Nvidia GTC: CEO Jensen Huang delivers keynote address

Key Concepts

• Accelerated Computing: A new computing model that leverages parallel processing, primarily through GPUs, to solve problems beyond the capabilities of traditional CPUs.
• CUDA: Nvidia's parallel computing platform and programming model that enables developers to use GPUs for general-purpose processing.
• AI Factories: Specialized data centers designed and optimized for producing AI tokens (computational units of AI) at high rates and low costs.
• Co-design: An approach where hardware and software are developed concurrently and in tandem to achieve maximum performance and efficiency, particularly for AI.
• Tokenization: The process of converting various forms of data (text, images, video, 3D structures, chemicals, etc.) into a numerical representation that AI can process and learn from.
• Virtual Cycle: A self-reinforcing loop of growth where increased usage and capability of a technology lead to more investment, further development, and wider adoption.
• Physical AI: AI that understands and interacts with the physical world, including laws of physics, causality, and permanence.
• Digital Twin: A virtual replica of a physical object, process, or system, used for simulation, optimization, and operation.

Summary

This transcript from Nvidia's GTC event, delivered by founder and CEO Jensen Huang, outlines a vision for the future of computing, driven by artificial intelligence and enabled by Nvidia's technological advancements. The presentation traces America's history of innovation, from the transistor and the microprocessor to the internet and personal computing, positioning AI as the next revolutionary leap.

The Dawn of a New Computing Era: Accelerated Computing and AI

Huang begins by highlighting America's legacy of innovation, citing key milestones like the transistor at Bell Labs, Hedy Lamarr's contributions to wireless communication, IBM's System/360, Intel's microprocessor, and Apple's personal computing revolution. He then pivots to the present, declaring that a "revolutionary new computing model" is emerging, which he believes will be Nvidia's most significant contribution to the computer industry.

Key Points:

• Moore's Law Slowdown: Huang notes that "dinard scaling has stopped," meaning the performance and power efficiency of transistors are no longer improving at the historical rate, necessitating a new approach.
• Nvidia's Solution: Accelerated Computing: For 30 years, Nvidia has been developing "accelerated computing," which combines the parallel processing power of GPUs with the sequential processing of CPUs.
• CUDA and Libraries: The success of accelerated computing is attributed to the invention of the GPU and the CUDA programming model, along with a vast ecosystem of CUDA X libraries. These libraries, such as CU Litho for computational lithography, sparse solvers for CAE, Co-op for optimization, Warp Python solver, QDF for data frames, and CoupNN for large language models, are presented as the "treasure of our company." They enable developers to harness the full potential of accelerated computing across various domains, from chip manufacturing to healthcare (Monai) and genomics.

Revolutionizing Industries: 6G, Quantum Computing, and Enterprise

The presentation then delves into specific industry transformations powered by Nvidia's technology.

1. Telecommunications and 6G:

• The Need for American Innovation: Huang expresses concern that wireless technology, once a domain of American leadership, is now largely deployed on foreign technologies. He sees the current "fundamental platform shift" as an opportunity for the U.S. to regain leadership.
• Nvidia ARC for 6G: Nvidia is launching a new product line called "Nvidia ARC" (Aerial Radio Network Computer). This system, built on Nvidia's Grace CPU, Blackwell GPU, and ConnectX networking, runs the Aerial CUDA X library to create a software-defined, programmable wireless communication system capable of simultaneous wireless communication and AI processing.
• Partnership with Nokia: Nvidia is partnering with Nokia, a major telecommunications player with 7,000 5G patents, to integrate Nvidia ARC into their future base stations. This aims to upgrade millions of base stations globally for 6G and AI capabilities.
• AI for RAN and AI on RAN: AI will be used to improve spectral efficiency in radio communications through reinforcement learning and real-time adjustments. Furthermore, "AI on RAN" will enable cloud computing capabilities to extend to the edge, creating an "edge industrial robotics cloud."

2. Quantum Computing:

• The Breakthrough of Logical Qubits: Huang highlights a "fundamental breakthrough" in quantum computing: the ability to create one "logical qubit" that is coherent, stable, and error-corrected. This logical qubit is composed of tens or hundreds of physical qubits.
• GPU-Quantum Computer Synergy: Nvidia's approach is to connect quantum computers directly to GPU supercomputers. This integration is crucial for error correction, AI calibration and control of quantum computers, and collective simulations.
• NVQLink and CUDA Quantum: Nvidia has developed NVQLink, a new interconnect architecture for direct connection between quantum processors and GPUs, capable of moving terabytes of data per second for quantum error correction. CUDA Quantum is an open platform for quantum GPU computing, enabling researchers to orchestrate quantum devices and AI supercomputers.
• DOE Partnership: The Department of Energy is partnering with Nvidia to build seven new AI supercomputers to advance national science, underscoring the importance of computing in scientific discovery.

3. Artificial Intelligence: The New Industrial Revolution

The core of the presentation focuses on AI as the "new industrial revolution."

