NVIDIA Changed Generative AI Forever (SIGGRAPH 2023 Supercut)

it all started over 50 years ago with a simple question [Music] what if a computer could make pictures how could we use them [Music] and what would they look like Generations from now hello folks I'm Mr computer image [Music] thank you [Music] [Music] [Applause] [Music] [Music] foreign welcome in I hope you're hungry [Music] [Music] this was the Showcase demo it had two and a half million polygons or so two rays per pixel a couple of bounces per Ray the demonstration was frankly at the time incredibly beautiful that was five years ago now five years later racer RTX 250 million polygons 100 times more geometry 10 Rays per pixel about 10 balances per Ray and using dlss and for like seven out of eight pixels Computing only one out of eight and as a result we're able to render this at 4K scale it up to 4K 30 Hertz hit it not bad for real time thank you while we were Reinventing computer graphics with artificial intelligence we were inventing the GPU all together for artificial intelligence the GPU when I came to see you last time five years ago most people would say that this is what a GPU looks like and in fact this is the GPU that we announced this is the touring GPU but this is what a GPU is today this GPU is altogether one trillion transistors this GPU has 35 000 parts it's manufactured by a robot like an electric car it weighs 70 pounds consumes 6000 Watts and this GPU revolutionized computer science altogether and 12 years later after 12 years working on artificial intelligence something gigantic happened degenerative AI era is upon us the iPhone moment of AI if you will where all of the Technologies of artificial intelligence came together in such a way that it is now possible for us to enjoy AI in so many different applications thousands of papers in just the last several years have been written about this area of large language models and generative AI here are some examples of some amazing things this is the Adobe Firefly Adobe Firefly does owl painting imagine the space around the image that we never captured move AI does mocap from just video this is on the upper right you could you decide which one's real I'm going with the left viscom does sketch to image Guided by language prompt this one's really cool there are a lot of people who know how to sketch and from the sketch and some guidance from your language you could generate something photorealistic and rendered the future of computer Graphics is clearly going to be revolutionized and this is really cool Wonder Dynamics not not only is the name of the company call but they do pose and lighting detection and replace the actor with a CG character they're just it just goes on and on and on the generative AI era has arrived well what's really profound though is that when you take a step back and ask yourself what is the meaning of generative AI why is this such a big deal why is it changing everything well the reason for that is first human is the new programming language we've democratized computer science everybody can be a programmer now because human language natural language is the best programming language and it's the reason why chat GPT has been so popular everybody can program that computer large language model is a new Computing platform because now the programming language is human and what your program the computer understands larger language models and generative AI is the new killer app these three insights has gotten everybody just insanely excited and because for the very first time after 15 years or so a new Computing platform has emerged like the PC like the internet like mobile cloud computing a new Computing platform has merged and this new Computing platform is going to enable all kinds of new applications but very differently than the past this new Computing platform benefits every single Computing platform before it for the very first time this new Computing platform not only enables new applications in this new era but helps every application in the old era this is the reason why the industry is moving so fast but one particular area is extremely important which is the basic scale out of the cloud the basic skill out of the cloud historically was based on off-the-shelf CPUs x86 CPUs while general purpose Computing is a horrible way of doing generative Ai and so we created a brand new processor for the era of generative Ai and this is it this is the Grace Hopper we announced Grace Hopper in fact just only recently several months ago and today we're announcing that we're going to give it a boost we're going to give this processor a boost with the world's fastest memory called hbm3e the world's fast memory fastest memory now connected to Grace Hopper we're calling it gh200 the chips are in production it has four petaflops of Transformer engine processing capability and now it has five terabytes per second of HB and 3E performance so this is the new G h200 based on the architecture Grace Hopper and a processor for this new Computing era the CPU now has 144 cores the GPU has 10 terabytes per second of frame buffer bandwidth you could take just about any large language model you like and put it into this and