GTC Spring 2021 Keynote

Category: Main Keynote · Year: 2021 · ▶ Watch

Speakers: Alex - Product Manager, NVIDIA Maxine · Dr. Milan Nedeljkovic - Member of the Board of Management, BMW AG Production · Jensen Huang - Founder & CEO, NVIDIA · Perry Nightingale - WPP

Switch language → 中文

Segments (15)

  • 16:06 · Introduction
    • Jensen Huang introduces the core themes: AI, accelerated computing, data centers, and the metaverse.
  • 21:35 · Computer Graphics & RTX
    • Showcasing the advancements in computer graphics driven by RTX and AI.
  • 24:39 · NVIDIA Omniverse
    • Introduction to Omniverse, a platform for connecting 3D worlds and creating digital twins.
  • 33:37 · BMW Digital Twin
    • BMW showcases their use of Omniverse to create a digital twin of their manufacturing factory.
  • 39:07 · The Data Center is the New Unit of Computing
    • Discussing the evolution of data centers and the need for DPUs to offload infrastructure tasks.
  • 44:30 · DGX and AI Supercomputing
    • Announcements of new DGX systems and software for training massive AI models.
  • 52:20 · Clara Discovery & Healthcare
    • Highlighting AI’s impact on drug discovery and genomics.
  • 55:47 · Quantum Computing
    • Introducing cuQuantum to accelerate quantum circuit simulations on GPUs.
  • 01:01:00 · NVIDIA Grace CPU
    • Unveiling the Grace CPU, an Arm-based processor designed for giant-scale AI and HPC.
  • 01:05:30 · Expanding the Arm Ecosystem
    • Partnerships with AWS, Ampere, Marvell, and MediaTek to bring GPUs to Arm platforms.
  • 01:07:02 · The Waves of AI
    • Outlining the progression of AI from cloud to enterprise, edge, and autonomous systems.
  • 01:14:13 · Morpheus & Cybersecurity
    • Introducing Morpheus for AI-driven, real-time network security.
  • 01:16:56 · NVIDIA AI Enterprise & Edge
    • Software suites and platforms for deploying AI in enterprise environments and at the edge.
  • 01:24:00 · Conversational AI & Recommenders
    • Showcasing Jarvis for conversational AI and Merlin for recommender systems.
  • 01:34:36 · NVIDIA DRIVE & Autonomous Vehicles
    • Announcing new hardware and simulation tools for the autonomous vehicle industry.

Product Announcements (16)

  • [41:55] DOCA 1.0
    • SDK for programming BlueField DPUs
    • specs: Deep packet inspection, secure boot, TLS crypto offload
    • availability: Available today
  • [42:38] BlueField-3 DPU
    • Next-generation Data Processing Unit
    • specs: 22 billion transistors, 400 Gbps networking, 16x Arm A78 cores
    • availability: Expected 2022
  • [45:48] DGX Station 320G
    • Workgroup AI supercomputer-in-a-box
    • specs: 4x A100 80GB GPUs, 320GB memory, 8 TB/sec bandwidth
    • availability: $149,000 or $9,000/month subscription
  • [47:43] DGX SuperPOD (A100 80GB)
    • Cloud-native AI supercomputer
    • specs: Upgraded with 80GB A100s and BlueField-2 DPUs
    • availability: Not specified
  • [50:28] Megatron Triton
    • Software for training giant transformer models
    • specs: Optimal multi-GPU and multi-node parallelism
    • availability: Not specified
  • [57:28] cuQuantum
    • Acceleration library for simulating quantum circuits
    • specs: Optimized for tensor network and state vector solvers
    • availability: Not specified
  • [01:01:00] NVIDIA Grace CPU
    • Arm-based CPU for giant-scale AI and HPC
    • specs: 2000 GB/sec memory-to-GPU bandwidth, 300 SPECint per CPU
    • availability: Available 2023
  • [01:11:00] Aerial A100
    • AI-on-5G computing platform
    • specs: Integrates Ampere GPU and BlueField DPU on one card
    • availability: Not specified
  • [01:14:13] NVIDIA Morpheus
    • AI cybersecurity framework
    • specs: Real-time all-packet inspection using AI
    • availability: Not specified
  • [01:17:28] NVIDIA AI Enterprise
    • Enterprise-grade AI software suite
    • specs: Certified for VMware vSphere
    • availability: Not specified
  • [01:20:09] NVIDIA TAO Framework
    • Train, Adapt, and Optimize framework for AI models
    • specs: Federated learning, transfer learning
    • availability: Not specified
  • [01:20:38] NVIDIA Fleet Command
    • Cloud-native platform for orchestrating AI at the edge
    • specs: Secure boot, remote management
    • availability: Not specified
  • [01:24:00] NVIDIA Jarvis
    • Conversational AI framework
    • specs: Speech recognition, language understanding, translation
    • availability: Available today
  • [01:37:58] Hyperion 8 AV Platform
    • Reference architecture for autonomous vehicles
    • specs: 8 cameras, 4 fisheyes, 9 radar, 2 lidar
    • availability: Not specified
  • [01:38:39] NVIDIA DRIVE Atlan
    • Next-generation AV SoC
    • specs: 1000 TOPS, 400 Gbps networking, integrates Grace CPU
    • availability: Targeting 2025 models
  • [01:42:20] DRIVE Sim powered by Omniverse
    • Simulation platform for autonomous vehicles
    • specs: Physically accurate, scalable, cloud-native
    • availability: Available this summer

Specific Numbers (9)

Timestamp Metric Value Context
18:30 CUDA GPUs 1 Billion Installed base of CUDA GPUs shipped.
18:30 ExaFLOPS 250 GPU computing power in the cloud.
18:30 Developers 2.5 Million Number of developers in the NVIDIA ecosystem.
41:00 Members 10 Million GeForce NOW members across 70 countries.
42:38 Transistors 22 Billion Transistor count on the BlueField-3 DPU.
48:50 Parameters 175 Billion Size of the GPT-3 language model.
01:01:20 GB/sec 2000 Memory-to-GPU bandwidth provided by the Grace CPU architecture.
01:03:20 Exaflops 20 AI performance of the upcoming Alps supercomputer.
01:38:39 TOPS 1000 Performance of the DRIVE Atlan SoC.

