GTC 2019 Keynote (San Jose)

Category: Main Keynote · Year: 2019 · ▶ Watch

Speakers: Eyal Waldman - CEO, Mellanox Technologies · Jensen Huang - CEO, NVIDIA · Jensen Huang, CEO, NVIDIA · Matt Garman - Vice President of Compute Services, Amazon Web Services · Microsoft Engineer - Microsoft · OmniSci Engineer - OmniSci

Switch language → 中文

Segments (30)

  • 00:00 · I am AI Intro
    • A video montage showcasing the diverse applications of NVIDIA AI across various industries.
  • 02:48 · Accelerated Computing Growth
    • Jensen Huang highlights the massive growth in NVIDIA’s developer ecosystem and CUDA downloads.
  • 06:15 · Introducing CUDA-X and PRADA
    • NVIDIA consolidates its acceleration libraries into CUDA-X, driven by the PRADA philosophy.
  • 15:05 · Chapter 1: Graphics and Real-Time Ray Tracing
    • A showcase of real-time ray tracing capabilities using RTX hardware and the Unity engine.
  • 21:00 · Turing Architecture
    • Deep dive into the Turing GPU architecture, featuring RT cores, Tensor cores, and advanced shading techniques.
  • 27:00 · RTX in Gaming: Dragonhound and Quake II
    • Demonstrations of RTX technology enhancing visual fidelity in modern and classic games.
  • 34:50 · RTX for Offline Rendering
    • How RTX servers drastically reduce the time and cost of offline rendering for film and design.
  • 41:15 · Announcing NVIDIA Omniverse
    • Introduction of an open collaboration platform that connects various 3D design tools in real-time.
  • 49:00 · GeForce NOW Cloud Gaming
    • A brief mention of NVIDIA’s cloud gaming service bringing high-end graphics to all PC users.
  • 50:00 · Cloud Gaming Partnerships
    • Jensen discusses cloud gaming and announces the GeForce NOW alliance with Softbank and LGU+.
  • 54:00 · RTX Server and Pod
    • Introduction of the RTX Server and RTX Server Pod for datacenter graphics and cloud gaming.
  • 56:25 · Project Sol Demo
    • A real-time cinematic rendering demo showcasing the power of RTX.
  • 60:45 · Chapter 2: AI & HPC
    • Focus shifts to data science, introducing the CUDA-X AI ecosystem and RAPIDS partnerships.
  • 73:55 · AI Conversational Search
    • Discussion on the complexity of conversational AI, followed by a live demo of Microsoft Bing.
  • 84:35 · Clara AI Toolkit
    • Announcement of the Clara AI Toolkit to accelerate AI applications in radiology.
  • 87:30 · Data Science and OmniSci Demo
    • Highlighting data science as the new HPC challenge, featuring a live data analytics demo with OmniSci.
  • 1:40:00 · Data Analytics and RAPIDS
    • Jensen discusses the RAPIDS open-source framework and its impact on accelerating data science workflows.
  • 1:42:40 · Data Science Workstations
    • Introduction of new workstations optimized for the growing field of data science.
  • 1:45:30 · Data Science - The New HPC Challenge
    • Comparing supercomputing, hyperscale, and the unique computational demands of data science.
  • 1:56:25 · Mellanox Partnership
    • Eyal Waldman joins to discuss the importance of networking and the NVIDIA-Mellanox acquisition.
  • 2:01:50 · Enterprise Data Science Solutions
    • Announcing DGX Pods and T4-powered enterprise servers for scalable data science.
  • 2:06:40 · AWS Partnership and G4 Instances
    • Matt Garman discusses ML on AWS, customer success stories, and new T4-powered G4 instances.
  • 2:13:30 · Robotics and Jetson Nano
    • Introduction of the $99 Jetson Nano computer and the Isaac robotics platform.
  • 2:21:15 · Autonomous Vehicles and DRIVE AP2X
    • Updates on the DRIVE platform, including Release 9.0 and the Safety Force Field path planning.
  • 150:00 · DRIVE AV Safety Force Field
    • Jensen Huang announces DRIVE AV Safety Force Field for path planning and prediction.
  • 151:45 · DRIVE Constellation
    • Announcement of the availability of DRIVE Constellation, a virtual AV test fleet.
  • 152:30 · DRIVE Constellation Demo
    • A demonstration of the DRIVE Constellation simulator showing various driving conditions.
  • 156:20 · Autonomous Driving Demo
    • Video demonstration of an autonomous vehicle driving on a highway with perception overlays.
  • 157:55 · Toyota Partnership
    • Announcement of a partnership with Toyota TRI-AD for autonomous vehicles.
  • 158:25 · Keynote Summary
    • Jensen Huang summarizes the key announcements from the entire keynote.

