GTC Sept 2022 Keynote with NVIDIA CEO Jensen Huang
Computing is advancing at incredible speeds. The engine propelling this rocket is accelerated computing, and its fuel is AI. Today, we will show you new advances in NVIDIA RTX, NVIDIA AI, and NVIDIA Omniverse How these platforms propel new breakthroughs in AI, new applications of AI, and the next wave of AI for science and industry We will announce new chips that power these new applications, and, for the first time, cloud services to extend the reach of NVIDIA platforms.
I want to thank all the sponsors, partners, researchers, developers, and customers who joined us to make this event amazing. Welcome to GTC! I want to show you something amazing. RacerX is a fully interactive simulation built with NVIDIA Omniverse. RacerX is physically simulated. The lighting, reflections, and refractions are ray traced - nothing is pre-rendered and baked in. The parts and joints of the cars are individually modeled.
Their physical properties affect driving dynamics. Things in the environment are not static props, but rigid body, cloth, and fluid simulations. Smoke, fire, and dust are volumetric simulations. RacerX is a simulation.
Future games will not have pre-baked worlds. Future games will be simulations. RacerX is running one single GPU. Let me tell you how we did it. We introduced the programmable shading GPU nearly a quarter of a century ago.
GPUs revolutionized 3D graphics and created a medium with an infinite palette for artists. At Siggraph 2018, we launched NVIDIA RTX, a new GPU architecture that extended programmable shaders with two new processors. RT cores accelerate real-time ray tracing. Tensor cores process matrix operations central to deep learning. RTX opened a new frontier for computer scientists and a flood of new algorithms have appeared. A new era of RTX neural rendering has started.
Today, we're announcing Ada Lovelace, our 3rd generation RTX. One Ada GPU powers RacerX. This generation is named after mathematician Ada Lovelace who is often regarded as the world's first computer programmer. NVIDIA engineers worked closely with TSMC to create the 4N process optimized for GPUs. This process lets us integrate 76 billion transistors and over 18,000 CUDA cores, 70% more than the Ampere generation.
Ada evolves all three RTX processors: First, a new streaming multiprocessor with 90 Teraflops—over 2x the previous generation. Ada's SM includes a major new technology, called Shader Execution Reordering, which reschedules work on the fly giving a 2-3x speed-up for ray tracing. SER is as big an innovation as out-of-order execution was for CPUs. Second, a new RT Core, with twice the ray-triangle intersection throughput and two important new hardware units. A new Opacity Micromap Engine speeds up ray tracing of alpha-test geometry by a factor of 2x and a new Micro-Mesh Engine, which increases geometric richness without the BVH build and storage cost. And third, a new Tensor Core with the Hopper FP8 Transformer Engine and 1.4 petaFLOPs of tensor processing.
Raw ray tracing horsepower is not enough to ensure high frame rates. Ray tracing is notoriously hard to parallelize because rays bounce in every direction and intersect surfaces of various types. GPUs are highly parallel and most efficient when processing similar work at the same time. Ray tracing workloads lead to different threads processing different shaders or accessing memory that is hard to coalesce or cache.
Shader Execution Reordering (SER) improves execution efficiency by rescheduling shading workloads on the fly to better utilize the GPU's resources. We're seeing up to 2-3x increase in ray tracing and 25% in overall game performance. Like accelerated computing, computer graphics is a full-stack challenge. Breakthroughs require innovation in architecture and design, as well as algorithms. For example, NVIDIA's RTXGI uses ray tracing to do real-time, multi-bounce indirect lighting.
RTXDI uses ray tracing to do direct illumination from millions of lights, casting shadows from all lights. RTXDI is used for emissive surfaces such as billboards, TV screens, and neon tubes. NVIDIA Real-Time Denoisers (NRD) is a spatio-temporal denoising technique that takes an incomplete ray-traced image and infers the ground-truth, reducing the number of rays needed.
And DLSS, Deep Learning Super Resolution, is one of our greatest achievements. Ray tracing requires insane amounts of computation. Each frame of a CGI movie takes hours to render. We want to do this in real-time. NVIDIA RTX opened the world to real-time ray tracing.
RT Cores do BVH traversal and ray-triangle intersection testing, which saves the SMs from spending thousands of instructions on each ray. But even with RT Cores, frame rates were too low for games. We needed another breakthrough. Enter deep learning. DLSS uses a convolutional autoencoder AI model and takes the low-resolution current frame and the high-resolution previous frame, to predict, on a pixel-by-pixel basis, a higher resolution current frame.
The AI model is trained to predict an ultra-high resolution 16K reference image. The difference between the predicted and reference image is used to train the neural network. The process is repeated tens of thousands of times until the network can predict a high-quality image.
Ada introduces DLSS 3, a new AI that generates entirely new frames, rather than just pixels. DLSS 3 has four components - a new Optical Flow Accelerator, game engine motion vectors, a convolutional autoencoder AI frame generator, and our Reflex super-low-latency pipeline. DLSS 3 processes the new frame and the prior frame to discover how the scene is changing. The Optical Flow Accelerator provides the neural network with the direction and velocity of pixels from frame to frame. Pairs of frames from the game, along with the geometry and pixel motion vectors, are then fed into a neural network that generates the intermediate frames.
DLSS 3 generates entirely new frames, without processing the graphics pipeline, boosting game performance by up to 4X over brute-force rendering. Because DLSS 3 generates a new frame without involving the game, it benefits both GPU- and CPU-limited games. For CPU-limited games, like physics-heavy or large-world games, DLSS 3 allows the Ada GPU to render the game at a much higher framerate than the CPU is able to compute the game.
DLSS 3 is one of our greatest neural rendering inventions. Here is Cyberpunk 2077 shown in an all new Max Ray Tracing Mode with SER and DLSS 3. Pushing state-of-the-art graphics into the future takes an incredible amount of computational horsepower. In a modern game like Cyberpunk, we run over 600 ray tracing calculations for every pixel just to determine the lighting. That's a 16x increase from the time we first introduced real-time ray tracing 4 years ago. Yet the number of transistors we have available in a GPU to do these calculations has not increased at this rate.
