GTC Sept 2022 Keynote
Category: Fall Keynote · Year: 2022 · ▶ Watch
Speakers: Jensen Huang - CEO, NVIDIA
Segments (17)
- 00:00 · Introduction
- Jensen Huang introduces the keynote focusing on RTX, AI, and Omniverse.
- 01:00 · Racer RTX Demo
- A showcase of a fully interactive, physically simulated, and ray-traced environment built in Omniverse.
- 04:30 · Ada Lovelace Architecture
- Announcement of the 3rd generation RTX architecture, Ada Lovelace, featuring new RT and Tensor cores.
- 08:40 · DLSS 3
- Introduction of DLSS 3, which uses AI to generate entirely new frames, boosting performance significantly.
- 13:40 · Portal with RTX & RTX Remix
- Unveiling of a ray-traced remaster of Portal and the RTX Remix modding platform.
- 17:35 · GeForce RTX 40 Series
- Official announcement and pricing for the GeForce RTX 4090 and RTX 4080 GPUs.
- 21:55 · NVIDIA Omniverse
- Updates on the Omniverse platform, its ecosystem, and real-world digital twin applications.
- 32:50 · Omniverse Cloud
- Launch of Omniverse Cloud services and the Graphics Delivery Network (GDN).
- 37:10 · Autonomous Vehicles & DRIVE Thor
- Introduction of the DRIVE Thor centralized car computer, replacing Atlan.
- 49:00 · Robotics & Edge AI
- Announcements for Jetson Orin Nano, IGX Orin for industrial/medical, and updates to Isaac Sim.
- 57:25 · Accelerated Computing Software
- Updates on RAPIDS, Triton Inference Server, and the new CV-CUDA library.
- 01:05:00 · Quantum Computing
- Overview of cuQuantum and the QODA platform for hybrid quantum-classical computing.
- 01:08:45 · Large Language Models
- Introduction of NeMo LLM and BioNeMo cloud services for customizing large language models.
- 01:15:40 · Grace Hopper Superchip
- Details on the Grace Hopper architecture and its massive benefits for recommender systems.
- 01:24:20 · OVX & L40 GPU
- Announcement of the L40 GPU for data centers and the second-generation OVX systems.
- 01:25:40 · Omniverse ACE
- Demonstration of the Avatar Cloud Engine for building interactive, AI-driven digital humans.
- 01:33:00 · I am AI
- Closing video highlighting the diverse and transformative applications of AI.
Product Announcements (14)
- [05:15] Ada Lovelace Architecture
- NVIDIA’s 3rd generation RTX GPU architecture.
- specs: TSMC 4N process, 76 billion transistors, Shader Execution Reordering, 3rd-gen RT cores, 4th-gen Tensor cores.
- availability: N/A
- [08:40] DLSS 3
- AI-powered performance multiplier that generates entirely new frames.
- specs: Uses Optical Flow Accelerator, motion vectors, and a convolutional autoencoder.
- availability: N/A
- [13:40] Portal with RTX
- A free DLC reimagining the classic game with full ray tracing.
- specs: Built using Omniverse and RTX Remix.
- availability: November
- [15:30] RTX Remix
- An Omniverse application for modding classic games with RTX features.
- specs: Captures game assets into USD, uses AI to upres textures and convert materials.
- availability: Shortly after Ada launch
- [17:35] GeForce RTX 4090
- Flagship Ada Lovelace consumer GPU.
- specs: 24GB G6X memory, 2-4x faster than 3090 Ti.
- availability: $1,599, available October 12th
- [18:30] GeForce RTX 4080
- High-end Ada Lovelace consumer GPU.
- specs: Available in 16GB and 12GB G6X variants, 2-4x faster than 3080 Ti.
- availability: 16GB starting at $1,199, 12GB starting at $899, available in November
- [32:50] Omniverse Cloud
- Infrastructure-as-a-Service for designing, publishing, and experiencing metaverse applications.
- specs: Includes Omniverse Cloud, Replicator, and Farm. Powered by NVIDIA GDN.
- availability: N/A
- [38:10] NVIDIA DRIVE Thor
- Centralized car computer SoC replacing Atlan.