Product Announcements (20)

  • [06:15] CUDA-X
    • A unified suite of GPU-accelerated computing libraries.
    • specs: Combines domain-specific libraries (RTX, HPC, AI, DR, IS, CL, ME) into a single stack compatible with all NVIDIA GPUs.
    • availability: Available now via NGC.
  • [20:30] Unity RTX Experimental Package
    • Real-time ray tracing support for the Unity game engine.
    • specs: Enables developers to build physically accurate lighting and reflections in real-time.
    • availability: Available April 4th.
  • [26:00] New Turing GPUs and Laptops
    • Expansion of the Turing architecture lineup.
    • specs: Includes lower-tier GPUs without RT cores and 40 new gaming laptop models.
    • availability: Starting at $219 for the entry-level GPU.
  • [30:50] Quake II RTX
    • A fully path-traced version of the classic game Quake II.
    • specs: Features dynamic time of day, volumetric lighting, and physically based materials.
    • availability: To be open-sourced soon.
  • [41:15] NVIDIA Omniverse
    • An open collaboration platform for 3D production pipelines.
    • specs: Uses Universal Scene Description (USD) and Material Definition Language (MDL) to sync tools like Maya, Unreal, and Substance in real-time.
    • availability: Early access available at developer.nvidia.com/nvidia-omniverse.
  • [54:05] RTX Server
    • Datacenter graphics server design
    • specs: 40 Turing GPUs in 8U
    • availability: N/A
  • [55:00] RTX Server Pod
    • Modular design for enterprise and cloud edge datacenters
    • specs: 32 RTX servers, 1280 GPUs in 10 racks, up to 10,000 concurrent users
    • availability: N/A
  • [68:45] CUDA-X AI
    • End-to-end AI and data science ecosystem
    • specs: Integrates RAPIDS, TensorRT, cuDNN, and other libraries
    • availability: N/A
  • [84:35] Clara AI Toolkit
    • Toolkit for building and deploying AI applications for radiology
    • specs: AI-assisted annotation, transfer learning, AI deployment
    • availability: Available at developer.nvidia.com/clara
  • [1:43:30] Data Science Workstation
    • A workstation optimized for data scientists.
    • specs: Powered by NVIDIA GPU and CUDA-X AI, Dual Quadro RTX 8000 with 96 GB Memory.
    • availability: Available from top computer makers (Dell, HP, Lenovo).
  • [1:50:50] NVIDIA T4 Tensor Core GPU
    • A GPU designed for scale-out enterprise servers.
    • specs: 70 watts, fits in standard servers, 4x T4 provides ~260 Teraflops FP16.
  • [2:01:50] NVIDIA DGX Pod
    • A hyper-converged infrastructure reference architecture for AI.
    • specs: Integrates compute, storage, and networking from partners like DDN, Dell EMC, NetApp, Pure Storage, Arista, Cisco, Mellanox.
    • availability: Can be installed in 1 day.
  • [2:02:50] T4 Enterprise Servers
    • Enterprise servers optimized for data science.
    • specs: Powered by NVIDIA T4 and CUDA-X AI, NGC Certified.
    • availability: Available from Cisco, Dell EMC, Fujitsu, HPE, Inspur, Lenovo, Sugon.
  • [2:12:10] Amazon EC2 G4 Instances
    • New cloud instances on AWS.
    • specs: Featuring NVIDIA T4 Tensor Core GPUs, designed for ML inference, graphics, and video transcoding.
    • availability: Coming soon.
  • [2:14:30] Jetson Nano
    • A small, low-cost AI computer for robotics and edge devices.
    • specs: CUDA-X acceleration stack, high-resolution sensor support, runs all CUDA-X AI models.
    • availability: $99
  • [2:16:20] Isaac Open SDK
    • A software development kit for robotics.
    • specs: Includes Isaac Robot Engine, Isaac Sim, and Isaac Gym.
    • availability: Available at developer.nvidia.com/isaac-sdk
  • [2:23:00] DRIVE AP2X Release 9.0
    • High function L2+ autopilot system software.
    • specs: On-ramp to off-ramp, surround perception, localization to HD maps, real-time mapping.
  • [2:27:50] Safety Force Field
    • A computational method for autonomous vehicle path planning.
    • specs: Mathematically verifiable, designed to computationally avoid causing harm.
    • availability: Open platform.
  • [150:00] DRIVE AV Safety Force Field
    • Path planning and prediction software for autonomous vehicles.
    • specs: Designed for AV safety and drive comfort, mathematically verifiable, validated in simulation, open platform.
    • availability: Announced
  • [151:45] DRIVE Constellation
    • Virtual AV Test Fleet simulator.
    • specs: Bit-accurate, hardware-in-the-loop simulator, tests corner and rare conditions, cloud-based workflow.
    • availability: Available Now