That's the power of RTX— we can deliver a 16x increase in 4 years with AI. Some pixels are calculated, most are predicted. Let's enjoy Microsoft Flight Simulator, a game that is CPU-limited because of the realistic physics and the giant worlds. It is one of the longest running franchises.
This is Flight Sim's 40th anniversary! With Ada and DLSS 3, Flight Simulator is silky smooth. I am thrilled to announce Portal RTX. NVIDIA Lightspeed Studios used Omniverse to remaster one of the most celebrated games in history. Portal was developed by Valve and released in 2007. It was GDC's Game of the Year and exhibited at the Smithsonian. Portal RTX is both nostalgic and futuristic at the same time.
Let's take a look. Portal with RTX is a mod. We did it with Omniverse.
Modding is a massive culture. Everyone is a creator. There are millions of modders and billions of mods are downloaded each year. In fact, 9 of the 10 most popular competitive games owe their existence to mods We created an Ominverse application called RTX Remix for game modding. Start the game and capture the game into USD, which is loaded into Omniverse. Here, we're using Elder Scrolls Morrowind from Bethesda, one of the top modded games of all time.
Once in RTX Remix, the AI-assisted toolset has deep learning models to up-res textures and assets and an AI model to convert materials to have physically accurate properties. The rich Omniverse ecosystem of creative tools can then be used to enhance the game assets. When done, export the RTX Mod pack and play the game with the RTX renderer. RTX Remix is incredible tech and the most advanced game modding tool ever created.
Portal RTX and RTX Remix will be available shortly after Ada launch. RTX neural rendering algorithms run on programmable shaders, RT cores, and Tensor Cores to create amazing images. The total processing throughput of Ada is a massive leap over the Ampere generation and the performance shows. For rasterized games, Ada is up to 2x faster and it's 4x faster for ray-traced games. Ada is incredibly energy efficient - over twice the performance at the same power compared to Ampere.
And you can really push Ada - we've overclocked Ada past 3GHz in our labs. RTX reinvented graphics. Now Ada is paving the way for future games to be fully simulated worlds, like RacerX.
Today, we are announcing our highly anticipated NVIDIA Ada Lovelace GPU The GeForce RTX 4090 NVIDIA engineers pushed technologies on every front: New SM with Shader Execution Reordering New RT Core with Opacity Micromap and Micro-Mesh Engines New Tensor Core with FP8 Transformer Engine And a pixel processing 300 TOPS Optical Flow Accelerator for DLSS 3 Overall, 4x more processing throughput. Compared to the world's reigning GPU champion, the 3090Ti 4090 is 2x faster on Microsoft Flight Simulator... 3x faster on Portal RTX...
and 4x faster on RacerX. The GeForce RTX 4090, the new heavyweight champ, is $1,599. Available October 12th. The GeForce RTX 4080 comes in 16GB and 12GB versions.
4080 is three times the performance of 3080 Ti on RacerX. 4080 starts at $899. The best gaming platform in the world just got better.
Our 30-Series, starting at $329, are the best GPUs in the world serving mainstream gamers. RTX 4090 and 4080 GPUs, starting at $899, deliver the ultimate performance to enthusiasts. The Ada Lovelace generation advances all three RTX processors, the engines of neural rendering. Ada is a quantum leap for gamers and paves the way for creators of fully simulated worlds, like Omniverse. Welcome to the Omniverse! The next evolution of the internet, called the metaverse, will be extended with 3D. Today's internet connects websites described in HTML and viewed through a browser.
The metaverse, the 3D internet, connects virtual 3D worlds described in USD and viewed through a simulation engine. Omniverse is a platform for building and operating metaverse applications. Omniverse is useful wherever digital and physical worlds meet.
One of the most essential uses of Omniverse is robotics, the next wave of AI, where AI interacts with the physical world. Omniverse is like nothing ever built. Omniverse is a real-time, large-scale 3D database. It is a shared 3D world.
Omniverse is a network built on USD. It connects 3D worlds. Omniverse is a computing platform. You can write applications that run on Omniverse. These applications are like portals into the Omniverse virtual world. We've written some Omniverse applications like Create and View for collaboration, Replicator for synthetic data generation, and Isaac Sim and Drive Sim for digital twin simulations. We are releasing a major update to Omniverse: Support for Ada Lovelace GPUs, with a giant leap in ray tracing and large-scene performance.
New GAN-based and Diffusion model-based neural rendering tools. OmniGraph is a graph execution engine to procedurally control behavior, motion, and action. A major update of Omniverse physics to process kinematics of complex multi-part objects.
New Cloud XR for Ada's amazing raytracing in VR. The world's first SimReady asset library for Synthetic Data Generation and Digital Twin simulations. One of the most popular applications of Omniverse is Replicator, which is used to generate synthetic data to train self-driving cars, robots, and all kinds of computer vision models. New Omniverse JT Connector. This is big. Siemens invented JT, the industry-standard language of product lifecycle management and the interoperability format of CAD systems, like NX, Creo, Catia, and Inventor. The JT Connector opens the industrial and manufacturing worlds to Omniverse.
Omniverse is an enterprise platform that is useful from the beginning of design and styling, through engineering and manufacturing, marketing to operations— the entire product lifecycle. At the core of making movies or games, building cars or any consumer product, or building and operating factories and warehouses, is a super complex 3D pipeline. Let's take a look at the creative and design process.
Many specialized tools are used by expert designers, artists, and engineers working in different teams, organizations, or geographies, passing work around. Different data formats are stored all over the company. A visual effects pipeline can use 5 to 7 tools; each specializes in a certain aspect of the workflow, such as modeling, rigging, animation, or simulations. DNEG, an award-winning visual effects and animation studio, uses Autodesk Maya, SideFX Houdini, Adobe Substance 3D, and Foundry Nuke in their workflow. Just as the internet connects websites, Omniverse connects 3D worlds. Omniverse lets DNEG's creators bring in data from different tools in their full fidelity and collaborate interactively in a shared Omniverse world.
Teams can create a Custom 3D Pipeline to orchestrate and manage the complex workflow using simple out-of-the-box Omniverse tools. These Omniverse capabilities are valuable to nearly every industry. NVIDIA RacerX was created by 30 artists, using 11 tools collaborating across 12 time zones, and done in only 3 months.
