VivaTech 2025 黄仁勋主题演讲
文章语言:
简
繁
EN
Share
Minutes
原文
会议摘要
Nvidia introduces the Grace Blackwell system, a supercomputer designed for advanced reasoning and decision-making tasks, marking a significant advancement in AI technology. This system accelerates decision-making for complex problems, supports quantum algorithm acceleration, and embodies the transition from perception-based AI to agentic AI, capable of understanding, reasoning, planning, and executing tasks. Additionally, Nvidia is partnering globally to integrate AI into various industries, emphasizing safety, reliability, and industrial transformation.
会议速览

Tokens, generated by a novel factory, serve as AI's foundational elements, transforming images into scientific data, exploring unknown territories, detecting Earth's hidden dangers, enhancing agricultural productivity, aiding medical diagnoses, and conserving wildlife.

Tokens are revealed as transformative elements that decode the laws of physics, enhancing efficiency and enabling technological advancements that bring joy and comfort. They signify significant leaps in human capability, allowing exploration beyond known limits, starting from a doctor's visit and extending to the cosmos.

At the inaugural GTC Paris, NVIDIA's CEO highlights the evolution of accelerated computing, emphasizing significant advancements in libraries for various applications, including semiconductor design, molecular dynamics, and quantum computing. He announces the integration of CUDA Q with Grace Blackwell 200, marking a pivotal moment in quantum-classical computing collaboration.

The speaker congratulates the quantum computing industry for its accomplishments and discusses the significant progress in AI, transitioning from perception and generative capabilities to agentic AI, which involves understanding, reasoning, and planning. This new wave of AI is revolutionizing robotics and is rooted in advancements started with GeForce and accelerated computing applications.

The dialogue explores the advancement from traditional GeForce graphics cards to complex, thinking machines like GB 200 and GB 300, emphasizing the role of digital twins in designing, planning, and operating physical entities digitally before their physical realization. The discussion highlights the creation of a 'thinking machine' capable of reasoning and planning, akin to human thought processes, and the technological innovations required to assemble a giant virtual GPU with high energy efficiency and low power consumption.

The dialogue discusses the evolution from the Hopper system, featuring 8 GPUs connected via NVLink, to the Blackwell node, which offers superior performance and liquid cooling. The introduction of the MVLink system allows for the disaggregation of traditional motherboard components, enabling more efficient scaling up of AI supercomputers. This innovation results in a system capable of achieving 30-40 times more performance than previous generations, necessary for handling the increased computational demands of reasoning models in AI.

The dialogue details the intricate engineering process behind the Grace Blackwell supercomputing systems, highlighting the technological advancements and global collaboration required for their mass production. It emphasizes the system's high performance and efficiency, surpassing previous supercomputers while consuming significantly less power, and discusses the potential for these systems to drive future discoveries and innovations.

The dialogue discusses the introduction of the Grace Blackwell system, emphasizing its adaptability for both advanced AI operations and traditional IT systems. It highlights the RTX Pro server, designed to bridge AI-native architectures with conventional enterprise IT, offering compatibility with various operating systems and applications, thus transforming data center capabilities globally.

AI data centers, or AI factories, are redefining the concept of data storage by solely focusing on generating intelligent tokens for various industries, transforming them into critical revenue-generating facilities. These facilities, equipped with a vast number of computers, are not traditional data centers but are designed and scaled to produce AI, significantly impacting economies worldwide. Leaders across nations are recognizing the importance of integrating AI factories into their infrastructure, leading to a new industrial revolution driven by artificial intelligence. Europe is notably increasing its AI computing capacity by tenfold within two years, addressing shortages and enhancing AI capabilities for researchers and startups.

Nvidia is partnering with countries worldwide to develop AI ecosystems, establishing technology centers in seven nations for collaborative research and startup engagement. The UK ecosystem, recently visited, is built on the Nvidia stack, showcasing partnerships with major players in computing, software, and cloud services. This reinvented computing stack supports a new software development paradigm, integrating DevOps, ML ops, and AI ops, with support from solution integrators and providers. Significant partnerships exist in the UK, Germany, Italy, and France, aiming to upskill local economies and integrate AI capabilities into enterprises.