- specs: 2000 TOPS, integrates Hopper Transformer Engine, multi-domain computing isolation.
- availability: Starting in 2025 models
- [50:15] Jetson Orin Nano
- Compact robotics computer.
- specs: 80x faster than the previous Jetson Nano.
- availability: N/A
- [51:10] NVIDIA IGX Orin
- Industrial-grade edge AI platform.
- specs: Designed for industrial and medical environments requiring high safety and security.
- availability: N/A
- [01:02:30] CV-CUDA
- Open-source GPU-accelerated library for computer vision.
- specs: Provides highly performant CV kernels with zero-copy interfaces.
- availability: Early access in October
- [01:11:00] NeMo LLM Service
- Cloud service for customizing large language models.
- specs: Uses prompt learning to adapt foundation models to specific tasks.
- availability: Early access in October
- [01:13:10] BioNeMo LLM Service
- Digital biology framework for LLMs.
- specs: Understands chemicals, proteins, DNA, and RNA sequences.
- availability: Early access in October
- [01:24:45] NVIDIA L40 GPU
- Data center GPU based on Ada Lovelace.
- specs: 48GB frame buffer, designed for Omniverse and virtual world simulation.
- availability: In full production
Specific Numbers (13)
| Timestamp | Metric | Value | Context |
|---|---|---|---|
| 05:45 | Transistors | 76 Billion | Number of transistors in the Ada Lovelace GPU. |
| 05:50 | CUDA Cores | 18,000+ | Number of CUDA cores in Ada, 70% more than Ampere. |
| 06:00 | TFLOPS | 90 | Shader TFLOPS performance of the new Ada Streaming Multiprocessor. |
| 06:55 | PFLOPS | 1.4 | Tensor processing performance of the new Ada Tensor Core. |
| 17:35 | GB | 24 | Memory capacity of the RTX 4090. |
| 17:35 | $ | 1599 | Starting price of the RTX 4090. |
| 18:30 | $ | 1199 | Starting price of the RTX 4080 16GB. |
| 18:30 | $ | 899 | Starting price of the RTX 4080 12GB. |
| 38:10 | TOPS | 2000 | Performance of the DRIVE Thor SoC. |
| 01:20:30 | Cores | 72 | Number of Arm Neoverse V2 cores in the Grace CPU. |
| 01:20:40 | TB/s | 3.2 | Bandwidth of the scalable coherency fabric in the Grace CPU. |
| 01:20:50 | MB | 117 | L3 cache size in the Grace CPU. |
| 01:20:55 | GB/s | 500 | Memory bandwidth of the LPDDR5X memory in the Grace CPU. |
Benchmark Claims (6)
- [05:40] Power Efficiency: 2x
- vs: Ampere generation
- gain: Ada Lovelace delivers twice the power efficiency of the previous generation.
- [10:55] Game Performance (DLSS 3): Up to 4x
- vs: Brute force rendering
- gain: DLSS 3 boosts game performance by up to 4 times compared to native rendering.
- [17:40] Overall Performance: 2-4x faster
- vs: RTX 3090 Ti
- gain: The RTX 4090 is 2 to 4 times faster than the previous flagship.
- [38:20] Automotive Compute: 2x throughput, >2x performance
- vs: DRIVE Atlan
- gain: DRIVE Thor provides double the throughput and more than double the performance of the previously announced Atlan.
- [50:20] Robotics Compute: 80x faster
- vs: Jetson Nano
- gain: Jetson Orin Nano is 80 times faster than the original Jetson Nano.
- [01:16:00] Large Language Model Processing: 30x faster
- vs: Ampere
- gain: Hopper processes large language models up to 30 times faster than Ampere.
Customer Stories (7)
- [27:50] Heavy.AI and Charter Communications
- Used Omniverse to create digital twins of their 4G and 5G networks.
- outcome: Optimized placement and operations of cell towers, potentially saving billions.
- [28:40] Lowe’s
- Used Omniverse to design, build, and operate digital twins of their retail stores.
- outcome: Optimized operations, enhanced shopping experience, and allowed associates to explore designs in mixed reality.
