years of owning that, the AI software will continue to get better. And through over-the-air updates, the vehicle will get better over time. A complete paradigm shift in terms of how we experience a vehicle over time. Exciting, for me, as both the consumer of the vehicle, but also being the company driving the enablement of this. If we think about the key elements of the AV development workflow, it starts with data collection and labeling data. But today, of course, synthetic data generation can also be used instead of relying solely on ground truth data from data collection cars. Mapping plays a critical component in mapping the world to help propel a self-driving vehicle. Training networks, deep neural networks, is critically important here, anywhere from 20 or more DNNs for Level 2 and Level 2+ and ADAS up to 30, 40. We even have [ 8 ] customers publicly pronounced they have 48 deep neural networks to propel their self-driving efforts at Tesla. And then simulation is so important. There's no way you could possibly have enough data collection and AV cars around the world to be able to run every one of the simulations in the real world. So simulation can allow us to test billions of miles in a data center, run through everything in both software in the loop and hardware in the loop testing to ensure every one of the possible global driving scenarios are tested. Now I mentioned some of these deep neural network that are being developed. And here, you see an example of many of them, from obstacle or object detection to distance, time to collision, the opposite of object detection, where do I have free space to drive. We're applying lidar, everything that we need to be able to see about path and lean, planning, reading traffic signs, applying maps, understanding what we're seeing from a high beam, parking, being able to understand how do I approach an intersection, weight perception is something we as humans do quite well. It's difficult for an AI to do. Again, reading traffic light. Gestures and poses can be quite difficult to discern. And that's important both externally where you see someone directing traffic here, but also internally in the vehicle, where we also want to look at things like gaze and eye openness, how do we train models to detect a distracted or drowsy driver. So here, you're seeing roughly 20 or more deep neural networks. But as I said, this is quite complex. In other talks, you've heard me go into great detail about the scale that's required, supercomputer scale for the development of these networks. Today, I won't be going into that, but I will touch more on the simulation side of things and in particular, synthetic data generation so... And here, you can get a sense of these deep neural networks running in real time. So let's talk briefly about NVIDIA Drive Sim because it's a broad software platform that incorporates a variety of different modes in which you, as consumers, can apply it to your AV development and testing efforts. First is synthetic data generation. It's super important to be able to create the world and add synthetic data to the ground truth data that you're collecting from your data collection cars. Of course, you would not want to put norm on a vehicle -- on a bicycle, excuse me, and have me ride directly in front of the vehicle and capture that real-world data when, of course, I'd get run over. So rare and dangerous corner cases are certainly one application of synthetic data generation. But we see customers who might be missing data in many markets or examples, maybe we're missing enough traffic signs in certain markets where synthetic data generation can allow you to create tens of thousands, if not millions of images where you have gaps in your data set, and synthetic data is super critical for that so that we can accelerate the development efforts, but more importantly, improve the efficacy of the AI models for AV. And then, of course, as we develop those models, it's continuous integration, continuous development. Once we feel that the AI models are effective, we want to start testing them in a simulation environment. And that testing can include both software in the loop, where we can run at various speeds from real time to super real time for really complex scenarios, maybe even run it slower than real time, to hardware in the loop, so that we have perfect bid accuracy, that end state driving computer. Instead of being in an actual vehicle on the road, we place it in a fleet in the data center. And now we can have hundreds of vehicles driving in the data center, running all types of simulation through every possible driving scenario, from tests like NHTSA and NCAP, to every possible environmental condition, to introducing transient situations like cut-ins, to even applying testing for faults, where you might have a vehicle introduced that's not on the road and be able to see what would happen. So NVIDIA Drive SIM is the software platform that can enable both that software-in-the-loop testing as well as that hardware-in-the-loop testing. And then finally, we can use DRIVE Sim for a virtual vehicle. Simulating factors, including the in-cabin experience here, in the middle, you see depicted one version of an AR heads-up display. We can test all the various different form factors using DRIVE Sim and DRIVE IX Sim, even early upfront when we're developing the cockpit, before we invest time and more importantly, money into designing a cockpit