During a visit by President Macron, significant announcements are made regarding AI advancements and partnerships in France, including a collaboration with Schneider to build AI factories digitally and a partnership with a young European AI company to develop an AI cloud. The dialogue emphasizes the rapid evolution of AI technology, the importance of digitalizing factories using AI, and the dedication to enhancing open-source AI models for enterprise applications.

A partnership with Perplexity integrates regional models into their reasoning search engine, allowing queries tailored to specific cultures and languages. Agentic AI, characterized by its capability to reason, plan, and execute tasks autonomously, represents a significant advancement from one-shot AI. An example demonstrates how a single prompt can initiate a complex, multi-step plan involving various AI agents to start a food truck business in Paris. The need for high-performance systems is highlighted due to the exponential increase in data processing required by agentic AI. Companies are encouraged to build specialized agents using provided platforms, tools, and frameworks, showcasing the future of AI integration in businesses.

Nvidia introduces DGX Lepton, a system enabling deployment of AI microservices across various clouds, including public, regional, and private, simplifying management and scalability for diverse computing needs.

The dialogue highlights the evolution of AI supercomputers, starting with the DGX 1 built in 2016, and progresses to discuss the development and capabilities of Lepton, a platform enabling on-demand access to a global network of GPUs across various clouds and regions. It emphasizes the ease and reliability Lepton provides for developers in managing multi-cloud GPU clusters, deploying AI models, and monitoring progress in real time. The discussion also showcases several companies and applications built on NVIDIA's technology, illustrating the widespread adoption and integration of AI solutions across different industries.

The discussion highlights the significance of Europe's contribution to AI, specifically mentioning the Synapse 1 computer from 1992 which ran neural networks 8000 times faster than CPUs at the time. A partnership with Siemens aims to fuse European industrial capabilities with AI, leading to the Industrial AI Revolution. This collaboration spans various areas including design, simulation, digital twins of factories, and AI operations, presenting a significant opportunity to revolutionize industries through smart software.

The narrative outlines the progression from the inception of the Industrial Revolution, marked by innovations like Watt's steam engine, through the electrification era led by figures such as Faraday and Siemens, to the current era of AI and robotics. It highlights the transformation of manufacturing, the advent of intelligent autonomous agents, and the development of a collaborative workforce addressing global labor shortages.

The dialogue highlights the utilization of digital twins in Omniverse by various companies, including BMW, Mercedes Benz, Schaefer, a French train station, and Toyota, to design, plan, and simulate the effectiveness of their factories and warehouses before physical implementation.

The speaker discusses the significance of digital twins in enabling robots to learn realistic operations, highlighting the necessity for photo-realistic simulation and adherence to physical laws. They announce the development of the world's first industrial AI cloud in Europe, designed for real-time design, simulation, and robotic learning. The ecosystem supports the digital twin revolution and the AI-driven future of robotics, with Nvidia providing the AI supercomputers for both Omniverse digital twins and robotics, emphasizing safety and high-performance requirements.

Nvidia Drive, leveraging the Halo safety system, enables the construction of safe autonomous vehicles (AVs) through diverse software stacks, sensors, and redundant computers. Utilizing Nvidia Omniverse and Cosmos, developers generate realistic synthetic training data to address real-world edge cases, enhancing AV models' ability to perceive, reason, and make safe decisions. An independent classical stack monitors performance, calling for emergency stops when necessary, while redundant sensors and computers ensure operational safety even in case of failures. In critical situations, the system executes minimum risk maneuvers, such as pulling over, to maintain safety.

As of June 11, 2025, the future of autonomous driving and humanoid robotics is poised for a significant transformation powered by AI. With a focus on safety, companies are developing software-defined AI-driven systems for autonomous vehicles, marking a decade of progress in the field. Similarly, advancements in AI are enabling the generation of video from prompts, leading to the potential for humanoid robots with advanced motion and articulation abilities, promising to create one of the largest industries ever, particularly in European countries known for their manufacturing prowess.

Discussing the hindrance of widespread robot deployment due to complexity and cost, the dialogue highlights the current limitation to large corporations. It introduces a solution through humanid AI, emphasizing simple teaching toolkits and the development of advanced robotic computers like Thor, aimed at making robotics accessible to smaller businesses and entities.