The team used Omniverse to connect and collaborate and created a Custom 3D Pipeline to orchestrate and manage their work. The architectural design process commonly uses apps like Revit, Rhino, SketchUp, ArchiCAD, and 3ds Max. KPF, an award-winning architectural design firm connects architects, designers, and clients in Omniverse.
Clients can review progress by interacting with a full-fidelity and photo-realistically rendered model. Auto and other physical product makers commonly use Siemens, Autodesk, Ansys, and PTC tools. GM is creating a digital twin of their Michigan Design Studio in Omniverse where designers, engineers, and marketers can collaborate. Omniverse gives them a single source of truth that is physically simulated and faithfully rendered. Omniverse is a virtual world used to design, build, and operate things in the physical world. In the future, everything made will have a digital twin, used by the makers to emulate the behavior and performance of a product in the physical world.
No physical product operated by software can be deployed in large numbers without testing the software on its digital twin. Let me give you some examples. Telcos will deploy millions of 5G microcells and towers over the next 5 years. Optimizing the placement and operations of new and existing cells and towers can save billions for the 2 trillion dollar telecom industry. Charter, a $50 billion dollar telecommunications provider and HeavyAI, a super cool interactive data analytics platform company, are using Omniverse to create digital twins of their 4 and 5G networks at metro and nationwide scales.
HeavyAI uses Omniverse to combine multi-billion-point LIDAR scans, satellite imagery, buildings, trees, and surroundings and their RF propagation model to simulate the performance of Charter's radio network in real time. Lowe's is one of the world's largest home improvement retailers with over 2,000 stores and 300,000 retail associates. Lowe's is using Omniverse to design, build, and operate digital twins of their stores to optimize operations and enhance the shopping experience.
And with the amazing Magic Leap 2 headset connected to Omniverse, associates can enter a mixed physical and Omniverse world to explore new store designs. The $300 billion-dollar global railroad market helps move the world a critical part of the global supply chain. Deutsche Bahn, the national railway company of Germany uses Omniverse to build and operate a digital twin of 5,700 stations and 33,000 km of track. Omniverse is also used to train AI models that can continuously monitor the railways and trains and recognize hazards and situations that can affect network operations.
Deutsche Bahn expects to increase capacity and efficiency of the railway, and reduce its carbon footprint without building new tracks. Each of these stories are amazing. Be sure to check their videos on the NVIDIA homepage. There are many great examples of how companies use Omniverse to create digital twins of factories, logistics warehouses, automated manufacturing lines, and industrial plants. Let's take a look. Connections to Omniverse are growing fast.
There are already 150 connectors to Omniverse. These are the tools and platforms used by the world's $100 trillion industries. These connectors open Omniverse to companies that span the largest industries from retail, transportation, telecommunications, manufacturing, media and entertainment, consumer and luxury goods, to supply chain and logistics.
Here are some examples of companies who are using Omniverse to connect their teams, visualize their data in full fidelity, generate synthetic data to train AI models, and simulate digital twins. Omniverse is a new computing platform and requires a new computer system. There are three elements to the Omniverse computer: RTX computers for creators, designers, and engineers.
OVX servers to host connections to the Nucleus database and run virtual world simulations. And third, the NVIDIA GDN, portals into the Omniverse. Through GeForce Now, we've built a global GDN, a Graphics Delivery Network, that reaches 100 regions with super-fast and responsive RTX graphics. Whereas a Content Delivery Network, or CDN, allows internet video to be streamed efficiently, NVIDIA GDN can stream interactive graphics efficiently. Between NVIDIA RTX PCs, NVIDIA GPUs in the cloud, and NVIDIA GDN, we can build an Omniverse Computer that covers the planet.
Let me show you how Rimac, a hypercar and advanced EV platform company, uses Omniverse to connect their design, engineering, and marketing pipelines to create the amazing Nevera. Rimac designs on RTX workstations and publishes the Omniverse-based car configurator application to run on NVIDIA GDN, which is then streamed to be enjoyed on any device. NVIDIA Omniverse Cloud is a suite of software and infrastructure-as-a-service for designing, publishing, and experiencing metaverse applications from anywhere, on any device. Let's see how Rimac, technology pioneer of AI-enabled hypercars and advanced electric solutions, can enable collaborative workflows for their 3D teams, and deliver advanced 3D experiences to their audiences with Omniverse Cloud.
The 3D models of cars such as Rimac's are designed by large teams of uniquely skilled individuals each using their own specialized tools. Here, designers connect to Omniverse Nucleus Cloud, their shared database engine. With USD workflows, they can aggregate full CAD fidelity datasets from their favorite tools including Autodesk Alias, Houdini, Adobe Substance 3D Painter, and Omniverse Create. Each designer either works on their individual RTX computers or streams Omniverse apps from OVX servers in the cloud. Once complete, Rimac integrates the ground-truth USD model into an advanced, real-time 3D configurator and publishes it to the NVIDIA Graphics Delivery Network— a planetary-scale network of distributed data centers that stream high-performance, low-latency 3D experiences to edge devices. Today's configurators are limited by pre-rendered options that can only be shown in a single scene.
With the end-to-end USD pipeline, Rimac fans and consumers are configuring the actual engineering dataset— not a simplified model— in full fidelity with physically accurate materials and real-time RTX ray tracing. This unlocks all possible design permutations, without Rimac having to manually render every option as a layer. Omniverse Cloud connects 3D artists, designers, and metaverse viewers at every edge of the planet and delivers the possibility to build and operate advanced 3D internet applications on any device. Today we're announcing NVIDIA Omniverse Cloud, an infrastructure-as-a-service that connects Omniverse applications running in the cloud, on-prem, or on a device. In addition, Replicator and Farm will also be available in the cloud.
Farm is a scaling engine for a render farm. Omniverse Cloud, Replicator, and Farm containers are available on AWS today. We are also offering them as managed services. Sign up today for early access. Robotics is the next wave of AI.
Breakthroughs in deep learning have opened the door to creating systems that can perceive their surroundings, plan a sequence of actions, and perform useful tasks, in real-time, every time. Everything from the way software is developed to how it runs is completely different. A new type of processor had to be created. NVIDIA Xavier was the world's first robotics processor designed for deep learning.