of the future. We can simulate it and test it in all its different form factors and get a sense for what that would look like prior to sinking money into developing that new cockpit. So here, let me give you a sense and play a video so that you can get a sense of the synthetic data that could be created using the DRIVE replicator module. Look at the difficulty here at night, and we're also seeing RADAR and LiDAR being produced at the same time. We're seeing many emergency vehicles, a difficult lighting condition at night. Could you imagine trying to tell a data collection car to go out and get this seen? We can introduce the idea of creating data for occluded pedestrians, the idea of domain randomization. Here, you're going to see us traversing a parking lot in a variety of different placement of cones and carts, something that would be difficult to do in the real world. So with just one test drive, we can collect a variety of different data sets. And again, we can test every one of the sensors from -- and create synthetic data for every one of the sensors. So from camera to RADAR to LiDAR, DRIVE Replicator can create that scenario, drive it, and then create the synthetic data, including domain randomization so that you could rapidly accelerate the data assets that you have to be able to improve your AI training. So let's talk about a few success stories here at NVIDIA. We're drinking our own champagne and using DRIVE Replicator in the development and creating synthetic data to be able to develop our own deep neural networks. So going left to right, let's talk about the examples. First one is we call it PathNet. It's our DNN that detects path and lane planning. And one of the things that happens in the real world, but certainly is not within protocol to be able to send a data collection car out there, would be vehicles that are actually driving over the lane lines, whether that's the center line or the [ end ] line. And of course, we can't send a data collection car out in the real world to do this, but it's something we need to be prepared for. The second thing that you see visually depicted here are difficult positions of the sun. Here, we're seeing one at sunrise, and that can be so very difficult to be able to pick up. So we used synthetic data generation to both create images, both of these difficult sun conditions, as well as vehicles that would be driving over the lane lines to be able to improve the efficacy of our model here. And what we actually saw was an improvement in detecting the mean average distance to failure of 10%. Quite significant. Secondly, as we move from one version of Hyperion or sensor configuration to the next, we wanted to be able to speed up the development of data. So before we even add the actual sensors in, we had models to be able to simulate data of those sensors, and we're able to cut months out by simulating these sensors and creating data assets of those sensors prior to actually having them. A third example are traffic lights. We had -- it's very difficult to have issues with lights at angles, occluded lights. You might have a tree hanging over. And so here again, we created tens of thousands of images to be able to improve the ability to spot traffic lights and got almost a 3% improvement to the models. And then lastly, traffic signs. In Europe, we wanted to be able to create a larger set of data assets of traffic signs. And so we created nearly 100,000 images across 6 categories and saw an improvement in our ability to spot traffic signs and appropriately read traffic signs in Europe of nearly 2%. Now 2%, 3% and of course, for path, you see a 10% improvement. When we're talking about safe self-driving, every little bit of improvement that we can do to the AI models is critically important for safety. So here, you can see some terrific early examples of synthetic data generation, improving our models, and we'll continue to apply this as we continue to advance the development of these models for the industry. So let's pivot and change gears a bit and talk about the Metaverse and what we, at NVIDIA, are building in Omniverse for Digital Twins and Immersive Experiences. So what is an Omniverse Digital Twin? First and foremost, we use USD to create an interoperable and extensible single source of truth. Through connectors to all of the various applications from design to marketing renderings, to industrial engineering, we create one source to access all of that data live and in real time. The second thing that is so important is that we create a physically accurate replica of whatever it is, whether that's the vehicle, whether it's the function of an assembly line, we'll talk about the factory use case, whether it's human ergonomics, and the physics are accurate. It's so important for a Digital Twin that the physics are actually accurate. Thirdly, and as I talked through here, I'm going to give you a variety of different examples where AI can be an enabler to Digital Twins and be running both pre and post as we test these Digital Twins. And lastly, it's not only important that it's physically accurate, that everything is in perfect synchronization when we're running simulations, when we're looking at an entire factory. So we start with the Digital Twin, now let me talk about the raft of use cases we see for these Digital Twins in automotive. So from left to right, and really here, we're depicting them as you might expect a workflow. The first