• AI Beyond Chatbots: Huang clarifies that AI is much more than chatbots; it's a fundamental reinvention of the computing stack, moving from hand-coded software on CPUs to data-intensive machine learning training on GPUs.
• The Tokenization Paradigm: AI's ability to "tokenize" almost anything – words, images, video, 3D structures, chemicals, cells – allows it to learn, translate, and generate information, mirroring human interaction.
• AI as "Work," Not Just Tools: Unlike previous software that provided "tools" for humans to use, AI acts as "workers" that can perform tasks and augment human productivity. This opens up new economic segments and drives growth, especially in the face of labor shortages.
• AI Factories: The immense computational demands of AI necessitate "AI factories" – specialized data centers designed to produce valuable AI tokens at high rates and cost-effectively. These are distinct from traditional data centers.
• Three Scaling Laws for AI: The recent acceleration in AI is driven by three scaling laws:
  1. Pre-training: Learning basic skills from vast amounts of human-created information.
  2. Post-training: Developing skills to solve specific problems.
  3. Thinking (Inference): The most computationally intensive stage, where AI reasons and solves problems on behalf of humans.
• The Virtuous Cycle of AI: Smarter models lead to more usage, which in turn requires more compute, making the models even smarter. This "virtuous cycle" is now in full swing for AI, mirroring Nvidia's earlier success with CUDA.
• Extreme Co-design: To overcome the limitations of Moore's Law and meet the exponential demands of AI, Nvidia employs "extreme co-design," rearchitecting hardware, systems, software, model architectures, and applications simultaneously.
• Grace Blackwell and NVLink 72: The Grace Blackwell GB200 NVLink 72 system is presented as a prime example of extreme co-design, delivering 10x performance and the lowest cost token generation. This system integrates 72 GPUs into a single rack-scale computer.
• Manufacturing in America: A significant announcement is Nvidia's commitment to manufacturing its AI infrastructure, including Blackwell, in America, starting with facilities in Arizona, Indiana, and Texas. This is framed as a return to reindustrialization and a boost for national security and jobs.
• Vera Rubin Supercomputer: The next generation of Nvidia's rack-scale computer, Vera Rubin, is showcased, offering 100 petaflops of performance and a cableless, liquid-cooled design. It incorporates new processors like BlueField 4 for enhanced memory handling (KV caching) and context processing.
• Omniverse DSX for AI Factories: Nvidia Omniverse DSX is introduced as a digital twin platform for designing, building, and operating gigascale AI factories, enabling collaboration among hundreds of companies and optimizing for power, cooling, and compute density.

Advancing Physical AI and Robotics

The presentation extends to the realm of physical AI, where robots and autonomous systems are being developed.

• Three Computers for Physical AI: Training and operating physical AI requires three distinct computing platforms:
  1. Training Computer: (e.g., Grace Blackwell GB200 NVLink 72) for training AI models.
  2. Omniverse Computer: For simulations and digital twins, enabling robots and factories to learn and be validated virtually. This computer excels in generative AI, computer graphics, and sensor simulation.
  3. Robotics Computer: (e.g., Jetson Thor) for operating robots in real-world scenarios, including self-driving cars and industrial robots.
• Robotic Factories and Digital Twins: Nvidia is working with partners like Foxconn to build state-of-the-art robotic facilities using digital twins for planning, design, and operation. This approach is crucial for managing the complexity of modern manufacturing.
• Humanoid Robots and Autonomous Vehicles: Nvidia is collaborating with companies like Figure, Agility, and Johnson & Johnson on humanoid robots for various applications, including warehouse automation and surgical procedures. The development of robo taxis, powered by the Nvidia DRIVE Hyperion platform, is also highlighted as a significant inflection point, with partnerships with Uber aiming to create a global network of autonomous vehicles.
• Disney's Blue Robot: A demonstration of Disney's "Blue" robot, trained and simulated within Nvidia's Omniverse platform, showcases the potential for realistic and interactive robotic learning.

Open Models and Ecosystem Growth

Nvidia's commitment to open-source models and its expansive ecosystem are also emphasized.

• Importance of Open Source: Huang stresses the critical role of open-source models for startups, researchers, and companies worldwide, enabling domain expertise to be embedded into AI. Nvidia is dedicated to leading in open-source contributions.
• Ecosystem Integration: Nvidia's libraries, software stack, and GPUs are integrated across all major cloud providers (AWS, Google Cloud, Microsoft Azure, Oracle) and enterprise SaaS platforms (ServiceNow, SAP, Synopsis, Cadence).
• New Partnerships: New collaborations are announced with Crowdstrike for AI-powered cybersecurity and Palantir for accelerating data processing and insight generation for government and enterprises.

Conclusion

Jensen Huang concludes by reiterating the two major platform transitions: from general-purpose computing to accelerated computing, and from classical software to AI. These simultaneous shifts are driving Nvidia's extraordinary growth. He highlights new platforms for 6G (ARC), robotics and cars (Hyperion), AI factories (DSX), and AI-enabled factories (MEGA). The overarching message is one of continued innovation, American leadership in technology, and the transformative power of AI to reshape industries and society. The event concludes with a song emphasizing collaboration, progress, and shared vision.