TLDR;
Jensen Huang's GTC keynote in Washington D.C. highlights NVIDIA's advancements in accelerated computing and AI, emphasizing their role in driving a new industrial revolution. Key announcements include partnerships with Nokia for 6G technology, advancements in quantum computing with NVQLink, and collaborations with the Department of Energy to build AI supercomputers. NVIDIA is focusing on extreme co-design to enhance AI model performance and reduce costs, showcasing the Grace Blackwell NVL72 as a significant achievement. The company is also re-industrializing America by manufacturing AI components domestically and introducing AI factories powered by Omniverse DSX.
- NVIDIA is pioneering a new computing model with GPUs and CUDA, essential for solving problems beyond traditional CPUs.
- AI is transforming industries, requiring a new computing stack and AI factories to produce valuable tokens.
- Extreme co-design and strategic partnerships are crucial for driving performance and reducing costs in AI infrastructure.
America, the Land of Innovation [0:00]
The video celebrates America's history of innovation, highlighting key inventions and technological advancements that have shaped the modern world. From the transistor at Bell Labs to the personal computer by Apple and the internet's foundation through ARPANET, these breakthroughs have propelled various industries and improved lives. NVIDIA GPUs, invented in America, are now at the heart of the AI revolution, serving as essential infrastructure for companies and nations. American innovators are driving progress in various fields, promising advancements in clean energy and space exploration.
Intro [4:29]
Jensen Huang, NVIDIA's founder and CEO, expresses his pride in America and introduces the topics to be covered at GTC, including industry trends, scientific advancements, computing innovations, and future technologies. He thanks NVIDIA's partners for sponsoring the event and acknowledges the significance of GTC as the "Super Bowl of AI." Huang highlights NVIDIA's invention of a new computing model, marking a significant shift in the industry.
GPU to CPU Accelerated Computing Shift [6:03]
NVIDIA invented a new computing model to address problems that general-purpose computers couldn't solve, anticipating the slowdown of transistor performance and power scaling. For 30 years, NVIDIA has advanced accelerated computing by inventing the GPU and CUDA programming model. Accelerated computing requires reinventing algorithms and libraries, making it a long-term endeavor.
CUDA-X Libraries [8:09]
CUDA-X libraries are essential for accelerated computing, offering redesigned algorithms that enable ecosystem partners to leverage its benefits. These libraries, including CU Litho for computational lithography and cuOpt for numerical optimization, have taken years to develop and are used by industry leaders like TSMC and ASML. The libraries cover various applications, such as sparse solvers for CAE, Warp Python solver for CUDA, and QDF for accelerating SQL data frame databases.
CUDA Opens New Markets [11:05]
CUDA-X libraries open new markets for NVIDIA by enabling simulations and applications across various industries, including healthcare, life sciences, manufacturing, robotics, autonomous vehicles, computer graphics, and video games. The simulations demonstrate the beauty of mathematics and deep computer science, showcasing NVIDIA's journey from its first application in 1993 to its current capabilities. Jensen Huang thanks NVIDIA employees for their contributions to this incredible journey.
Telecommunications [15:34]
Telecommunications is crucial for the economy and national security, but the U.S. has fallen behind in wireless technology. NVIDIA aims to change this by innovating with American technology during a fundamental platform shift. Computer technology is the foundation of every industry and the most important instrument of science.
Nokia to Build AI-Native 6G on New NVIDIA Arc [17:34]
NVIDIA partners with Nokia to build on new technology based on accelerated computing and AI, positioning the U.S. at the center of the next revolution in 6G. NVIDIA introduces the NVIDIA Arc (Aerial Radio Network Computer), built from the Grace CPU, Blackwell GPU, and ConnectX Melanox Connect X networking. Nokia will integrate NVIDIA Arc into their future base stations, upgrading millions of base stations worldwide with 6G and AI.