task, of course, is how do we design and engineer the next vehicle. I can tell you that with the shift from ICE to EV, there's a great deal of work that we're seeing from our customers who as they're designing that new EV platform and many going full EV, they're applying Omniverse to the design and engineering collaboration. It's not uncommon for a very large brand with many makes and models to have hundreds of design reviews a year, and not uncommon for it to take weeks to be ready for these design reviews. What if through Omniverse, by having live connectors into all the design data, we could have each of those individual applications being used in real time, do a real-time design review, make changes and see those propagate in real time and cut really the preparation time really for weeks to same day or dates. This allows for so much better iteration, which, of course, creates better designs. Simulation is super critical. You heard me already talk about AV simulation with DRIVE Sim, where we can test the entire AV software stack, every possible global driving scenarios. If we had 100 of those base driving scenarios, 100 environmental conditions, 100 transient situations, 100 faults, that's 100 million scenarios to test and simulation for autonomous vehicles. And then, of course, the exciting new opportunity for cockpit interior design. Again, all of this happening in DRIVE Sim. Factories of the future build is an absolutely terrific use case. We're seeing great adoption immediately following this. We're going to give you a sense of BMW's factory of the future. And here, we're applying the AI to train robots. Another terrific use case for synthetic data generation for robot training, whether pick and place or the actual robots doing the assembly. Simulating the movement in the factory, including even human ergonomics, so a fully operational Digital Twin. We can have millions of square feet of plants being traversed through a digital factory, digital first, simulate first. Then we're seeing great opportunities to sell both the vehicle as well as sell the functions of autonomy and be able to have an immersive experience in the configurator. Imagine VR experience, and I'm going to show you one example of an online configurator and talk through what the world could look like in a 3D VR/AR before you even buy the vehicle and ultimately also monitor the fleets through a service dashboard using Omniverse, and whether that's fleet management or an individual vehicle in an AR -- XR/AR experience. So to build this out a little bit more, we have a video to follow, but just to introduce the video on BMW, we look at them pulling in all of the data from the physical world to the design of the vehicle, to the plan of the factory, to the humans that will need to work in the factory, to the robots that will traverse and move parts throughout the factory. When we talk about a factory Digital twin, all of this needs to be done in all -- we bring in PhysX to ensure that the physics are accurate, material definition language to ensure that we have the right materials. We apply the AI for how a robot would move, for example. All of this happens in real time. So why don't we let you take a look. For those who like to see it visually, let's walk you through a video of BMW's factory of the future. [Presentation]

Pretty amazing to see the level of detail. Look at the parts here just in looking at the powertrain itself, but simulating everything from the human ergonomics to the robots' movement in the factory. Quite an impressive work that BMW has done. So now let's talk about not the factory side of things, but the consumer side of things. And of course, you want to be able to get the best information about your products into the hands of the consumers in the best way that they can consume it. So whether we're talking about 3D models replacing physical products in terms of being able to present on the web and online and through a variety of other experience, introducing the idea of an AI assistant into an online experience and ultimately tying that all together into configurators and virtual showrooms of the future, which we're really excited. And this is already being done, but we're going to see a terrific expansion of this. So let me show you another video of what Rimac has done in terms of early-stage development of an online configurator for their, of course, high-performance product. [Presentation]

Pretty impressive, isn't it? And imagine there what you saw at the end, having a commercial potentially be filmed in Omniverse since traditional mannerisms that we have today. So let's talk a little bit more about how this can manifest itself with the connected customer and showroom experiences. So imagine today, you can go into a showroom and have a VR experience, and in fact, you can do this today with the Lucid Motors showroom where you can go in and see the various configurations of the vehicle. Today, though, what's missing is you're talking to an assistant and asking them to make choices on an iPad to change the exterior or colors to see what that would look like, change different configurations of the interior. So it's terrific what you can do to experience it today. But in the future, imagine also adding an immersive test drive. Here, you see we're re-spotting the lanes. You can see the animal that we spotted on the side of the road. And imagine you're actually, with the VR/AR experience, driving. And