The dialogue discusses the development of human robotics, focusing on the Thor processor and the challenge of acquiring training data. This issue is addressed through Omniverse, a digital twin world that simulates physics for AI training. Partnerships with Disney Research and DeepMind aim to create advanced physics simulations for robotics learning.

Through extensive training in varied virtual scenarios, the individual develops physical movement skills that seamlessly translate to real-world environments, showcasing the effectiveness of virtual reality in skill acquisition.

The keynote presenter instructs a robot to behave, suggests taking a group photo, humorously invites the robot home, acknowledges its intelligence, and envisions a future where such robots are commonplace, albeit lacking physical capabilities to serve drinks.

An industrial revolution driven by AI is underway, highlighted by advancements in robotics and the creation of Blackwell, a machine designed for reasoning. The demand for inference workloads is exponentially increasing, leading to the establishment of AI factories focused on generating tokens. Europe is significantly expanding its AI infrastructure, demonstrating a strong commitment to the AI sector.
要点回答
Q:What are tokens and how do they contribute to AI?
A:Tokens are the building blocks of AI, acting as a new kind of factory generator that transforms various types of data, such as images into scientific data, and aid in the exploration of new possibilities in AI development.
Q:What are some examples of applications where Nvidia's technology is used?
A:Nvidia's technology is used in various applications including molecular dynamics, computational lithography in semiconductor design, decision making with optimization problems, geometry and physics solvers, and it supports classical machine learning among others.
Q:What advancements have been made in the field of quantum computing according to Nvidia?
A:Nvidia has announced advancements in quantum computing with the demonstration of the world's first logical qubit and the opening of an inflection point in quantum computing. They have been working on CudaQ, a new library that enables a quantum-classical accelerated computing approach, and have announced that their entire quantum algorithm stack is now accelerated on Grace Blackwell 200.
Q:What is the relationship between AI and Nvidia's GPUs?
A:Nvidia GPUs have been instrumental in the development of AI, starting with the revolutionization of deep learning with the AlexNet in 2012. GPUs have continued to advance AI through deep learning research, and now support the new wave of AI, including multimodal capabilities and agentic AI, which involves embodied AI and the ability to generate locomotion.
Q:What is the concept of digital twins and how is it related to simulations?
A:Digital twins are virtual representations of physical entities or systems, which can be designed, planned, optimized, and operated before they are built in the physical world. This concept is closely related to simulations, as everything can now be turned into a digital twin thanks to the scale and speed of modern simulations.
Q:What is the design philosophy behind Nvidia's Grace Blackwell 200?
A:The design philosophy behind Nvidia's Grace Blackwell 200 is to create a 'thinking machine' that is capable of reasoning, planning, and performing tasks autonomously, similar to how humans use their own minds to generate thoughts and ideas.
Q:What is the Hopper system and how did it relate to AI?
A:The Hopper system was an early computing system that put Nvidia on the map in the field of AI. It was quickly replaced by the more advanced Blackwell node, demonstrating the rapid advancement in AI computing power.
Q:What is the significance of the Mv link and nvlink systems in AI computing?
A:Mv link and nvlink are innovative interconnect systems that allow the scaling out of computing power by connecting more CPUs with Ethernet. Nvlink Mv link is a memory semantics interconnect that directly connects CPUs of different Mv link systems, forming a compute fabric, not a network, to improve performance and efficiency.
Q:What are the technical specifications of the Blackwell system's spine?
A:The Blackwell system's spine, made of 100% copper, is designed to directly connect all Mv link chips to GPUs without blocks, allowing them to communicate at the same time. It supports an astonishing bandwidth of 130 TB per second, more than the peak traffic of the world's internet.
Q:How much faster is the Blackwell system compared to the previous Hopper system?
A:The Blackwell system is a giant leap in performance compared to the Hopper system, offering up to 30-40 times more performance per generation. This is crucial as reasoning models used in AI such as ChatGPT require a lot more computational capability.
Q:What is the architectural impact of the Blackwell system on AI and revenue generation?
A:The architectural design of the Blackwell system, including the high bandwidth capabilities, enables faster thinking and output, supporting more users simultaneously. This results in significantly higher revenues for AI-based factories, making the Blackwell system pivotal for the growth of the AI industry.