Specific Numbers (17)

Timestamp Metric Value Context
03:30 NVIDIA Developers 1.8 Million A 50% growth from the previous year.
03:30 CUDA Downloads 12 Million A 50% growth from the previous year.
05:30 Performance Increase 40x Improvement in accelerated computing performance from 2010 to 2019.
21:00 Turing Transistors 18 Billion Transistor count on the high-end Turing RTX architecture.
21:00 Tensor Core Performance 130 TFLOPS Compute capability of the Tensor cores on Turing.
34:50 3D Creators 9 Million The addressable market of creators getting access to RTX in 2019.
38:30 Render Cost Savings $220,000 The cost difference between a 25-node CPU render farm ($250k) and a 1-node RTX Server ($30k).
54:05 GPUs per server 40 Number of Turing GPUs in an 8U RTX Server
55:00 GPUs per pod 1280 Number of GPUs in a 10-rack RTX Server Pod
55:00 Concurrent users 10,000 Supported users per RTX Pod
83:50 TensorRT Downloads 300k Growth from 50k in 2017 to 300k in 2018 (6X increase)
1:41:15 Query Time 4 minutes Reduced from 8 days using the new accelerated platform.
1:42:40 Data Scientists 3 million Estimated number of data scientists worldwide.
1:47:45 Supercomputer Compute Load 1 billion petaflops Measured in instances of compute, not time.
1:50:50 T4 Power Consumption 70 watts Power draw of the T4 GPU.
2:13:40 Jetson Developers 200,000 Number of developers using the Jetson platform.
2:14:30 Price $99 Cost of the new Jetson Nano computer.

Benchmark Claims (5)

  • [05:30] Accelerated Computing Performance: 40x
    • vs: 2010 baseline performance
    • gain: 40x improvement over 9 years through full-stack optimization.
  • [38:30] Offline Rendering (Incredibles 2): 6 Hours (1 RTX Server)
    • vs: 38 Hours (25 Dual Skylake CPU nodes)
    • gain: Over 6x faster rendering at a fraction of the hardware and power cost.
  • [1:44:30] End-to-End Data Science Workflow: Significantly faster
    • vs: CPU
    • gain: Visual chart shows a massive reduction in time for data prep, training, and end-to-end tasks.
  • [2:04:25] Acceleration of Data Science Clusters (End to End): 3 minutes
    • vs: 10x CPU Nodes (35 minutes)
    • gain: Over 10x faster using 10x T4 Nodes.
  • [2:05:50] Deep Learning Scaling (ResNet-50): Linear scaling
    • vs: Standard Ethernet
    • gain: Using RoCE RDMA prevents the performance plateau seen with standard networking as server count increases.

Customer Stories (9)

  • [37:20] Image Engine
    • Utilized NVIDIA rendering technology for complex visual effects.
    • outcome: Achieved high-fidelity, photorealistic rendering for film and television production.
  • [38:30] Pixar
    • Benchmarked rendering a frame of Incredibles 2 on CPU nodes versus an RTX Server.
    • outcome: Reduced render time from 38 hours to 6 hours and cut hardware costs drastically.
  • [53:30] Softbank and LGU+
    • Partnered with NVIDIA for cloud gaming
    • outcome: Announcing GeForce NOW alliance
  • [77:25] Microsoft Bing
    • Implemented AI conversational search using NVIDIA technology
    • outcome: Improved search accuracy and user experience with complex, multi-part queries
  • [86:50] MGH, NIH, OSU, DKFZ
    • Adopted Clara AI for radiology workflows
    • outcome: Reduced annotation time from hours to minutes, deployed clinical models in less than 24 hours
  • [2:09:50] Western Digital
    • Used P3s and Voltas on AWS for material science properties and magnetic/heat flows.
    • outcome: Improved the quality of their disk drives.
  • [2:10:15] Celgene
    • Used AI for drug design, moving from an on-premise cluster to AWS.
    • outcome: Reduced processing time from 2 months to 6 hours.
  • [2:11:10] Lyft
    • Ran all 50 million monthly rides on AWS, using AI/ML on P3s and Voltas with SageMaker.
    • outcome: Calculated fares, optimized drop-off/pick-up, and performed fraud detection.
  • [157:55] Toyota (TRI-AD)
    • Partnered with NVIDIA to create the future of autonomous vehicles.
    • outcome: Collaboration on AV Core Systems, Driving Simulation, In-Car Computer, and AI for AV.