Quantum Computing [21:37]
In 1981, Richard Feman envisioned a quantum computer to simulate nature directly. Forty years later, a breakthrough: the creation of one logical cubit that is coherent, stable, and error-corrected. Quantum computers require connecting to GPU supercomputers for error correction, AI calibration, control, and simulations.
The Future of Quantum Computing [24:00]
Building a quantum computer involves using cubits, which are fragile and sensitive to noise. Quantum error correction is essential, requiring extra cubits to detect and correct errors without damaging the primary cubits. NVQLink, a new interconnect architecture, directly connects quantum processors with NVIDIA GPUs to facilitate quantum error correction.
Announcing NVIDIA QLink Quantum-GPU Interconnect [26:17]
NVIDIA announces MVQLink, enabling quantum computer control, calibration, error correction, and hybrid simulations by connecting QPUs and GPU supercomputers. It is scalable for future quantum computers with tens to hundreds of thousands of cubits. CUDA Q supports QPU, allowing computation to move between QPU and GPU in microseconds.
Department of Energy Partnering with NVIDIA to Build 7 New AI Supercomputers [28:40]
The Department of Energy (DOE) is partnering with NVIDIA to build seven new AI supercomputers to advance the nation's science. Computing is the fundamental instrument of science, undergoing platform shifts towards accelerated computing, AI, and quantum computing. These supercomputers will leverage AI, principled simulation, and quantum computing to enhance scientific research.
AI - The New Industrial Revolution [30:32]
AI is more than just a chatbot; it has reinvented the computing stack, shifting from hand-coded software on CPUs to machine learning on GPUs. This new stack requires significant energy and infrastructure to generate tokens, the computational units of AI. AI can tokenize various forms of data, enabling it to learn, translate, and generate content.
Three Scaling Laws [42:13]
AI models have become smarter through pre-training, post-training, and thinking, increasing the demand for computation. Smarter models lead to increased usage, creating a virtuous cycle where more intelligence requires more computation. AI models are now good enough to be paid for, leading to two exponentials: the compute requirement of the three scaling laws and the increasing demand from more users.
Extreme Co-Design [50:03]
Extreme co-design is essential to drive down costs and improve AI performance, involving the re-architecting of computer architecture, chips, systems, software, model architecture, and applications. NVIDIA scales up AI by creating entire rack-scale computers and scaling out with Spectrum Ethernet technology. This extreme co-design results in significant performance improvements.
Grace Blackwell NVL72 - "A Thinking Machine" [51:57]
The Grace Blackwell NVL72 is an extreme co-designed computer that integrates 72 GPUs into one giant fabric, allowing every expert AI model to communicate with each other. This system significantly improves speed and efficiency compared to previous models.
Extreme Co-Designed Blackwell NVL72 [53:30]
The Grace Blackwell NVL72 allows for more efficient processing by distributing experts across GPUs, resulting in a 10x performance increase compared to the H200. The Grace Blackwell MVLink72 generates the lowest cost tokens due to its high token generation capability and optimized total cost of ownership.
CSP Capex Spend [58:03]
The top six CSPs (Amazon, Corewave, Google, Meta, Microsoft, and Oracle) are investing heavily in capex, with the Grace Blackwell MVLink72 being in volume production to meet this demand. Two platform shifts are occurring: from general-purpose computing to accelerated computing and from classical software to AI. NVIDIA's GPU can handle both, making it a safe choice for investment.
Exceptionally Strong Demand for Grace Blackwell NVL72 [1:01:10]
NVIDIA is experiencing extraordinary growth for Grace Blackwell, driven by two exponentials. The company has visibility into half a trillion dollars of cumulative Blackwell and early Rubin ramps through 2026. Blackwell is growing five times faster than Hopper, with 20 million Blackwell GPUs in the early parts of Rubin.