Since then, every two years, we've announced a giant leap in performance. The driving force of the increased performance is the simultaneous increase in number and resolution of sensors, and more sophisticated AI models needed to expand driving domains and improve safety. Safety is the #1 priority in robotics, and it demands diversity and redundancy in sensors and algorithms. All of that demands more processing.
Last year we introduced Atlan, a 1,000 TOPS SOC. Today, we are announcing that Atlan is no more... And will be replaced by Thor— twice the throughput of Atlan and more than twice the delivered performance.
There are three big reasons we did that: Grace, Hopper, and Ada Lovelace. The incredible transformer engine of Hopper, and the fast evolution of Vision Transformers, is too important not to include in our next robotics processor. The invention of multi-instance-GPU in Ada is a great opportunity to centralize the car computer, reducing hundreds of dollars of cost. And the CPU of Grace is too amazing to pass up.
As all parallel processing algorithms are off-loaded and accelerated by our GPU, the remaining workload tends to be single-thread limited. Grace has excellent single-threaded performance. So, engineers at NVIDIA scrambled to create Thor. The amount of new technology in Thor is insane.
Thor can be configured in multiple ways. can dedicate all of its 2,000 TOPS and 2,000 TFLOPS to an autonomous driving pipeline. Or, it can be configured to use a portion for cabin AI and infotainment and a portion for driver assistance. Thor's multi-compute-domain isolation lets concurrent time-critical processes run without interruption.
On one computer, you can simultaneously run Linux, QNX, and Android. Thor centralizes numerous computers and simultaneously offers a leap in capability while reducing cost and power. Today, parking, active safety, driver monitoring, camera mirrors, cluster, and infotainment are different computers.
In the future, these functions will no longer be separate computers, but will delivered by software that runs on Thor and improves over time. Building robotic computers requires two computers: an AI factory in the datacenter that processes data, trains the AI, simulates the digital twin, and maps the world, and an AI computer in the car that processes the sensors to perceive the environment, stay clear of obstacles, and drive the car to its destination. NVIDIA DRIVE is an end-to-end platform for autonomous vehicle development and deployment. For development, DRIVE includes Replicator Synthetic Data Generation, NVIDIA AI infrastructure, DRIVE Sim, and DRIVE Map. For deployment, DRIVE includes the full-stack driving and in-cabin AI applications, the AI computer, and the Hyperion reference autonomous vehicle system.
Let me show you some new, exciting capabilities in the NVIDIA DRIVE platform. First, an AI pipeline called Neural Reconstruction Engine that came out of NVIDIA research is now a major feature in Drive Sim. Creating scenarios for simulation is laborious and difficult to scale. Our researchers have developed an AI pipeline that constructs a 3D scene from recorded sensor data. The 3D scene is imported into DRIVE Sim and can be augmented with human-created content or AI-generated content. This video-to-3D-geometry pipeline runs on NVIDIA OVX systems and enables us to create simulation scenarios on a global scale.
Let's take a look. We're supercharging DRIVE Sim with a Neural reconstruction Engine, to enhance simulation with AI and data captured by the fleet. In a matter of minutes, the neural reconstruction engine can reconstruct a full 3D digital twin of a recorded drive from sensor data. Objects are harvested and reconstructed using AI.
Large scale asset and scene libraries are created from real world drives. The reconstructed scenes and assets are loaded into Omniverse and are ready to use in DRIVE Sim. Using NVIDIA DRIVE Map, we can place dynamic objects such as vehicles As well as pedestrians. Recorded data can now be turned into fully reactive and modifiable simulation environments for closed loop testing. New scenarios can be authored using synthetic and harvested scenes and assets.
We can generate synthetic ground truth data to train perception networks. These authored scenarios can be used for end-to-end testing Including challenging scenarios NVIDIA DRIVE Sim, powered by Omniverse and AI. DRIVE Sim is essential to building our driving system from beginning to fleet operations. It's an integral part of our CI/CD, continuous integration/continuous deployment, engineering process. One of the important DRIVE Sim features is Hardware-in-the-Loop, which lets us run the entire in-car software stack right there in the AI factory. The AI car computer has no idea it's actually inside a simulation and not in the car and driving on the road.
DRIVE Sim with Hardware-in-the-Loop can also simulate the inside of the car. Future cars will not just have simple dashboards, but surround displays where digital and physical design come together. Car designers, software engineers, and electronics engineers can collaborate in DRIVE Sim while running all the actual computers and software stacks. DRIVE Sim will be their virtual design studio. Let me show it to you.
We are extending NVIDIA DRIVE SIM's platform to create a digital twin of the car's interior, with full hardware-in-the-loop support. NVIDIA Omniverse is the platform that brings this all together. Designers and engineers can work side by side to perfectly integrate the physical design with the digital interface before the actual car exists. - Let's try a new cluster layout - Sure. - Can you resize the screen? - Ok, lets try this.
- This looks good. DRIVE Sim, built on Omniverse, will accelerate the development of new AI cockpit and in-vehicle infotainment systems, plus improve usability by testing them in a digital twin of future cars. Safety is the number one priority in robotics. And as I mentioned, building robotics systems requires building two computers. An AI Factory in the datacenter, and an AI computer in the car.
Safety architecture, design, and methodologies are pervasive throughout our systems, from datacenter to the car. NVIDIA has invested 15,000 engineering years in safety systems and processes. Five million lines of code have been safety-assessed. Our DRIVE chips and platform are designed for ASIL D operation and ISO 26262 safety. We've dedicated ourselves to an end-to-end safety approach that extends from the AI factory to the fleet.
We're making excellent progress developing the DRIVE end-to-end autonomous driving system. Let me show you our progress from Replicator synthetic data generation to AI model advances, DRIVE Sim with hardware-in-the-loop, DRIVE Map autonomous fleet mapping, urban and highway driving, and parking. Let's take a look.
- Hi Pavithra, you have an event at Plaza de Cesar Chavez Park at 6pm. Would you like me to navigate there? - Yes, please. - Beginning drive pilot.
