Tokens play a crucial role in detecting diseases early, decoding the intricacies of life, and understanding biological functions, aiding in the conservation of valued species.

At GTC, the discussion highlights the transformative impact of AI and robotics in maximizing potential and enhancing human capabilities, marking the beginning of a new era in technology exploration.

Nvidia highlights significant advancements in artificial intelligence, showcasing how AI has transformed computer graphics and computing models, leading to generative computing and enhanced performance in various industries.

The recent breakthroughs in agentic AI, characterized by its ability to perceive, reason, and act, and physical AI, which understands the physical world, are transforming computing layers. These advancements open new market opportunities and enhance problem-solving capabilities across various industries.

The dialogue discusses the three fundamental challenges in AI development: solving the data problem, training models without human intervention, and achieving scalable algorithms. It highlights the significant increase in computational requirements due to agentic AI's ability to reason, breaking problems down step-by-step, which generates a substantially higher number of tokens compared to previous AI models.

The dialogue highlights the significant increase in computational requirements for AI models, emphasizing the role of reinforcement learning and synthetic data generation in teaching AI to reason. It discusses the challenge of generating vast amounts of data for training, particularly without relying heavily on human input, and showcases the potential of reinforcement learning in solving complex problems across various domains. The speaker underscores the industry's response to these challenges, indicating a major shift in AI training methodologies.

The dialogue highlights the significant growth in AI infrastructure, particularly among the top four CSPs (Amazon, Azure, GCP, Oci), noting the increased demand and computational needs for AI models. It predicts a trillion-dollar data center build-out by the end of the decade, driven by the shift from general-purpose computing to AI-accelerated computing and the recognition of capital investment in software's future. This transformation is characterized by the emergence of AI factories focused on generative computing for various applications.

The dialogue highlights NVIDIA's development and impact of various accelerated libraries, including cuNumeric, cuLitho, and cuOpt, which significantly enhance performance in fields such as computational lithography, 5G radio processing, and mathematical optimization, aiming to open-source some of these tools to further accelerate industry advancements.

The announcement highlights significant advancements in accelerated computing through partnerships with major systems companies, introducing QDSS sparse solvers and accelerated libraries for structured data, physics simulations, and more, marking a tipping point in the industry's shift towards CUDA-powered solutions.

The speaker discusses the evolution of AI, emphasizing its origins in the cloud due to infrastructure and computational needs. They highlight the importance of full-stack technology and the role of cloud service providers in nurturing AI development. The speaker expresses excitement about AI's expansion into various sectors, including enterprise, manufacturing, robotics, and self-driving cars, and mentions the emergence of GPU clouds. They particularly focus on the transformative potential of AI in radio networks and edge computing, announcing a collaboration with major companies to build a full stack for radio networks in the U.S., aiming to revolutionize communications and adapt radio signals using AI and reinforcement learning.

NVIDIA has been deeply involved in the development of self-driving car technology for over a decade, partnering with major companies like Tesla and GM. Their technology, including GPUs and AI-driven systems, is utilized in both data centers and vehicles for training, simulation, and on-road operations. NVIDIA is particularly proud of its commitment to automotive safety, having developed a comprehensive safety assessment process for every line of code and securing over 1000 patents related to safety in autonomous vehicles. They employ advanced AI techniques such as model distillation, closed loop training, and synthetic data generation to enhance the adaptability and safety of autonomous vehicles, ensuring they can navigate complex scenarios robustly.

Nvidia achieves unprecedented scale-up in computer architecture through the Grace Blackwell MvLink 72 rack, marking a fundamental shift from integrated to disaggregated MvLink and from air-cooled to liquid-cooled systems, enabling a 1x flops supercomputer in a single rack.

The dialogue discusses the extreme computing challenge of AI inference, emphasizing the need for high efficiency and performance in generating tokens for AI responses. The speaker explains the significance of token generation speed and volume, highlighting the balance required between latency and throughput in AI systems to ensure optimal service quality, revenue, and profitability.

The discussion centers on the challenge of optimizing AI services to provide high-quality customer experience while maximizing data center efficiency and revenue. It explores the trade-offs between speed, computation power, bandwidth, and the complexity of AI models, highlighting a demo that compares traditional and reasoning models in solving complex problems.