it will inference like crazy the inference cost of large language models will drop significantly because look how small this computer is and you could scale this out in the world's data centers you can connect this with ethernet you can connect it with the finiband and of course there's all kinds of different ways that you can scale it up this is two gpus but what if we would like to scale this up into a much much larger GPU run it please foreign [Applause] this is actual size by the way this is actual size and it probably even runs crisis the world's largest single GPU one exoflops four petaflops per grasshopper 256 connected by mv-link into one giant system and so this is a modern GPU so next time when you order a GPU on Amazon don't be surprised if this shows up okay so that's how you take race Hopper and scale it up into of course a giant system future Frontier models will be built this way and so let me show you this this is how you would do it and so now you would have a single Grace Hopper in each one of these nodes this is the way Computing was done in the past for the last 60 years ever since the IBM system 360 and for the last 60 years that's the way we've been doing Computing well now general purpose of computing is going to give way to accelerated Computing and AI Computing every single application every single database whenever you interact with an app with a computer you'll likely be first engaging a large language model that large language model will figure out what is your intention what is your desire what are you trying to do given the context and present the information to you in the best possible way well if you were to have an ISO budget way of processing that workload it would take let me just choose the number 100 million dollars and 100 million dollars would be a reasonably small data center these days 100 million dollars will buy you about 8 800 x86 gpus and we can take about five megawatts to operate that and I normalize the performance into 1X using the exact same budget with accelerated Computing Grace Hopper it would consume only three megawatts but your throughput goes up by an order of magnitude basically the Energy Efficiency the cost efficiency of accelerated Computing for generative AI applications is about 20x 20x in Moore's Law and just the current way of scaling CPUs that would be a very very long time and so this is a giant Step Up in efficiency and throughput so this is ISO budget let's take a look at this now again and let's go through ISO workload suppose your intention was to provide a service and that service has so many number of users and so your workload is fairly well understood plus or minus and so with ISO workload this 1X 100 million dollars using general purpose computing and using accelerated Computing Grace Hopper it would only cost eight million dollars eight million dollars and only 260 kilowatts so 20 times less power and 12 times less cost this is the reason why accelerated Computing is going to be the path forward and this is the reason why the world's data centers are very quickly transitioning to accelerated Computing and some people say and you guys might have heard I don't know who said it but the more you buy the the more you save and and that's that's wisdom let's talk about a couple new things and so today I want to talk about Omniverse and generative Ai and how they come together the first thing that we already established is that graphics and artificial intelligence are inseparable that Graphics needs Ai and AI needs Graphics Graphics needs Ai and AI needs graphics and so the thing that we could do is we could create a virtual world that is physically simulated physics simulator that allows an artificial intelligence to learn how to perceive the environment using a vision Transformer Maybe and to use reinforcement learning to understand the consequences of its physical actions and learn how to animate and learn how to articulate to achieve a particular goal and so one mission of a connected artificial intelligence system and a virtual world system that we call Omniverse is so that the future of AI could be physically grounded the number of applications is really quite exciting because as we know the largest Industries in the world are heavy industry and those heavy Industries are physics-based physically based and so first application is so that AI can learn in a virtual world an application the second reason why AI is and computer Graphics are inseparable is that AI will help also to create these Virtual Worlds well let's take a look at what wpp the world's largest Ad Agency and byd the world's largest electric vehicle maker how they're using Omniverse and generative AI in their work play it please wpp is building the next generation of car configurators for automotive giant byds denza luxury brand powered by Omniverse cloud and generative AI open USD and Omniverse Cloud allows denza to connect High Fidelity data from industry-leading CAD tools to create a physically accurate real-time digital twin of its N7 wpp artists can work seamlessly on this model in the same Omniverse Cloud environment with their preferred tools from Autodesk Adobe and side effects to deliver the next era