- You have arrived. Robotic computers are the newest types of computers. These are real-time systems that sense the environment, reason about their surroundings, and plan actions consistent with their goals. Orin is our second-generation processor, designed specifically for robotics, and it is a homerun. Orin is the computing engine of autonomous vehicles and has been selected by over 40 companies building cars, trucks, robotaxis, and shuttles.
The fundamental processing pipeline for autonomous vehicles can be applied to all kinds of robotics systems. Jetson is our robotics computer with a million developers and used by some 6,000 companies, including 1,000 startups. Amazon, Boeing, Canon, Microsoft, Shopify, and Teradyne are among the many companies building robots with Jetson. We are working with the industry on amazing robots— from pick-n-pack robotic arms, to agriculture robots, AMRs (autonomous mobile robots), to last-mile delivery robots. Today, we are announcing the Jetson Orin Nano, a tiny robotics computer that is 80x faster than our previous super-popular Jetson Nano. Jetson Orin Nano runs the NVIDIA Isaac robotics stack and features the ROS 2 GPU-accelerated framework.
There are robots that move, and there are robotic systems that watch things that move. Metropolis is our edge AI platform. It's been downloaded 1 million times and has more than 1,000 application partners around the world. Metropolis makes sense of cameras, lidar, and other IoT sensors to make warehouses, factories, retail stores, and cities safer and more efficient. Orin is also the robotics processor of our industrial-grade IGX Edge AI platform that Metropolis runs on.
Today, we're announcing that Siemens, the world's largest industrial automation company, is adopting Metropolis and Orin IGX for their industrial edge computing platform. Each year, two million medical devices are made by 16,000 companies. In the past, medical instruments were built from bespoke sensor-processing chips dedicated to one function. Driven by the same breakthroughs that have enabled self-driving cars and robots, future medical instruments will also be software-defined and powered by artificial intelligence. The same fundamental robotics pipeline of sensor processing, environment reconstruction, detection and segmentation applies to medical imaging systems.
Orin IGX is also an ideal computing platform for medical imaging and robotics. And running on Orin IGX is NVIDIA Clara Holoscan, a low-latency imaging processing platform. Clara Holoscan, like all NVIDIA robotics platforms, includes libraries for data processing, AI model training, simulation, and the robotics application. More than 70 leading medical device companies— including Siemens Healthineers, Intuitive, and Olympus— startups and medical centers are developing on Clara Holoscan. Today we're announcing that Activ Surgical, Proximie, and Moon Surgical will build their surgical robotics systems on NVIDIA Clara Holoscan running on the Orin IGX Platform. There are no robots with universal skills, but building robotic systems share common methods and tools.
Robotics requires two computer systems: one being the AI factory for developing the robot AI, and the other an AI computer to operate the robot. For several fast-growing application domains, we created platforms that include domain-specific libraries for data generation, AI training, simulation, mapping, and the runtime stack. We've built one called Isaac for AMRs, autonomous mobile robots. AMRs can be used to move things in warehouses, distribution centers, and factories, or be autonomous cleaning, roaming security, hospitality, and last-mile delivery bots. AMRs don't need fixed, pre-programmed tracks, and can autonomously navigate from waypoint to waypoint, avoiding and maneuvering around obstacles. AMRs are basically self-driving cars for unstructured environments.
The Isaac platform consists of useful tools that can benefit all AMR developers: Sim Ready Asset libraries of things that are often found in warehouses and factories; Replicator for synthetic data generation; Isaac Sim with connectors to the ROS and other popular robotics ecosystems; Isaac ROS engine with AI and computer vision algorithms and runtime that connects with the ROS Bridge; and cuOpt, a real-time fleet task-assignment and route-planning engine. CuOpt is an amazing GPU-accelerated optimization engine. It's like having a pre-Quantum computer. Let's take a look at the Isaac Platform in action.
Making the right operational decisions for your warehouse environment is key to maximizing output while minimizing both upfront and ongoing costs. Using NVIDIA cuOpt in Isaac Sim on Omniverse Cloud, you can make dynamic, data-driven decisions at any point during the layout and operation of your warehouse. Isaac Sim in the cloud lets us tap into the accelerated performance of NVIDIA OVX to generate thousands of environments and layouts in a fraction of the time that it would take using a single-GPU system. CuOpt can then provide optimized task assignments and routing solutions informed by a collision-based occupancy map that will dictate the movement of the Autonomous Mobile Robot, or AMR, fleet within the warehouse. Distributed teams can access designs from any device and easily vary parameters such as budget, speed of delivery, and robustness to identify the ideal layout for their operational needs. High volumes of synthetic data, generated by Isaac Replicator in the cloud, are then used to train the AMR's perception model.
Navigation is validated with software-in-the-loop in a true-to-life, simulated environment. Once the AMRs are operating in the physical warehouse, cuOpt can reoptimize as obstructions and environmental changes are identified. Explore how you can accelerate your logistics planning and optimization workflows with NVIDIA cuOpt in Isaac Sim on the cloud today. Today, we are announcing that Isaac is going to the cloud. On NGC, you can get the cloud-ready Omniverse VMI (virtual machine image) and Isaac containers and deploy it on any public cloud. For AWS, we've gone one step further and are publishing optimized Omniverse AMI and Isaac containers in the Marketplace.
For ROS developers using Robomaker, you can pull the Isaac container from the Marketplace and connect Isaac Sim to your ROS ecosystem and robotics simulation and scale your CI/CD pipeline. Isaac will run on the many options of NVIDIA GPU instances in AWS, Azure, GCP, and OCI. NVIDIA is dedicated to advancing science and industry with accelerated computing. The days of no-work performance scaling are over. Unaccelerated software will no longer enjoy performance scaling without a disproportionate increase in cost. With nearly three decades of a singular focus, NVIDIA is expert at accelerating software and scaling compute by a Million-X, going well beyond Moore's law.
Accelerated computing is a full-stack challenge, demands deep understanding of the problem domain, optimizing across every layer of computing, and all three chips: CPU, GPU, and DPU. Scaling across multi-GPUs and multi-nodes is a datacenter-scale challenge and requires treating the network and storage as part of the computing fabric. And developers and customers want to run software in many places— from PCs to supercomputing centers, enterprise datacenters, cloud, to edge.