The dialogue discusses optimizing AI models with trillions of parameters by distributing workloads across GPUs using tensor, pipeline, and expert parallelism. It highlights the importance of homogeneous architectures like NVLink for handling complex computing phases, such as context processing and decoding, in large language models.

NVIDIA introduces Dynamo, an open-source operating system designed to manage complex AI workloads in data centers, akin to orchestrating an AI factory, offering dynamic resource allocation for tasks like prefill and decoding.

The discussion explores the optimization of AI factory throughput and energy efficiency using the Hopper and Blackwell systems. It highlights the trade-offs between AI intelligence and volume production, emphasizing the importance of energy-efficient compute architectures in power-limited data centers. Innovations like MvLink 8 and Dynamo are introduced to enhance performance and efficiency.

The discussion highlights the significant performance improvements of Blackwell over Hopper in AI operations, emphasizing the balance between throughput and quality. Blackwell achieves 25 times the performance in one generation under ISO power conditions, demonstrating its superiority in handling next-generation workloads, particularly in reasoning models where it outperforms Hopper by 40 times.

The discussion highlights the superiority of investing in advanced technology, specifically comparing a 100 MW factory's efficiency using Hopper versus Blackwell chips, emphasizing significant improvements in production capacity and cost-effectiveness.

The dialogue highlights the complexity and scale of building state-of-the-art AI factories, emphasizing the use of digital twins for planning and optimization before physical construction. Nvidia's Omniverse Blueprint enables collaborative engineering teams to simulate and iterate designs efficiently, significantly reducing errors and accelerating the build-up process. The speaker also outlines a roadmap for upcoming AI infrastructure, including the Blackwell Ultra and Vera Rubin systems, showcasing the strategic planning required for such large-scale projects.

The dialogue discusses NVIDIA's GPU advancements, clarifying that each Blackwell chip comprises two GPU dies, and outlines future plans including the Ruben Ultra project, which aims for extreme scale-up with 576 MvLink connections and significant increases in performance and bandwidth. It also highlights the transition to Spectrum X networking technology, aiming to enhance Ethernet networks with Infiniband-like properties for better performance and manageability, exemplified by the creation of the largest single GPU cluster with Spectrum X.

The discussion outlines the challenges and solutions involved in scaling GPU usage in data centers, particularly focusing on the energy and cost implications of using transceivers. It introduces the invention of the world's first MRI micro mirror and the implementation of silicon photonics, which significantly reduce power consumption and enable the handling of massive GPU volumes. These advancements pave the way for more efficient, high-capacity data center architectures, aiming to save tens of megawatts of power and facilitating the deployment of millions of GPUs.

The transformation of enterprise computing through AI involves a new computing stack, including changes in processors, operating systems, and data access methods. AI agents will become integral to the digital workforce, significantly altering how enterprises operate and necessitating the development of a new line of computers.

A new line of high-performance computers, including the DGX Station and Spark, is announced for data scientists and researchers, marking the era of AI computing. These systems, capable of 20 petaflops and equipped with 72 CPU cores, HBM memory, and PCI Express slots, will be manufactured by major OEMs like HP, Dell, Lenovo, and Asus. The company is also revolutionizing computing, networking, and storage, introducing an open-source AI model that can be run on various platforms, including the new computers and in the cloud. Additionally, partnerships with global companies are highlighted for integrating AI frameworks. The discussion shifts to robotics, emphasizing the need for autonomous robots due to a global labor shortage. Nvidia's technology enables a continuous loop of robot AI simulation, training, testing, and real-world experience, with the introduction of Isaac GRO N1, a generalist foundation model for humanoid robots, showcasing advancements in embodied AI and robotics.

The core challenges in AI include data creation, model architecture, and scaling laws. To address these in robotics, Omniverse serves as the operating system for physical AI, enhanced by Cosmos for generating infinite, grounded environments. Newton, a collaboration between DeepMind, Disney Research, and Nvidia, introduces a specialized physics engine designed for fine-grained simulations, accelerating AI training for robotics.

After showcasing advanced robotics and AI technologies, including real-time simulation and robot training methods, the speaker announces that Group N 1 robotics platform is now open sourced, highlighting significant progress in AI and robotics. The event also covers the production status of Blackwell, emphasizing the increasing demand for AI computation and infrastructure for cloud, enterprise, and robots.