you can say, show it to me driving at night, show it to me driving on a highway, show it to me driving in a city, and experience the autonomous functions before you buy the vehicle. What a terrific way to, a, sell more vehicles; and b, sell more subscriptions. We can test all of the various aspects of the configuration. Today, you're seeing the iPad experience. In the future, imagine just being able to have conversational AI applied. And you could just speak out loud into your VR/AR headset with a microphone, and just speak out loud what you want to see, show me the exteriors black, show me the interiors, this package. In the future, we see great opportunities for AR for things like service. Here, we're seeing we're in the actual showroom. But imagine in the future, maybe having a Facetime-like experience where the dealer has assessed some repair, but now has this experience where they show you an AR, the brake pads that need repair, something where -- today, that would just be a phone call and they're explaining it to you. In the future, you can actually see it for yourself. So terrific opportunities for Omniverse here in the connected customer and showroom experiences. So now let's talk about AI and automotive, and I'll start with where AI can be applied to these Digital Twins. And then I'll go through sort of the art of the possible, of a variety of different ways that we're seeing AI applied in automotive. So starting with these opportunities for AI and HPC across these digital twins that I described, and we'll go from left to right, from design, imagine having the AI models to have the optimal supply chain forecast and demand modeling, be a part of you designing the vehicle in the right configuration so that you can maximize for customer choice and profit at the same time, from an engineering standpoint, optimizing around price as well as performance, building the optimal configurations. One of the greatest opportunities that we see here is Generative AI for design, and I'll say more about that. And then traditional HPC for computer data engineering, CFD and CSM. In simulation, SDG, of course, to improve the AI. You heard me talk a little bit about imagine for configurators down the road, even having recommenders, being able to test for how would that intelligent cockpit experience happen here, and then down the road in a configurator, have a recommender providing dynamic pricing and offers. If I go back in the middle of the build, the SDG for the AI for robotics, I've already talked about, AI for logistics, for parts delivery, optimal prep time, whether it's walk time, assembly time, applying AI for predictive maintenance to be able to get the most out of the machines on the floor. We're seeing great opportunities already being done in the industry, applying AI for visual inspection and motion tracking to both improve the quality of vehicles coming off the line as well as improve the actual factory operations themselves and the time to produce vehicles. And then finally, in service and fleet management, moving from traditional data science and machine learning to the AI for predictive maintenance. Ultimately, if we're out of warranty, having recommenders around loyalty. So a number of things that can be built into these Digital Twins so that the Digital Twins is also AI-enabled. Now why is AI important? Really, look no further than what leaders are doing today that have transformed their industries, what Netflix has done, what Uber is doing in terms of the ride hailing experience, and music. And our expectations of customers, it's changed forever. We expect to have intelligent recommendations provided to us. We expect the AI to understand us and to interact with us in a way that is personalized, that is intelligent. That's the expectation going forward. But the consumer [indiscernible] things is not the only one. McKinsey did a study recently, and they looked at what is the value of AI in automotive across all of the use cases. If we look at while the consumer side and my opening comments around some of these leaders, but we look at manufacturing use cases. Look at nearly $300 billion of opportunity in supply chain management and predictive maintenance, over $150 billion in opportunity in maximizing yield and throughput and energy optimization, nearly $100 billion in opportunity for logistics and warehouse optimization, nearly $50 billion in opportunity in optimizing around product features and product development. And then there is that $300 billion in these intelligent customer interactions, from sales and marketing, to servicing vehicles, to even finance and IT. So a significant amount of value that's available to you. And now let's kind of [indiscernible] these in an art of the possible way through a variety of use cases that I'm seeing our customers adopt and apply AI and move from, as I said, these traditional data science and machine learning to full AI. The first is in the connected vehicle, where we're seeing vehicle management, whether it's a fleet or predictive maintenance for an individual vehicle, being able to alert you that it's time for maintenance early before you're going to have a vehicle stranded on the side of the road; to mobility, whether it's ride hailing or eventually, in the future, as a service offerings, where dynamic pricing and route optimizations -- think of it as inventory. Do I have a vehicle nearby? And so AI absolutely applying here. And