Different applications want to run in different locations and in different ways. Today, we are going to talk about accelerated computing across the stack: New chips and how they will boost performance far beyond the number of transistors. New libraries and how it accelerates critical workloads to science and industry. New domain-specific frameworks to help develop performant and easily deployable software. And new platforms to let you deploy your software securely, safely, and with order-of-magnitude gains.
The NVIDIA platform now boasts 3.5 million developers. 12,000 start-ups are founding their companies on NVIDIA. We accelerate some 3,000 apps through 550 SDKs and AI models. And, collectively, the industries we serve represent some $100 trillion of value. Over the past 12 months, we updated more than 100 SDKs and introduced 25 new ones. With each new update, we enhance the performance and throughput of our fleet of computers.
With each addition, we solve new challenges for developers and extend accelerated computing into new opportunities. Omniverse Cloud, new OV Connectors like JT, Drive Sim, Isaac ROS and Isaac Sim, Sim Ready Assets and Replicator, and Clara Holoscan open new worlds to accelerated computing. Let me update you on a few more of our most popular and newest acceleration libraries.
RAPIDS is our open-source suite of SDKs that accelerate data science and data engineering. With RAPIDS, developers process DataFrames, SQL arrays, machine learning, and graph analytics and do in minutes what used to take hours, or process 10-100x the amount of data. and do in minutes what used to take hours, or process 10-100x the amount of data.
More than a quarter of the world's top companies are using RAPIDS. There are over 100 integrations with popular open-source and commercial software toolkits. And RAPIDS is the foundation of Merlin, Morpheus, cuOpt, Clara, Triton, and other NVIDIA SDKs. We are expanding the reach of RAPIDS. 10 million Python developers on Windows can now access RAPIDS through WSL. And RAPIDS now supports the Arm Server Based System Architecture.
Apache Spark is the most popular data processing engine and is used by 80% of Fortune 500 companies. NVIDIA RAPIDS has a plugin for Spark 3 and integration with Delta Lake and Apache Iceberg. Enterprises can transparently accelerate Spark dataframe and SQL operations.
AT&T is using RAPIDS-accelerated Spark to reduce processing time for their advertising pipeline. AT&T sped up processing by 4x and reduced cost by 70%. IRS is using RAPIDS-accelerated Spark to speed processing by 20x while reducing cost in half. Adobe Intelligent Services platform was able to reduce cost by two thirds. Data processing is a giant part of cloud spend— tens of billions of dollars each year and billions of kilowatts of power.
NVIDIA RAPIDS-accelerated Spark is now integrated into GCP Data Proc and ready to help save millions on your data processing spend. Triton is our open-source hyperscale inference serving software that supports any AI model; supports GPUs, x86 and Arm CPUs, and AWS Inferentia; and supports all major frameworks. Triton has been downloaded over three million times, an increase of 300% from last year. Available in all major public clouds and integrated into the top MLOPS platforms, Triton is used by over 35,000 companies. Airtel, the 2nd largest wireless provider in India, uses Triton to serve ASR and speech analytics for customer service. Nio, the leading premium EV car company in China, uses Triton to test their autonomous driving models against 100TB of data daily.
GE Healthcare has standardized on Triton for their datacenter and embedded systems. Microsoft uses Triton to process real-time grammar checking. Amazon uses Triton to process real-time spell-checking on their sites. We introduced over 50 new features in the last year. New model orchestration to automatically load models on demand, efficiently allocating GPU and memory resources.
A major new feature is large language model inference with multi-GPU and multi-node execution. Inference time is accelerated from many minutes to a second. Graph databases store objects and their relationships as nodes and edges in a graph. Whether for fraud detection, drug discovery, supply chain optimization, product recommendation, contact tracing, customer 360 journey analysis, or social media, it is all about finding patterns and relationships. The largest companies in financial services, retail, healthcare, and manufacturing have their most valuable data stored in graph databases. Graph sizes can easily reach trillions of nodes and edges.
And now with deep learning, we can learn patterns and relationships on a giant scale and train a predictive model for that database. NVIDIA now accelerates two of the most popular graph neural network frameworks Deep Graph Library and PyTorch Geometric. Amazon, American Express, Entos life sciences, Meituan, and Pinterest are among the leading users of this fast-growing deep learning approach. Over 80% of internet traffic is video. Increasingly, AI and computer graphics will augment the video. What used to be just streaming encoded video will include image processing like relighting, reposing, blurring backgrounds, super resolution, AI inference, and computer graphics for AR.
Special effects that used to require offline processing are now going into the cloud for live video. These operations are ideal for GPUs and can tap into the NVIDIA GPUs already in cloud datacenters for AI inference. But we need a new cloud-native library. CV-CUDA is an open-source, GPU-accelerated library for imaging and computer vision. CV-CUDA provides a set of highly performant computer vision kernels, with zero-copy interfaces to integrate efficiently with other libraries and deep learning frameworks. CV-CUDA can accelerate the end-to-end processing throughput by over 10X.
Early access application is now online. To advance quantum computing research, the community needs a platform to design and validate the quantum hardware and algorithms. The community needs a superfast pre-quantum computer that produces known-good results. NVIDIA's cuQuantum is an SDK for quantum circuit simulation. With cuQuantum, a 32-node DGX Pod can simulate a 40 qubits quantum computer.
cuQuantum has been adopted broadly— from AWS, Google, IBM, Oracle, startups all over the world, and supercomputing centers. AWS integrated cuQuantum into its Braket quantum computing service and saw 900x speed-up and 3.5x reduction in cost. Oracle is building a quantum simulation virtual machine for OCI cloud. When quantum computing arrives, it will likely be an extension of classical accelerated computing. NVIDIA QODA is an open, quantum processor-agnostic platform for hybrid quantum-accelerated computing.
Just as accelerated computing required a decade-long reformulation of software and algorithms quantum-accelerated computing will need the same. QODA offers researchers a programming model for quantum-accelerated computing that researchers can use to discover new science today. QODA running on DGX with cuQuantum fully emulates a quantum-accelerated computer. We're working with world-class quantum researchers to develop systems and new algorithms on QODA.
AI continues to make exponential advances with new algorithms and new frameworks to develop them. The rich programmability, performance, and availability of NVIDIA makes our platform ideal for AI research. JAX has become a super popular library for machine learning research.