of automotive digitalization and immersive experiences today's configurators require hundreds of thousands of images to be pre-rendered to represent all possible options and variants open USD makes it possible for wpp to create a super digital twin of the card that includes all possible variants in one single asset deployed as a fully interactive 3D configurator on Omniverse Cloud gdn a network that can stream High Fidelity real-time 3D experiences to devices in over 100 regions were used to generate thousands of individual pieces of content that comprise a global marketing campaign the USD model is placed in a 3D environment that can either be scanned from The Real World using lidar and virtual production were created in seconds with generative AI tools from organizations such as Adobe and Shutterstock this Innovative wpp solution for byd brings generative Ai and collab rendered real-time 3D together for the first time powering the next generation of e-commerce kind of love that everything everything was rendered in real time nothing was pre-rendered every single every single scene that you saw was rendered in real time nothing was changed you literally take the cad you drag it into Omniverse you tell an AI synthesize and generate an environment and all of a sudden the car up here is wherever you like it to be so this is a one example of how generative Ai and human designs come together to create these incredible applications and so how do we do this applications generative AI models are making tremendous breakthroughs in fact the hopper GPU is impossible to design by humans we needed AIS and generative models to help us find the way to design this thing in such a high performance way and so it augments our design Engineers it makes it possible for us to create some of these amazing things at all and of course the productivity of the teams go up tremendously well we would like to do this in just about every single industry now we just need some powerful machines we have powerful machines in the cloud of course dgx cloud has many many uh many footprints around the world but wouldn't it be great if you had a powerful machine under your desk and so today we're announcing our latest generation Lovelace GPU the most powerful GPU we've ever put in the workstation is now oh gosh darn it I just put my fingerprints on there did you guys can you guys see that that was what that's not me could you hey can I have this clean in the future my bad sorry everybody uh my bad they work so hard it goes into these amazing workstations and these amazing workstations packs up to four of these gpus it packs up to four Nvidia RTX 6000 is the most powerful gpus ever created and it run real-time Ray tracing for Omniverse as well as train fine tune and inference large language models for generative AI incredibly fast and critically powerful and it produces answers in seconds not minutes uh for in some of the services that are out there okay and so another incredible machine are the servers and these servers as you know getting gpus in the cloud these days is no easy feat and now you can buy it okay you could have your your company buy it for you and put it in the data center and there's a whole bunch of these servers a whole bunch of different configurations I don't know if you guys could see this this is a server that has up to eight of the l40s Ada Lovelace gpus and of course these are not going to be used for Frontier models these are really used for mainstream models today that you can download from hugging face or Nvidia could work with your company to create you could use in just about all kinds of applications around your company and you can fine tune it with these gpus the fine tuning of a gpt3 model okay so this is GPT 340 billion parameters takes about seven hours for about a billion tokens and so 15 hours in a workstation with four gpus of course takes less with agpus and just in fine tuning this is one and a half times faster than our last generation a100 and so l40s is a really terrific GPU for Enterprise scale fine-tuning of mainstream large language models these amazing new Enterprise systems that are in production today all right let's change gears and talk about what's going on at siggraph this year I'm pretty sure all of you have already heard about open USD USD openusd is a very big deal open USD is a framework a universal interchange for creating 3D worlds for describing for compositing for simulating for collaborating on 3D projects open USD is going to bring together the world onto one standard 3D interchange and has the opportunity to do for the world and for computing what HTML did for the 2D web finally an industry standard powerful and extensible 3D interchange that brings the whole world together a really big deal now let's take a look at why it is such a Visionary thing that Pixar did I forget exactly when they invented it but they open sourced it in 2015 and they've been of course using this framework for over a decade building amazing 3D content well the 3D pipeline is incredibly complicated you got designers and artists and Engineers they all specialize in some part of the 3D workflow it could be modeling and texturing materials physics