then the AI in the cockpit. I would arguably say that we're going to see 2 applications of loyalty for autonomous vehicles. The first loyalty will be, is the AV software networks running well? Do I enjoy my autonomous drive experience? The second is, as we move to these AI cockpits of the future, is, does that brand have AI infused throughout the cockpit? Will I have an intelligent experience, whether I've got my child in the backseat who's gaming, or my spouse is watching a movie. And are they getting an intelligent interaction of what next movie they might want to watch if we're on a long drive and every other interaction in app in the cockpit. And loyalty, we know that for large brands, just 1% of loyalty can equate to $700 million of opportunity. So the implications here are huge. Another very interesting application of AI is around connected service. Here, we're seeing a very interesting service application where you can drive in and it can do that early assessment of the vehicle and what need for repair is being done. But there's the other interactions in the dealership as well, where conversational AI, if you're in a call center, can be providing recommendations, potentially even doing auto scheduling, applying any recommendations and/or loyalty throughout the experience that you have with the dealership. And we know that this opportunity is also in the billions. And if you just capture 1% of the $75 billion of opportunity that's out there, this could mean almost $1 billion to your brand. Now let's talk about the AI and Digital Twins for design and HPC, where there's nearly $50 billion of opportunity in the entire industry, nearly $500 million that could be available to you to apply AI upfront, maximize design and engineering collaboration, an exciting new space around generative design. We already have several customers who are experimenting with this. Imagine if you were to say, show me 2 different brands and how -- if they were to have a vehicle together, what might that vehicle look like; bring in different aspects of designs from different brands or makes and models. And get an early jump start on what that vehicle design could look like. And then iterate from there, cut time out of the design process. And then absolutely significant opportunities over the years for the HPC for CAE and CFD and CSM, as I've already described. In manufacturing, a raft of opportunity here, almost $300 billion of opportunity across the industry. Visual quality inspection is an area where we're seeing increasing application of AI, a number of partners of NVIDIA, by the way, creating terrific software opportunity, and NVIDIA accelerating the use of those kinds of applications and huge opportunities from driving down warranty and quality costs. And then there's, of course, predicting, not just the maintenance of vehicles and connected vehicle, but predicting the right time to do maintenance, to minimize equipment downtime in every one of those machines on the factory floors. And finally, autonomous robots. And that BMW example, they've applied both synthetic data generation as well as train the robots to be able to traverse the plants, and did it before the plant was built because they could build that Digital Twin. So again, terrific opportunities for AI, whether you already have the factory and you want to introduce new robots or you want to apply it all the way up front in a Digital Twin before the factory is even built. Then one final art of the possible opportunity is around the AI for supply chain. With today's number of makes and models and now the mix of ICE and EV vehicles, forecasting demand is more challenging than ever before. We're seeing terrific opportunities to apply AI to improve the forecasting models that we may have traditionally built in the past. Logistics is important. Whether it's starting at the factory level and being able to have the right parts planning in what to order upfront based on the optimal configurations and what's expected to be ordered from these forecasts, to then being able to have logistics applied within the plant, and ultimately to routing and whether that's routing in a plant or routing for mobility. Terrific opportunities here. NVIDIA has introduced a new software capability we call [ CUDA ] that solves for this really difficult logistics optimization challenge. And then lastly, there's so much, again, with all the different configurations that we have, so much opportunity to optimize, whether it's upfront, the price of the vehicle and the options on the vehicle, including what's the optimal price for autonomous subscription to optimizing for parts, to optimizing on spend. So as you can see, there's a variety in a broad set of applications for AI and automotive. And I've only scratched the surface here. What we'd love to do for any of you who are interested is have a deep dive with you to understand your use cases and how we can accelerate the use of AI to drive value to the brand, but also increase the speed at which you can apply AI to your business. So some concluding thoughts to wrap things up, and then we'll open this up to be able to take some Q&A and be able to have a good 10, maybe even 15 minutes of Q&A. So first of all, don't wait. Get started now. When we think about Digital Twins and Immersive Experiences, one of the first places to start is identifying, first, what are the most important use cases for you? Do you start with design