Blackwell: Made in America, Made for the World [1:03:00]
The age of AI has begun, with Blackwell as its engine, manufactured in America. The manufacturing process involves hundreds of chip processing steps, HBM stack assembly, and the integration of Blackwell dyes and HBM stacks on a custom silicon interposer wafer. The assembly is baked, molded, and cured, creating the GB300 Blackwell Ultra Super Chip.
NVIDIA Extreme Co-Design Delivers X-Factors on One-Year Rhythm [1:06:48]
NVIDIA is manufacturing in America again, with full production of Blackwell in Arizona. Extreme co-design provides 10x generational improvements. The company is now working on multiple chips simultaneously to achieve these performance gains.
Vera Rubin Superchip, Vera Rubin Compute Tray, Vera Rubin CPX Compute Tray, BlueField-4, NVLink Switch Tray [1:08:20]
The Vera Rubin is NVIDIA's third-generation MVLink 72 rack-scale computer, featuring a cableless design. It delivers 100 petaflops, 100 times the performance of the DGX1 delivered to OpenAI nine years ago. The Vera Rubin compute tray includes a context processor for handling larger amounts of context for AI.
AI Factories [1:14:42]
NVIDIA is now designing entire AI factories, integrating more of the problem to solve and creating better solutions. This AI factory is being built for Vera Rubin, with technology that allows partners to integrate into the factory digitally.
Omniverse DSX [1:15:22]
NVIDIA Omniverse DSX is a blueprint for building and operating gigascale AI factories. It co-designs building, power, and cooling with NVIDIA's AI infrastructure stack. The digital twin acts as an operating system, with engineers using AI agents to optimize power consumption and reduce strain on the AI factory and the grid.
Open Models [1:19:08]
Open-source models have become capable due to reasoning capabilities, multimodality, and efficiency through distillation. They are essential for startups, researchers, and companies. NVIDIA is dedicated to open source, leading in contributions with 23 models in leaderboards across various domains.
NVIDIA is Everywhere [1:21:20]
AI startups build on NVIDIA due to its rich ecosystem, tools, and ubiquitous GPUs. NVIDIA integrates its libraries and models into AWS, Google Cloud, Microsoft Azure, and Oracle Cloud. The company also integrates NVIDIA libraries into world SAS, such as Service Now and SAP, to create agentic SAS.
Physical AI [1:26:30]
Physical AI requires three computers: one for training (Grace Blackwell Invink 72), one for simulation (Omniverse computer), and one for operation (Thor Jetson Thor robotics computer). These computers run CUDA, enabling advancements in physical AI.
America is Reindustrializing [1:28:51]
America is re-industrializing, with Foxcon building a robotic facility in Houston, Texas, for manufacturing NVIDIA AI infrastructure systems. The factory is born digital in Omniverse, with engineers assembling their virtual factory in a Seaman's digital twin solution.
Newton Simulation Platform [1:34:00]
NVIDIA is working with Disney research on a new framework and simulation platform called Newton, enabling robots to learn in a physically aware environment.
Humanoid Robotics [1:34:37]
Humanoid robotics is likely to be one of the largest consumer electronics and industrial equipment markets. NVIDIA is working with companies like Figure and Agility to develop robots for various applications.
NVIDIA Drive Hyperion and Uber Partnership [1:35:55]
NVIDIA announces the NVIDIA Drive Hyperion, an architecture for creating robo taxi-ready vehicles. Hyperion is designed into Lucid, Mercedes-Benz, and Stalantis. NVIDIA is partnering with Uber to connect these Nvidia Drive Hyperion cars into a global network.
Global Robotaxi Ecosystem [1:37:42]
Robo taxis are at an inflection point, with a large market for connecting and deploying them worldwide. The partnership with Uber aims to create a global network of Hyperion or Robo taxi cars.
Close [1:39:23]
Jensen Huang summarizes the key points covered, including the transition from general-purpose computing to accelerated computing and the AI platform transition. NVIDIA has new platforms for 6G (ARC), robotics cars (Hyperion), and factories (DSX and mega). He thanks the audience and emphasizes the importance of making America great again.