Benchmark Claims (4)

  • [19:40] NAMD Performance: 13x faster
    • vs: Performance 5 years ago
    • gain: 13-fold increase in performance due to full-stack optimization.
  • [58:05] Sycamore Quantum Circuit Simulation: < 10 minutes
    • vs: Days or years on traditional systems
    • gain: Simulated depth=20 circuit in record time using cuQuantum on Selene.
  • [01:01:40] 1 Trillion Parameter Model Training: 10x faster
    • vs: Today’s fastest servers
    • gain: Grace CPU architecture significantly reduces training time for massive models.
  • [01:25:00] Bilingual Evaluation Understudy (BLEU): 40 (EN-JP), 50 (EN-ES)
    • vs: Standard translation benchmarks
    • gain: High-quality, fluent translation capabilities in Jarvis.

Customer Stories (5)

  • [33:37] BMW
    • Used NVIDIA Omniverse to create a digital twin of their manufacturing facilities.
    • outcome: Achieved 30% more efficient planning processes.
  • [52:00] Naver
    • Adopted DGX SuperPOD to build a giant language model for Korean.
    • outcome: Creating advanced language understanding AI services.
  • [53:30] Oxford Nanopore
    • Used DGX to train models for DNA sequencing.
    • outcome: Achieved 99.9% detection accuracy of single nucleotide variants.
  • [55:12] Recursion
    • Built the BioHive-1 supercomputer using DGX SuperPOD.
    • outcome: Accelerating drug discovery by analyzing massive biological datasets.
  • [01:03:20] Swiss National Supercomputing Centre (CSCS)
    • Building the ‘Alps’ supercomputer powered by NVIDIA Grace CPUs.
    • outcome: Will deliver 20 Exaflops of AI performance for scientific research.

Key Technologies (5)

  • Omniverse: A platform for connecting 3D worlds and simulating physically accurate digital twins.
  • DPU (Data Processing Unit): Offloads and accelerates networking, storage, and security tasks from the CPU.
  • Transformer Models: Advanced neural network architecture used for natural language processing and other tasks.
  • Quantum Simulation: Using classical GPUs to simulate quantum circuits to advance quantum computing research.
  • Arm Architecture: Energy-efficient CPU architecture being adopted for high-performance computing and AI.

Demos Shown (6)

  • [22:49] Gameplay footage of Bright Memory, Black Myth: Wukong, and Death Stranding demonstrating RTX graphics.
    • True
  • [33:37] Digital twin of a BMW factory in Omniverse, showing human and robot simulation.
    • True
  • [01:14:50] Morpheus detecting leaked credentials in network traffic.
    • True
  • [01:26:45] Jarvis performing real-time speech recognition and translation.
    • True
  • [01:29:47] Maxine features including real-time translation, eye contact correction, and video compression.
    • True
  • [01:42:30] DRIVE Sim simulating a Mercedes-Benz vehicle navigating various environments.
    • True

Predictions / Commitments (4)

  • [43:10, 2024] BlueField-4 will have 64 billion transistors and 800 Gbps networking.
  • [48:50, By 2023] We expect to see 100 trillion parameter AI models.
  • [01:01:00, 2023] NVIDIA Grace CPU will be available.
  • [01:38:39, 2025 models] DRIVE Atlan will be targeted for production vehicles.

Companies Mentioned (9)

Bentley Systems · Hewlett Packard Enterprise (HPE) · AWS · Ampere Computing · Marvell · MediaTek · Google Cloud · VMware · Mercedes-Benz

Notable Quotes (3)

Software is writing software no human can. — Jensen Huang @ 16:20

The data center is the new unit of computing. — Jensen Huang @ 39:07

Three chips, yearly leaps, one architecture. — Jensen Huang @ 01:04:04

Key Topics

Accelerated Computing · Artificial Intelligence · Omniverse · Digital Twins · Data Centers · DPUs (Data Processing Units) · Transformer Models · Quantum Computing · Arm Architecture · Edge Computing · 5G · Cybersecurity · Conversational AI · Autonomous Vehicles

Takeaways

  • NVIDIA is positioning itself as a full-stack computing platform company, not just a GPU vendor.
  • Omniverse is a major strategic push to create industrial digital twins and the ‘metaverse’ for enterprise.
  • The data center architecture is shifting, requiring DPUs like BlueField to offload infrastructure tasks from CPUs.
  • AI models are growing exponentially, driving the need for massive systems like DGX SuperPODs and new software like Megatron.
  • NVIDIA is entering the CPU market with ‘Grace’, an Arm-based processor designed specifically for giant-scale AI.
  • NVIDIA is aggressively expanding the Arm ecosystem across cloud, edge, and PC markets.
  • AI is moving to the edge and 5G networks, supported by platforms like Aerial and Fleet Command.
  • Autonomous vehicle development relies heavily on simulation (DRIVE Sim) and massive compute power (DRIVE Atlan).