and engineering? Do you start with the factory? Are you going to start with the configurator? Any or all of them are going to require design assets, the vehicle assets? But your use case will define which data assets to start with first. So determine what's your path? What will you start with? What will you go to next? And that will allow us to look at collaboration across the use cases, but the right data assets to get started and start building your solution. Ensure that you're injecting physically accurate simulations into those Digital Twins. So whether that's a physically accurate twin of the vehicle that's driving so that we can test every one of the up-to-100-million different types of simulations for an autonomous vehicle, to testing for the operations in the assembly line, to the human ergonomics, ensure that you're applying physically accurate simulations, and get the most out of synthetic data. Once you've identified the Digital Twin, think about where can I infuse AI into that Digital Twin. And independent of Digital Twins, what are the variety of different areas that you could apply AI? I gave you just a few examples from connected vehicle to connected service, connected dealer, for cockpits, then connected customer, AI for design and HPC manufacturing and supply chain. Stack rank and ask yourself, where is my highest value to my business to apply AI? And what's my highest ability to execute? And do a simple 2x2 quadrant where should I get going on my AI journey. And we're happy to help you think through that ability to execute how can we make you successful. When it comes to Omniverse and Digital Twins, develop anywhere and develop once and deploy it anywhere. So you can work on the workstation, you can work in the data center, you can deploy on the cloud. And with NVIDIA's launch planned, everything can be available to you easily, whether that's Omniverse or all the SDKs that are resident within NVIDIA AI Enterprise. So our team in automotive is here to help. I hope I've whet your appetite for AV simulation and synthetic data generation, the Metaverse for Digital Twins and Immersive Experiences in the art of the possible for AI. And with that, we'll open it up to questions.

[Operator Instructions] Let's get started. Norm, the floor is yours.

Well, hey, thanks so much for attending the webinar. We've got a lot of spirited questions and have tried to answer some of them in the online responses already. I hope you've had a chance to peruse those. We've got a few more questions that are here right now, and then I'll monitor others that may come in as I'm talking. There are several questions around the simulation and synthetic data. I'll start with the first one, which is around can synthetic data be used for perception development and validation. And excuse me as I'm fighting a bit of a cough and cold due to the season. So firstly, the primary use that we're seeing today for our own use of NVIDIA as well as our customers, the first synthetic data, is indeed around perception development. And primarily, it's around several things. One is using it for rare and dangerous corner cases, examples like vulnerable road users, infants and toddlers and driveways, garages, parking lots, pedestrians on bicycles coming in front of vehicles, the kind of things that you wouldn't send the data collection car out. But in addition to that, it's quite a number of use cases where we're trying to scale up the volume of scenes that you would have a difficult time collecting either in a geography or just difficult ODDs, emergency vehicles, crashed vehicles on highways. It's hard to get enough scale of seeing crashed vehicles on highways, for example. And believe it or not, in some cases, examples like occluded traffic light and signs, things of that nature. So primary use cases are augmenting, complementing, boot strapping data for developing the AV networks themselves, and complementing it with ground truth data. So that's certainly the primary use. Now can synthetic data also be used for validation? Certainly. We expect this to be used largely starting with those rare and dangerous corner cases, but it can be used for validation as well. But again, I would tell you, the weight of usage is much more heavily on perception development today. Now there's a follow-on question, very related around how do you guarantee that synthetic data generation is similar [ enough ] to real images to hold some kind of explainable AI indoor generalization [indiscernible]. For the explainable AI side of things, what we can and are doing is measuring the efficacy of models that are developed on real-world ground truth data only, then measuring the efficacy of models that are developed using a mix of real-world ground truth data plus synthetic data. And then we can also look at what's the efficacy of a model that was developed on synthetic data only. And we think about traditional data science and machine learning. You can do [indiscernible] and [ challenger ] type bake-offs and pick the model that's most effective. But then to the question, this gives you the answer to explain AI. You can speak to the efficacy of the models based upon the various different approaches that are being taken. There's one more question that's come in around Sim. And the question is, can you elaborate on the similarities and differences between software-in-the-loop versus virtual vehicle simulation? I'm pointing back to an early slide