- [29:15] Deutsche Bahn
- Built a digital twin of their national railway network in Omniverse.
- outcome: Expects to increase capacity and efficiency, and reduce carbon footprint without building new tracks.
- [34:00] Rimac
- Published an Omniverse-based 3D car configurator for the Nevera hypercar.
- outcome: Allowed customers to configure the car in full fidelity with real-time ray tracing via the cloud.
- [51:20] Siemens
- Adopted the IGX Orin platform for their industrial edge computing.
- outcome: Brought advanced AI and robotics capabilities to industrial automation.
- [01:14:50] Broad Institute
- Integrated NVIDIA Clara libraries (Parabricks) into their Terra cloud platform.
- outcome: Sped up whole genome sequencing from 24 hours to 1 hour, cutting compute costs by more than half.
- [01:23:30] Deloitte
- Partnered with NVIDIA to bring Morpheus cybersecurity and Omniverse services to enterprises.
- outcome: Enabled clients to build modern multi-cloud applications for security, automation, and digital twins.
Key Technologies (5)
- Shader Execution Reordering (SER): Reschedules shading workloads on the fly to better utilize GPU resources, significantly speeding up ray tracing.
- DLSS 3: Uses AI to generate entirely new frames, increasing frame rates without processing the graphics pipeline.
- NVIDIA Omniverse: A computing platform for building and operating metaverse applications and digital twins.
- Universal Scene Description (USD): The 3D internet protocol that connects virtual worlds within Omniverse.
- Transformer Engine: Hardware acceleration specifically designed to speed up the training and inference of Transformer-based AI models.
Demos Shown (8)
- [01:00] Racer RTX, a fully interactive, ray-traced physics simulation.
- True
- [11:35] Cyberpunk 2077 running with DLSS 3, showing significant frame rate improvements.
- True
- [12:35] Microsoft Flight Simulator running with DLSS 3.
- True
- [14:10] Portal with RTX, demonstrating full ray tracing in the classic game.
- True
- [15:40] RTX Remix being used to mod The Elder Scrolls III: Morrowind with enhanced assets and lighting.
- True
- [44:45] DRIVE Sim being used for concurrent physical and digital cabin design.
- True
- [55:05] Isaac Sim using cuOpt to optimize robot routing in a warehouse.
- True
- [01:26:10] Omniverse ACE powering interactive digital avatars (Violet and Tokyo).
- True
Predictions / Commitments (4)
- [01:45, Future] Future games will not have pre-baked worlds. Future games will be simulations.
- [27:30, Future] In the future, everything made will have a digital twin.
- [38:10, 2025] DRIVE Thor will be available for automakers’ 2025 models.
- [51:50, Future] Future medical instruments will be software-defined and powered by artificial intelligence.
Companies Mentioned (5)
TSMC · Microsoft · Valve · Siemens · Arm
Notable Quotes (4)
Computing is advancing at incredible speeds. The engine propelling this rocket is accelerated computing, and its fuel is AI. — Jensen Huang @ 00:10
Future games will not have pre-baked worlds. Future games will be simulations. — Jensen Huang @ 01:45
Omniverse is a platform for building and operating metaverse applications. — Jensen Huang @ 22:20
The more you buy, the more you save. — Jensen Huang @ 01:17:10
Key Topics
Accelerated Computing · Artificial Intelligence · Ray Tracing · Digital Twins · Metaverse · Omniverse · Autonomous Vehicles · Robotics · Large Language Models · Quantum Computing · Drug Discovery · Recommender Systems
Takeaways
- NVIDIA is heavily investing in full-stack simulation and digital twins via the Omniverse platform.
- The Ada Lovelace architecture provides massive leaps in ray tracing performance and introduces AI frame generation with DLSS 3.
- Digital twins are moving from concept to reality across major industries like retail, telecommunications, and manufacturing.
- DRIVE Thor represents a consolidation of vehicle computing, handling autonomous driving and in-cabin AI on a single chip.
- Large Language Models (LLMs) are a major focus, with NVIDIA launching cloud services to make customizing them accessible to enterprises.
- The Grace Hopper Superchip is specifically designed to overcome memory bottlenecks in massive AI workloads like recommender systems.