Q:What is the process involved in manufacturing the Blackwell system?
A:The manufacturing process for the Blackwell system involves several steps, starting from a blank silicon wafer that goes through hundreds of chip processing and lithography steps. This results in the creation of 200 billion transistors, which are then tested, sorted, and assembled into system-in-wafer packages and eventually integrated into racks of GB 200 compute nodes.
Q:How is the Blackwell system's performance and architecture described?
A:The Blackwell system is described as an engineering marvel featuring a highly integrated design with 144 Blackwell GPUs in each system, interconnected via a custom mated backplane with over 5000 copper cables. The system is capable of delivering 130 TB per second of all-to-all bandwidth, showcasing its massive computational power.
Q:What is the significance of the Blackwell system in terms of global production and energy consumption?
A:The Blackwell system marks a new level of supercomputing, with production capabilities enabling the creation of AI supercomputers at unprecedented scales. The energy consumption of these systems, rated at 100kW, is significantly higher than previous top systems, indicating a new generation of AI supercomputers that are both more powerful and energy-intensive.
Q:How does the AI industry relate to the concept of AI data centers?
A:AI data centers are not traditional data centers; they are facilities designed to generate intelligent tokens through AI processes. These 'factories of AI' are revenue-generating and are transforming into a country's infrastructure, with the potential to revolutionize every industry. This marks the emergence of a new industrial revolution and a new 'intelligence infrastructure' that will be crucial for every country, society, and company.
Q:What are European telcos and cloud service providers doing in terms of AI infrastructure?
A:European telcos and cloud service providers are building AI infrastructure with Nvidia, including the construction of AI supercomputers and the planning of additional AI factories and gigafactories, which will increase the amount of AI computing capacity in Europe by a factor of 10 in just two years.
Q:What is the role of the AI technology centers in different countries?
A:The role of the AI technology centers in different countries is to conduct collaborative research, work with startups, and build the ecosystem for AI in those regions.
Q:What is the significance of the Nvidia stack in AI ecosystems?
A:The Nvidia stack is significant in AI ecosystems because it provides a comprehensive suite of tools and technologies that enable the development and deployment of AI applications, making it accessible across various platforms, including cloud services.
Q:How is Nvidia collaborating with other companies to reinvent the computing stack?
A:Nvidia is collaborating with companies like Siemens, Cadence, Red Hat, ServiceNow, Cisco, NetApp, and Nutanix to reinvent the computing stack by developing new models and technologies that integrate Nvidia's capabilities into the stack.
Q:What is the impact of Nvidia's AI technology centers on local economies and talent upskilling?
A:The AI technology centers are contributing to the upskilling of local economies and talent in the UK and other countries by working with researchers, developers, partners, and cloud service providers to foster an ecosystem that supports the development of AI capabilities.
Q:What new announcements are being made regarding AI factories and digital twins?
A:New announcements include a partnership with Schneider to build AI factories that are designed, built, and operated digitally, with plans to optimize them completely digitally in the future, even as digital twins.
Q:What is the purpose of the new AI cloud being announced in collaboration with Mistral?
A:The new AI cloud being announced in collaboration with Mistral is intended to deliver AI models and applications for the ecosystem, and to allow other AI startups to use models such as Mistral or any other model they develop.
Q:How is Nvidia enhancing open models through its Nemo tools?
A:Nvidia is enhancing open models through its Nemo tools by taking pre-trained models, potentially doing neural architecture search, using better data, and employing reinforcement learning techniques to extend the context and reasoning capabilities of these models.
Q:What does the partnership with Perplexity involve and what new features does it bring to language models?
A:The partnership with Perplexity involves connecting regional models to Perplexity's search engine, allowing users to ask and receive questions in the language, culture, and sensibility of their country. This partnership enhances the personalization and relevance of the AI models.
Q:How are AI agents changing the use of AI and what are their capabilities?
A:AI agents are transforming the use of AI by reasoning through problems, breaking them down into multi-step plans, using proper tools, working with other agents, and utilizing context from memory. They can integrate various AI models and provide comprehensive solutions to complex prompts.
