GTC November 2021 Keynote

Category: Fall Keynote · Year: 2021 · ▶ Watch

Speakers: Jensen Huang - CEO, NVIDIA

Switch language → 中文

Segments (15)

  • 00:00 · Pre-Show Countdown
    • A countdown timer with ambient music and partner logos before the keynote begins.
  • 13:58 · Introduction and ‘I am AI’
    • Jensen Huang introduces the keynote and presents the ‘I am AI’ video showcasing AI applications.
  • 17:49 · Accelerated Computing and CUDA
    • Discussion on the evolution of CUDA, full-stack optimization, and the expansion of accelerated computing.
  • 21:08 · NVIDIA ReOpt
    • Announcement of ReOpt, an accelerated solver for operations research and logistics optimization.
  • 22:35 · Quantum Computing
    • Introduction of the cuQuantum DGX Appliance for accelerating quantum circuit simulations.
  • 24:26 · Python Data Science
    • Announcement of cuNumeric to accelerate NumPy workflows for data scientists.
  • 25:40 · Data Center Scale Computing
    • Unveiling of the NVIDIA Quantum-2 InfiniBand platform and BlueField DOCA 1.2 for cloud-native supercomputing.
  • 27:30 · Cybersecurity with Morpheus
    • Introduction of Morpheus, a deep learning cybersecurity platform for detecting anomalies.
  • 29:29 · The Million-X Leap in Science
    • Exploring how physics-ML and AI are driving massive speedups in scientific simulations like drug discovery and climate modeling.
  • 37:39 · NVIDIA Modulus
    • Announcement of Modulus, a framework for developing physics-ML models for digital twins.
  • 39:30 · Omniverse and Digital Twins
    • Deep dive into Omniverse as the platform for creating 3D virtual worlds and digital twins of factories and cities.
  • 51:00 · AI Avatars and Project Tokkio
    • Showcasing conversational AI avatars using Riva, Maxine, and Omniverse, featuring ‘Toy Jensen’.
  • 61:23 · Edge AI and Robotics
    • Discussing the NVIDIA Unified Computing Framework (UCF), Metropolis, and Clara Holoscan for medical instruments.
  • 71:30 · Autonomous Vehicles
    • Updates on NVIDIA DRIVE, including Hyperion 8, DRIVE Sim Replicator, and DRIVE Concierge.
  • 81:40 · Earth-2 and Conclusion
    • Announcement of Earth-2, a digital twin of the Earth to simulate and predict climate change.

Product Announcements (15)

  • [21:08] NVIDIA ReOpt
    • An accelerated solver for operations research and logistics optimization.
    • specs: Massively parallel algorithm, dynamic rerouting, scales to 1,000s of locations.
    • availability: Available Now
  • [23:10] cuQuantum DGX Appliance
    • An appliance for researching quantum algorithms on classical supercomputers.
    • specs: Out-of-the-box optimized stack for Cirq, accelerates quantum simulators.
    • availability: Available Q1 2022
  • [24:58] NVIDIA cuNumeric
    • A drop-in accelerator for NumPy workflows.
    • specs: Zero code changes, automatic parallelism across multi-GPU and multi-node systems.
    • availability: Available Now on GitHub and Conda
  • [25:40] NVIDIA Quantum-2
    • A 400 Gbps InfiniBand networking platform for cloud-native supercomputing.
    • specs: 57 billion transistors, 64 ports at 400 Gbps, performance isolation, precision timing.
    • availability: Sampling Now
  • [26:08] BlueField DOCA 1.2
    • A zero-trust security framework for the BlueField DPU.
    • specs: Deep packet inspection, intrusion detection, telemetry, hardware authentication.
    • availability: Not explicitly stated
  • [27:40] NVIDIA Morpheus
    • An accelerated AI platform for next-gen SIEM and cybersecurity.
    • specs: 600x faster data processing, detects anomalies with 10s of millions of AI models in real-time.
    • availability: Early Access 2 Available Now
  • [37:39] NVIDIA Modulus
    • A physics-ML neural simulation framework.
    • specs: Trains physics-ML models using governing physics, multi-GPU/multi-node training, 100,000x speedups.
    • availability: Available Now
  • [44:20] Omniverse Showroom, Farm, AR, VR
    • New features and applications within the Omniverse platform.
    • specs: Showroom for demos, Farm for batch processing, AR/VR streaming.
    • availability: Available in Beta / VR Coming Soon
  • [62:09] NVIDIA Riva Speech AI
    • An SDK for building conversational AI applications.
    • specs: World-class quality, customizable voice, supports 7 languages, deployable anywhere.
    • availability: Enterprise Support Available Q1 ‘22 Globally
  • [71:00] Triton Inference Server 2.15
    • An updated inference server supporting multiple frameworks and platforms.
    • specs: Forest inferencing (ML and DL in one app), multi-GPU/multi-node inference.
    • availability: Available for Download Today
  • [85:18] NVIDIA Clara Holoscan
    • A computational sensing instrument platform for medical devices.
    • specs: Stream-computing platform, runs in data center or edge, powered by AGX Orin.
    • availability: Available November 15
  • [89:18] NVIDIA Isaac ROS GEMs
    • Hardware-accelerated packages for the ROS community.
    • specs: Accelerates ROS-native packages up to 10x, brings NVIDIA AI to ROS.
    • availability: Download Isaac ROS GEMs today
  • [91:08] Isaac Sim Replicator
    • A synthetic data generation engine for training robots.
    • specs: Generates automatically labeled data, domain randomization.
    • availability: Not explicitly stated
  • [93:20] DRIVE Hyperion 8
    • A functionally safe production-ready AV platform.
    • specs: Dual Orin SoCs, 12 cameras, 9 radars, 12 ultrasonics, 1 front lidar.
    • availability: Available Now
  • [100:00] DRIVE Concierge
    • An AI assistant for vehicles built on Omniverse Avatar.
    • specs: Always-on 3D avatar, natural language interaction, automated parking.
    • availability: Not explicitly stated