Key Technologies (16)

  • CUDA-X: A unified collection of GPU-accelerated libraries for AI, HPC, and graphics.
  • RTX (Real-Time Ray Tracing): Hardware-accelerated calculation of light paths to create photorealistic reflections, shadows, and global illumination.
  • Turing Architecture: GPU architecture featuring dedicated RT Cores for ray tracing and Tensor Cores for AI processing.
  • Variable Rate Shading (VRS): Optimizes rendering performance by varying the shading rate across different areas of a frame.
  • Mesh Shaders: A new geometry pipeline that allows for highly complex and detailed scenes.
  • Omniverse: A platform that enables real-time, multi-tool collaboration for 3D content creation using USD.
  • RTX Server: Provides scalable, high-performance graphics and compute for datacenters.
  • CUDA-X AI: A comprehensive software acceleration library for AI and data science.
  • Clara AI: A platform for developing and deploying AI in medical imaging.
  • RAPIDS: An open-source machine learning and data science framework.
  • CUDA-X: The underlying engine and acceleration stack for NVIDIA’s AI and data science platforms.
  • InfiniBand and Ethernet (Mellanox): High-speed, low-latency networking protocols for data centers.
  • RoCE RDMA: RDMA over Converged Ethernet, used to improve network efficiency and scaling for deep learning.
  • Safety Force Field: A computational path-planning algorithm for autonomous vehicles to ensure safety.
  • Safety Force Field: Provides path planning and prediction to ensure autonomous vehicle safety.
  • Hardware-in-the-loop simulation: Allows bit-accurate testing of autonomous vehicle software in a virtual environment.

Demos Shown (13)

  • [15:40] A high-fidelity rendering of BMW cars to demonstrate photorealism.
    • True
  • [19:05] Live manipulation of a BMW interior and exterior using real-time ray tracing in Unity.
    • True
  • [27:10] Gameplay footage of Nexon’s Dragonhound toggling RTX reflections and shadows on and off.
    • True
  • [30:55] A fully path-traced version of Quake II showing dynamic lighting, glass refraction, and volumetric effects.
    • True
  • [37:20] A complex visual effects rendering sequence by Image Engine.
    • True
  • [45:35] Live collaboration in NVIDIA Omniverse showing updates in Maya, Unreal Engine, and Substance Painter reflecting instantly in a shared viewer.
    • True
  • [56:25] Project Sol cinematic rendering
    • Yes
  • [77:25] Microsoft Bing conversational search on a mobile device
    • Yes
  • [89:15] OmniSci analyzing and visualizing WiFi access point data
    • Yes
  • [2:18:50] A small, green robot named Kaya powered by Jetson Nano driving on stage.
    • Yes
  • [2:24:20] Video demonstration of DRIVE AP2X features including MyRoute, WaitNet, Sensor Fusion, and DRIVE Sim.
    • Yes (pre-recorded video)
  • [152:30] DRIVE Constellation simulation showing multiple camera views of a car driving through different weather and lighting conditions.
    • True
  • [156:20] Video of a real car driving autonomously on a highway, alongside visualizations of the car’s perception and planning systems.
    • True

Predictions / Commitments (4)

  • [11:15, Future] We believe that in the future, they will all be high-performance computing customers.
  • [36:45, End of 2019] By the end of this year, we should have all of them [major design tools] in production with RTX.
  • [1:55:10, Current/Ongoing trend] East-West networking traffic in data centers is going up exponentially.
  • [1:55:55, Future] The way you design a data center is going to change; networking and compute will become one continuous computing fabric.