JAX has the familiarity of Python and NumPy and the ability to define and compose differentiable functions needed for machine learning. JAX has an ecosystem of 300+ libraries spanning NLU, RL, drug discovery, neural rendering, and physics simulation. DeepMind's groundbreaking AlphaFold was done with JAX. NVIDIA engineers are hard at work with Google Research and DeepMind to optimize JAX for a major release coming in Q4. The NVIDIA AI for JAX container will be optimized for multi-GPU and multi-node scaling, optimized for all CSP platforms, support for Hopper's FP8 Transformer Engine, and all with an easy one-command to run.
Go to NGC and sign-up for NVIDIA AI for JAX. Large language models are the most important AI models today. Based on the Transformer architecture, large language models are giant and learn to understand the meaning of language from the corpus of human knowledge, without supervision and without labeled datasets. Research in Transformers and large language models is the most vibrant in AI, with over ten thousand research papers published already this year.
Unlike CNNs, Transformers can learn patterns and relationships across great spans of space and time. Transformers led to the breakthroughs in natural language processing and large language models. There are several things that make large language models so remarkable: A single pre-trained model can perform multiple tasks, like question answering, document summarization, text generation, translation, and even software programming. Models can be tuned with just a few examples to perform tasks for which they were never trained. Large language models follow the scaling law, and their performance continues to scale as you increase the size of the model.
Large language models can learn the representation of almost any complex sequences, for example, even the language of chemistry and biology. Large language models are excellent few-shot learners. The same pre-trained model, when asked the same question in different contexts, can produce a different response. With just a few examples, we can direct the model to be aware of our context and better perform its task.
There are endless possible applications for large language models— summarize a story, report breaking news, paraphrase for different audiences, or even searching for documents. Retraining the entire large model for each task would be understandably costly. Another method, called prompt learning, is an effective way to direct a pre-trained large language model to perform specific tasks by training a companion model with only a few examples. NVIDIA NeMo LLM is a prompt learning framework.
Using the large language model's natural ability to fill in the blanks, we provide Nemo some examples, or a few shots of learning, to learn an additional token, a prompt, that when prepended to an input phrase, will produce the desired output phrase for that context. NeMo, given a small set of examples, trains a prompt encoder. And only the prompt embeddings are updated. The pre-trained large language model's parameters are frozen during prompt learning, which substantially reduces the cost and time of fine tuning the language model for each task. Today, we are announcing NeMo LLM service, a cloud service that trains a large language model to perform a task from examples you provide. Nemo includes a selection of pre-trained, community-built foundation models.
The NeMo API produces the learned prompt embeddings and an optimized microservice that can be deployed on-prem, in any cloud, for one GPU or multi-GPU, multi-node. The Nemo Service will make it easier for researchers, developers, and enterprises to apply these incredible large language model. Come try it. Sign up now for early access in October. We used NeMo to create the Megatron 530B Cloud API. 530B is an NVIDIA managed-service, running on our DGX SuperPod, that you can connect to your application.
It has already been prompt-tuned for a few tasks: summarization, paraphrasing, and story generation. Early access starts in October. Transformers and large language models are also at the core of the digital biology revolution. Just as Transformers can learn the relationships of words and sentences and understand human language, Transformers can learn to read and write in the native language of biology and chemistry. MegaMolBART learned the language of chemistry. ESM-1 and ProtT5 learned the language of proteins.
Once the language is learned, biochemical large language models can be prompt-tuned to generate chemicals of certain properties and proteins within a family. These generated target proteins and drug candidates can then be virtually screened. The drug discovery space is practically infinite.
The combination of proteins and chemicals is unimaginably large. Only about 50 drugs are approved each year, costing an average of a billion dollars to discover. Large language models give us a new tool to explore the infinite world of drug discovery. But setting up environments to train these models is hard and processing the data and training can take weeks to months Today, we are announcing BioNemo large language model Service, a digital biology framework for researchers and developers to create large language models that understand chemicals, proteins, DNA and RNA sequences. Included in the service are pre-trained ESM-1, ProtT5, and MegaMolBART. The output of BioNeMo can be used for downstream tasks such as generating new proteins and chemicals, or predicting structure, function, or reaction properties.
Early access for BioNeMo is in October. The Broad is one of the top medical research institutes in the world and the largest producer of human genomic information. The Broad Institute and NVIDIA are announcing today that the NVIDIA Clara libraries are now available on Broad's Terra Cloud Platform.
With NVIDIA Parabricks, whole genome sequencing can be sped up from over 24 hours to just one hour, while cutting the compute cost by more than half. Researchers using the Genome Analysis Toolkit, GATK, can get improved accuracy with deep learning. The new NVIDIA BioNeMo will also be available on Terra Cloud. Using BioNeMo, NVIDIA and Broad researchers are collaborating to create large language models for genomics. We're excited that NVIDIA's Clara engines will turbocharge the work of Terra Cloud's 25,000+ users. Large language models have set the stage for Hopper.
Named after Grace Hopper, a computer scientist in the U.S. Navy, who pioneered computer programming in the '50s and first coined the term "bug," Hopper is the engine of large language model AI factories. One of the most amazing breakthroughs of Hopper is a new Transformer Engine.
A new Tensor Core and software that uses FP8 and FP16 numerical formats dynamically processes layers of a Transformer network. And the performance is incredible. The training time goes from days to hours, or on larger models, from a month to a week. And prompt-tuning on models such as 530B using H100 and Nemo LLM service is 5x faster than A100. And as Kirk said to Spock, "Hours instead of days, minutes instead of hours." H100 inference is magnitudes faster than A100 for large language models.
For instance, when serving large language model chatbots, with less than 1 second latency, on equivalent scale systems, Ampere can serve 100 users while Hopper can serve more than 3,000—30x more. Hopper will make large language models, the most important AI of our time, accessible to everyone. The more you buy, the more you save. Hopper H100 is the new engine of the AI factory. H100 is available now on NVIDIA LaunchPad.
DGX H100 systems, with NVIDIA Base Command orchestration, can be ordered now. OEM partners will ship H100 systems starting next month. And all major cloud providers are making H100 available in just a few months. Hopper is in full production.