Specific Numbers (10)

Timestamp Metric Value Context
18:00 CUDA Downloads 30 Million Total CUDA downloads over the past 15 years.
18:00 SDKs 150 Number of SDKs available in the NVIDIA ecosystem.
21:20 Speedup 3 Seconds vs 5 Minutes ReOpt routing 1,000 packages compared to traditional methods.
23:50 Qubits 1,688 Record qubit scale simulated using cuQuantum on 896 GPUs.
26:00 Transistors 57 Billion Number of transistors in the Quantum-2 switch chip.
26:00 Bandwidth 400 Gbps Throughput per port on the Quantum-2 InfiniBand platform.
39:20 Speedup 100,000x Speedup achieved by Modulus for weather prediction compared to traditional simulation.
64:00 Parameters 530 Billion Size of the Megatron-Turing NLG model.
85:40 Transistors 17 Billion Number of transistors in the Orin SoC.
85:40 Performance 250 TOPS AI performance of the Orin SoC.

Benchmark Claims (3)

  • [21:20] Vehicle Routing: 3 Seconds
    • vs: 5 Minutes (traditional solver)
    • gain: 100x faster for routing 1,000 packages.
  • [23:50] Quantum Simulation (MaxCut): 1,688 Qubits
    • vs: Previous records
    • gain: Largest ever exact quantum circuit simulation.
  • [39:20] Weather Prediction (FNO): 0.25 Seconds
    • vs: 4 Hours on traditional simulation
    • gain: 100,000x speedup for a 7-day forecast.

Customer Stories (5)

  • [21:25] Domino’s
    • Used ReOpt for vehicle routing and logistics optimization.
    • outcome: Improved delivery efficiency for last-mile delivery.
  • [40:40] BMW Group
    • Built a digital twin of their Regensburg factory in Omniverse.
    • outcome: Optimized production scenarios, increased agility, and improved efficiency.
  • [49:00] Ericsson
    • Built a city-scale digital twin in Omniverse to simulate 5G network coverage.
    • outcome: Optimized antenna placement and network performance using physically accurate RF simulation.
  • [55:40] Siemens Energy
    • Developed a digital twin of a Heat Recovery Steam Generator (HRSG) using Modulus and Omniverse.
    • outcome: Predicted corrosion accurately, potentially reducing downtime by 70% and saving $1.7 billion annually.
  • [68:20] Microsoft
    • Integrated NVIDIA AI (Riva) into Microsoft Teams for live captioning.
    • outcome: Provided real-time transcription and translation in 28 languages for over 250 million users.

Key Technologies (5)

  • CUDA: General-purpose parallel computing platform and programming model.
  • cuQuantum: SDK for accelerating quantum circuit simulations on GPUs.
  • Physics-ML: Combines machine learning with physical laws to accelerate scientific simulations.
  • Omniverse: A platform for 3D design collaboration and building physically accurate digital twins.
  • Omniverse Avatar: A platform for creating interactive, AI-driven 3D avatars.

Demos Shown (5)

  • [21:35] ReOpt optimizing warehouse robot routing in an Omniverse simulation.
    • True
  • [40:40] BMW factory digital twin in Omniverse, showing robot planning and human ergonomics.
    • True
  • [41:00] Toy Jensen (Project Tokkio) answering questions about climate change and astronomy.
    • True
  • [63:20] Maxine translating speech into multiple languages with synchronized lip movements.
    • True
  • [76:00] DRIVE Hyperion 8 navigating urban streets and highways autonomously.
    • True

Predictions / Commitments (2)

  • [84:00, Future / Ongoing] We will build a digital twin of the Earth, Earth-2, to simulate and predict climate change.
  • [91:30, 2024] By 2024, the vast majority of new EVs will have substantial AV capability.

Companies Mentioned (6)

Ansys · Google · Check Point, F5, Fortinet, Juniper, Palo Alto Networks, VMware · BMW · Ericsson · Microsoft

Notable Quotes (4)

Accelerated computing starts with NVIDIA CUDA. — Jensen Huang @ 17:49

Omniverse is a foundational platform for digital twin virtual worlds. — Jensen Huang @ 39:30

The more you buy, the more you save. — Jensen Huang @ 73:15

We will build a digital twin of our Earth. — Jensen Huang @ 84:00

Key Topics

Accelerated Computing · CUDA · Quantum Computing · Data Center Infrastructure · Cybersecurity · Physics-Informed Machine Learning · Digital Twins · Omniverse · Conversational AI · Robotics · Autonomous Vehicles · Climate Science

Takeaways

  • NVIDIA is expanding its focus from GPUs to full-stack, data-center-scale computing with products like Quantum-2 and BlueField DPUs.
  • Omniverse is positioned as a critical platform for the future, acting as the ‘HTML of 3D worlds’ to build physically accurate digital twins for industries ranging from manufacturing to telecommunications.
  • Physics-ML is driving a ‘Million-X Leap’ in scientific simulation, enabling breakthroughs in areas like drug discovery and weather prediction that were previously computationally impossible.
  • AI is evolving from perception tasks to generative models and complex reasoning, as demonstrated by large language models (Megatron) and interactive avatars (Project Tokkio).
  • Robotics and autonomous vehicles are transitioning from R&D to deployment, supported by end-to-end platforms like Isaac, DRIVE Hyperion, and synthetic data generation via Omniverse Replicator.
  • NVIDIA is committing to tackling massive global challenges, culminating in the announcement of Earth-2, a digital twin of the planet designed to simulate and mitigate climate change.