simulation animation set design scene composition there are so many parts and so many different tools and because the tools are created by different companies and largely incompatible import and exporting data conversion is just part of the workflow and because they're incompatible and because there's all this import and exporting fundamentally the workflow has to be serialized it's impossible to paralyze that and this is one of the reasons why creating these incredible 3D animation movies are so expensive and take so much time could you imagine if every single tool was natively compatible with USD then as a result everybody can work in parallel The Interchange and conversion goes away and instead of a serialized model you have a paralyzed spoken Hub model and so this way of doing work of course is incredibly appealing and it's one of the reasons why the vision of open USD has taken off it's being adopted in film architecture engineering and construction manufacturing and so many different fields of Robotics well five years ago we started working with Pixar and we adopted USD as the foundation of Omniverse it's not a tool in itself it's a connector of tools okay so Omniverse is a connector well let's take a look at held the vision of open USD came came together and this is just a fantastic illustration let's starting from the left here I think this is a Adobe Stager Houdini this is a modeling system Maya or animation system modeling system this is Omniverse blender render man Pixar's Minuteman and Unreal Engine from epic a game engine literally all open USD one data set ingested into everybody's tools and it looks basically the same everybody's rendering system is a little different and so the quality of the rendering is is a little different from tool to Tool but one data set available and usable by every tool this is the vision of open USD so incredibly powerful now Omniverse as I mentioned before is not a tool it's a platform for tools it's not a tool it's a platform for tools okay and so we would like to put Omniverse in as many places as possible there's a whole bunch of different applications and we showed you earlier how we used AI workbench to train this model to fine tune this model we started with llama2 and we taught it we taught it we fine-tuned it for USD and so let's take a look at the video for USD developers building profiling and optimizing large 3D scenes can be a very complex process chat USD is an llm that's fine-tuned with USD functions and python USD code Snippets using Nvidia AI workbench and the Nemo framework this generative AI copilot is easily accessed as an Omniverse Cloud API simplifying your USD development tasks directly in Omniverse use chat USD for general knowledge like to understand the geometry properties of your USD schema or complete previously tedious repetitive tasks like generating code to find and replace materials on specific objects or to instantly expose all variants of a USD print chot USD can also help you build complex scenes such as scaling a scene and organizing it in a certain way in your USD stage build bigger more complex Virtual Worlds faster than ever with chat USB generated AI for USD workflows chat USD now everybody can speak USD and chat USB USD could be a USD teacher it could be a USD copilot and help you create your virtual world okay enhance your productivity incredibly we're super excited about the work that we're doing here we're so happy that we chose openusd as the foundation of Omniverse and all of the work that we've done to extend it into real time into physics-based applications and this is the beginning of a journey that we will finally be able to digitalize to bring software-driven artificial intelligence powered workflows into the world's heavy Industries the 50 trillion dollars worth of industries that are wasting enormous amounts of energy and money and time all the time because it was simply based built on technology that wasn't available at the time and so Omniverse for industrial digitalization well all of this momentum that we've already seen with openusd is about to get turbocharged Alliance for open USD was announced with Pixar Apple Adobe Autodesk Nvidia as the founding members the alliance's mission is to Foster development and standardization of open USD and accelerate this adoption so whatever momentum we've already enjoyed the vision that we've already enjoyed it's about to get kicked into Turbo Charge well I want to thank all of you for coming today and remember remember accelerated Computing and generative AI the more you buy the more you save [Applause] [Music]

Channel: Ticker Symbol: YOU

Views: 184,963

Keywords: nvidia, nvda, nvidia stock, nvda stock, nvidia gtc 2023, jensen huang, nvidia keynote, openai, chatgpt, gpt4, msft, microsoft stock, msft stock, goog, googl, goog stock, google stock, artificial intelligence stocks, nvidia stock news, semiconductor stocks, tsmc, tsm stock, asml, asml stock, gpt-4, nvidia news, jensen huang keynote, nvidia 2023, ai copilot, computex 2023, nvidia computex 2023, nvidia keynote 2023, omniverse, ai stocks, best ai stocks, nvidia siggraph 2023

Id: dvRsZ4-wUGw

Length: 26min 42sec (1602 seconds)

Published: Wed Aug 09 2023