Recommenders run the digital economy. They are the engines behind social media, digital advertising, e-commerce, and search. Recommenders process a gigantic amount of data. Compared to large language models, the amount of data moved for every unit of computation is an order of magnitude larger. Here's a chart that has Compute Intensity on the vertical axis, and Fast Memory Capacity on the horizontal axis.
You can see that Recommender systems require an order of magnitude more Fast Memory Capacity. Large language models will continue to grow exponentially and require more computation over time. Recommender systems will also grow exponentially and will require more Fast Memory Capacity over time. Large language models and Recommender systems are the two most important AI models today, and they have diverging computing requirements.
Recommenders can scale to billions of users and billions of items. For example, every article, video, and social media item has a learned numerical representation called embedding. Each embedding table can be 10s of TBs of data that requires multiple GPUs to process. Processing Recommenders requires data parallelism in parts of the network, and model parallelism in other parts of the network, stressing every part of the computer. Grace Hopper is ideally suited for recommender systems.
Whereas Hopper HGX systems are rich in tensor processing, each HGX node, with 8 GPUs, is limited to 640 GB of Fast Memory. On the other hand, a single Grace Hopper Superchip can be configured for 580 GB of Fast Memory. 500 GB connected to Grace. And 80 GB connected to Hopper. The two chips are connected by a superfast chip-to-chip NVlink Eight Grace Hoppers have the same tensor-throughput of Hopper HGX, but over 7X more Fast Memory capacity. Now with just a 120-node Grace Hopper system, you can process a 70 TB state-of-the-art recommender system. Grace Hopper will be a giant leap for recommender systems.
The amazing Grace CPU is the magic behind Grace Hopper. Let me tell you about it. Grace uses the Arm Neoverse V2 CPU core that delivers high single-threaded performance and twice the energy efficiency of server CPUs. 72 CPU cores are connected by a Scalable Coherency Fabric that delivers 3.2TB/s of bisectional bandwidth. 117 MB of L3 cache are distributed across the mesh fabric.
Grace system memory is LPDDR5X, which delivers 1.5X the bandwidth of DDR5, but most importantly at 1/8 the power. And finally, Grace is connected to Hopper through a chip-to-chip NVLink that gives Hopper high-speed access to Grace's large memory. Grace and Grace Hopper Superchips are designed for high performance computing applications that process a giant amount of data, like data analytics and recommender systems. NVIDIA and the world's leading system makers are building a broad array of servers. NVIDIA's bringing the full stack of our HPC, AI, and Omniverse platforms to Grace and Grace Hopper. Systems will be available in the first half of 2023.
Artificial intelligence is the most important technology force of our time, and is creating capabilities never before possible. In just a few years, AI has delivered breakthroughs in computer vision, speech recognition, natural language understanding, and recommender systems. These AI models still have to be built into applications— like self-driving cars, customer service chatbots, and e-commerce sites— and workflows that adapt the AIs to specific use-cases and help the models learn from experience. To make AI accessible to industries, we've developed domain-specific frameworks that companies can use to make AI applications.
All of these application frameworks are built on NVIDIA HPC, NVIDIA AI, and NVIDIA Omniverse platforms. Let me give you a few examples. NVIDIA MONAI is the most popular medical imaging AI framework. 15 of the top 20 academic medical centers and pharmaceutical companies, like Bayer and GSK, have adopted MONAI for radiology and pathology. NVIDIA Merlin is a framework that helps you easily build an end-to-end recommender system that is performant and scalable.
Companies including Capital One, Tencent WeChat, and Fidelity are building recommenders to improve the quality and engagement of their services. NVIDIA cuOpt is an Operations Research optimization API using AI to help developers create complex, real-time fleet routing. AT&T, BMW, and Domino's are using CuOpt. BMW has said that using cuOpt has saved 20% on their factory fleet orchestration & planning—millions of dollars. NVIDIA Morpheus is a GPU-accelerated SDK enabling cybersecurity developers to build AI pipelines that detect anomalies by filtering, processing, and classifying giant volumes of real-time data.
We're partnering with Booz Allen Hamilton to bring a Morpheus-powered cybersecurity platform to the public and private sectors. Artificial intelligence is impacting a large number of application domains and requires two basic approaches to computing. We are helping customers develop and deploy these AI applications in secure multi-cloud environments, whether it's scale-up batch processing AI factories, or scale-out interactive cloud services. For scale-up AI factories, we are in full production with Hopper.
For scale-out inference, interactive Metaverse applications, and Omniverse simulations, we have the OVX computer. In Omniverse simulation mode, one session runs on multiple GPUs. In inference and multimedia modes, multiple sessions can run on one GPU. Today, we're announcing the OVX computer with the new Ada Lovelace L40 datacenter GPU. With a massive 48 GB framebuffer, OVX with eight L40s will be able to process giant Omniverse virtual world simulations.
L40 OVX computers will be available from the world's leading computer makers and cloud providers. L40 GPUs are in full production. 80% of the world's internet traffic is video. User-generated content, like video conferencing or short-form videos, makes up a lot of it. Increasingly, these video streams will also be augmented by AI special effects and computer graphics.
And emerging metaverse applications, like avatars, will do computer vision, speech AI, language understanding, and computer graphics in real-time and at cloud scale. Nothing like this has ever been done before. To make this possible, we've been building acceleration libraries like CV-CUDA, pre-trained AI models in our Avatar Cloud Engine like Riva and Audio2Face, and a zero-code application framework and cloud runtime engine called UCF, universal computing framework.
Using UCF, we created a reference application called Tokkio for building customer service avatars. All of this is running on AWS. Let me show it to you. Interactive avatars are growing in popularity and use, from providing customer and sales support to marketing tools and teaching aids. But creating interactive avatars is very hard and typically requires expensive equipment, specific expertise, and time-consuming workflows.
NVIDIA ACE is a collection of cloud native AI-microservices that make it easier to create, customize, and deploy intelligent and engaging avatars built on any engine and deployable on any cloud. An interactive avatar needs to understand and communicate with you. From the more common text-driven chatbots that help users with site- or subject-specific queries to real-time fully-animated 3D avatars that can see and hear you.
And ACE has all the necessary AI building blocks to bring them to life. - Hi, I am Violet. How can I