Companies Mentioned (22)

Unity · Microsoft · Epic Games · Nexon · Intel · Autodesk · Softbank · LGU+ · Databricks · Google Cloud · Microsoft Azure · Accenture · ONNX Runtime · OmniSci · Dell, HP, Lenovo · Mellanox Technologies · DDN, Dell EMC, NetApp, Pure Storage, Arista, Cisco · Cisco, Dell EMC, Fujitsu, HPE, Inspur, Lenovo, Sugon · Amazon Web Services (AWS) · Toyota · AWS · Mellanox

Notable Quotes (6)

PRADA stands for PRogrammable Acceleration of multiple Domains with one Architecture. — Jensen Huang @ 14:00

Which one is real? Left or right? … This is not real. … That one is real. — Jensen Huang @ 16:45

The more you buy, the more you save… I think I was wrong. RTX servers are free. — Jensen Huang @ 39:50

Data Science is the new HPC. — Jensen Huang @ 1:45:30

The network is going to become really, really important. — Jensen Huang @ 1:56:10

We have more machine learning is done in AWS in the cloud than anywhere. — Matt Garman @ 2:09:40

Key Topics

Accelerated Computing · CUDA-X · Real-Time Ray Tracing · NVIDIA RTX · Turing Architecture · Unity Engine · Unreal Engine · Offline Rendering · NVIDIA Omniverse · 3D Collaboration · Path Tracing · Cloud Gaming · Cloud Gaming · RTX Server · CUDA-X AI · Data Science · RAPIDS · Conversational AI · Medical Imaging · Clara AI · Data Analytics · Data Science · RAPIDS · High Performance Computing (HPC) · Workstations · Networking · Mellanox · Enterprise Servers · Cloud Computing · AWS · Robotics · Jetson Nano · Isaac SDK · Autonomous Vehicles · DRIVE AP2X · Autonomous Vehicles · DRIVE AV · Safety Force Field · Path Planning · DRIVE Constellation · Simulation · Hardware-in-the-loop · Toyota Partnership · Accelerated Computing · CUDA-X

Takeaways

  • NVIDIA’s developer ecosystem is experiencing massive growth, driven by the adoption of accelerated computing.
  • CUDA-X unifies NVIDIA’s software libraries, providing a programmable architecture across multiple domains (PRADA).
  • Real-time ray tracing is now a reality, with major game engines like Unity and Unreal integrating RTX support.
  • The Turing architecture represents a massive leap in graphics, utilizing dedicated RT and Tensor cores to handle complex lighting and AI tasks.
  • RTX Servers offer disruptive cost and time savings for offline rendering, making traditional CPU render farms obsolete.
  • NVIDIA Omniverse aims to revolutionize 3D production pipelines by enabling seamless, real-time collaboration across disparate design tools globally.
  • NVIDIA is expanding its RTX technology from desktop gaming to cloud servers and enterprise pods.
  • CUDA-X AI consolidates NVIDIA’s software libraries to accelerate data science and AI workflows.
  • Strong partnerships with major cloud providers and integrators are driving the adoption of RAPIDS.
  • AI is significantly enhancing conversational search capabilities, as demonstrated by Microsoft Bing.
  • The Clara AI Toolkit aims to accelerate AI adoption and model deployment in the radiology field.
  • GPU-accelerated data analytics platforms like OmniSci provide real-time insights into massive datasets.
  • NVIDIA is heavily targeting the data science market with optimized hardware (workstations, servers) and software (RAPIDS).
  • Data science workloads require a new architecture that bridges the gap between traditional supercomputing and hyperscale cloud.
  • The acquisition of Mellanox highlights NVIDIA’s belief that high-speed, low-latency networking is critical for scaling AI and data science.
  • NVIDIA is partnering broadly across hardware OEMs, storage providers, and cloud providers (like AWS) to deliver end-to-end AI solutions.
  • The $99 Jetson Nano aims to democratize AI robotics development for makers, students, and edge applications.
  • NVIDIA’s autonomous vehicle strategy includes full-stack software (DRIVE AP2X) and verifiable safety models (Safety Force Field).
  • NVIDIA introduced DRIVE AV Safety Force Field to provide mathematically verifiable path planning for autonomous vehicles.
  • DRIVE Constellation is now available, offering a cloud-based, hardware-in-the-loop virtual testing fleet for AVs.
  • NVIDIA demonstrated the capabilities of DRIVE Constellation to simulate complex and rare driving conditions.
  • A major partnership with Toyota TRI-AD was announced to develop the future of autonomous vehicles.
  • The keynote concluded with a summary of NVIDIA’s full-stack approach, spanning from RTX graphics to data science and autonomous machines.