that laid out the different applications of DRIVE Sim [indiscernible]. So let's start with software-in-the-loop. Here, the primary use case for software-in-the-loop is testing the entire AV software stack. Now we're really trying to test whether the AV software stack would detect the kind of things that we would expect, objects, pedestrians, distance, time et cetera. The value of software-in-the-loop over hardware-in-the-loop, although while not a part of the question, I'll come on to that in a second, is that with software-in-the-loop testing, we can run at various different modes. We can run at real time, just like we would in hardware-in-the-loop. We can also run it at super real time, 2x or 3x faster than real time, to be able to test a variety of scenarios in an accelerated fashion. And finally, we can test software-in-the-loop in slower than real time for really complex scenarios. And so this is all about testing the AV stack in a variety of different modes and different speeds. Now I will tell you what we're seeing from an industry perspective is that roughly, if I had to normalize this crossing a variety of different customers, about 80% of the simulations that we're seeing that's being done is done in software-in-the-loop mode. Another 10% to 15% of the simulation is being done in hardware-in-the-loop, where you can run at real time because, of course, this is just an ECU that while it's in a fleet in a data center, it thinks it's on the road. So it runs it real time, and it means it's perfect fit accuracy until we test for enough of the global driving scenarios, feel confident in the result, and then roughly another 5% or so of the testing is done on real on-road testing. So now to the contrast with virtual vehicle simulation. Virtual vehicle simulation does have, a, aspect of it that can be similar to where you're testing the full AV stack. And specifically, this would be an augmented reality heads-up slide, which was pictured on that slide a while back. So you could be running the AV stack, and you could be seeing how it would show up in a central head unit or in a heads-up display where you're seeing the AV stack running. I'm seeing [ bounding ] boxes on objects, pedestrians, vehicles, et cetera, I'm seeing lane lines light up. So that can be done in virtual vehicle and has similarities to software-in-the-loop. The real difference though is that in virtual vehicle, I can be testing in-cabin AI, things like distracted and drowsy driver detection. I can also be testing the other side of this, and that's testing the type of AI that could be running well in the vehicles on the road. The other type of virtual vehicle simulation that we can do is all the way at the upfront design. So we're at the early part of evaluating different design configurations in the head unit. Where might I place the AR heads-up display? How large might it be? Is it in front of the steering wheel? Is it going to be in the center? Is it truly a heads-up display on the dash, on the windshield? So we can be testing these various different configurations. G4 style in the back seat, other access to things like apps and YouTube in the back seat. And before you invest a significant amount of money in the design of the cockpit, you can be evaluating this in a simulated environment and testing different configurations, including in a VR/AR experience. So the overall -- the overlap in fill and virtual vehicles, the AV stack and the AR heads-up display, the differences are, one is all about the AV full stack; the other is about in-cabin AI and about design and cockpit testing and simulation. One final question is if there are other examples of synthetic data being used for training to improve the modeling and the test. So we're seeing a large number. We have a number of active projects with ourselves and customers. I've talked about vulnerable road users, being able to collect large-scale data for things that we don't see commonly, like those emergency and stopped vehicles. Traffic lights and signs, believe it or not, is something that we're seeing large numbers that customers want to be able to ingest these data assets at scale. They simply don't have enough examples of them in their data set. So these are common examples. And again, well, so emergency and stopped vehicles, object detection, weight perception is another one that we're seeing. So intersections, so these are probably the most prominent examples. Now the number of examples can run the gamut, but these would be the most common ones that we're seeing right now from the customers who are using it today. Well, I think that exhausts all the questions that we have. Rick, do you want to close this out?

Yes, sure. Thanks, Norm. Once again, thanks for joining our webinar everyone. We hope you found this webinar informative. And don't forget NVIDIA will be hosting GTC, our Developers Conference, from March 20 to 23 this year. There is some phenomenal speakers lined up for you so you don't want to miss out, and you can register directly from our website. Before you exit, just a gentle reminder to complete the brief survey by clicking the widget at the bottom of your screen as it will help us tailor future webinars. And if you'd like to replay this webinar, an on-demand version of this will be available approximately 1 hour, and can be accessed using the same link. Thank you again for joining us, and have a good day or evening depending on where you are.