Q:What are the different cloud environments where AI services can be deployed?
A:AI services can be deployed in public clouds, such as Nvidia's NCPs like Mistral, private clouds due to security and privacy requirements, on personal computers, or on various cloud platforms such as AWS and GCP. Nvidia's architecture is designed to run on a 'cloud of clouds,' allowing deployment flexibility.
Q:What is Nvidia's DGX-LEpton and how does it facilitate deployment of AI models?
A:Nvidia's DGX-LEpton is a system that allows for the deployment of AI models across various cloud environments, including private and public clouds, and even on personal computers. Users can deploy AI models using one 'super cloud' interface, and the model will be hosted and run on the selected cloud environments. This facilitates running AI models anywhere, including on smaller machines like the DGX Spark.
Q:How does Nvidia's architecture facilitate the deployment of AI models across different platforms?
A:Nvidia's architecture, which is present in many different cloud setups, enables developers to deploy AI models consistently across various platforms such as microservices, AI agents, or整个系统。AI模型可以在不同的云环境、私有云、以及本地机器上运行,利用的是同一套模型架构和部署方法。
Q:What is the purpose of the AI supercomputer and how was it received by the market?
A:The purpose of the AI supercomputer, such as the DGX-1, is to provide developers with easy and reliable access to compute resources for AI development. Initially, there was confusion in the market regarding its purpose and whether it could run common software like Windows. Despite the initial lack of interest, the AI supercomputer was eventually sold to OpenAI, marking the start of its market presence.
Q:What is Lepton, and how does it assist with AI model deployment?
A:Lepton is a system that offers developers on-demand access to GPUs across various clouds, regions, and partners. It allows the deployment of AI models into multiple clouds or regions for fast distributed inference and supports the entire lifecycle of an AI agent. Lepton assists with AI model deployment by providing an interface that integrates with Hugging Face and allows for deployment into different cloud environments with a single click.
Q:What is the significance of industrial AI and the partnerships Nvidia has formed in this area?
A:The significance of industrial AI lies in its potential to revolutionize industries by integrating artificial intelligence with traditional industrial capabilities. Nvidia has formed partnerships with companies like Siemens to create an 'Industrial AI Revolution,' developing solutions from design to simulation to operations in factories. These collaborations leverage AI to enhance efficiency and reinvent various industrial processes.
Q:How are industrial AI and digital twins revolutionizing manufacturing?
A:Industrial AI and digital twins are revolutionizing manufacturing by enabling virtual design and simulation of factories, warehouses, and other physical infrastructure. This technology allows for real-time planning and optimization of manufacturing processes, enabling the creation of intelligent autonomous agents and robots that can operate collaboratively with human workers. Digital twins provide a realistic representation of the physical world, which helps robots learn and adapt, improving the overall efficiency and adaptability of manufacturing environments.
Q:What achievements has Nvidia's AI team, Av team, accomplished in the field of self-driving cars?
A:Nvidia's AI team, Av team, is recognized for winning the end-to-end self-driving car challenge at CVPR two years in a row, and they won again that year as well.
Q:How does Nvidia Drive ensure safety in autonomous vehicles?
A:Nvidia Drive, built on the Halo safety system, enables the construction of safe autonomous vehicles. It features diverse software stacks, sensors, and redundant computers for safety. The system starts with training on diverse data using Nvidia Omniverse and Cosmos to generate realistic synthetic data. An independent classical stack runs in parallel to monitor safety, and in case of anomalies, it can call the arbitrator for an emergency stop. Furthermore, diversity and redundancy are integrated into the sensor and compute architecture to maintain safety even with sensor or computer failures.
Q:How does Nvidia's Halo system contribute to the safety of autonomous vehicles?
A:Nvidia's Halo system contributes to the safety of autonomous vehicles by being integrated into products worldwide to build the next generation of safe AVs. It covers the entire architecture from chip design and system design to software development and testing. The system is the first to be software-defined, featuring a completely software-driven stack for AVs, developed over nearly 10 years, making it world-renowned for its safety capabilities.
Q:What is the next gigantic opportunity mentioned in the speech?
A:The next gigantic opportunity mentioned in the speech is the future of autonomous driving, which is expected to be powered by AI and will be a massive industry.
Q:What is the path for AI to revolutionize robotics according to the speech?
A:The path for AI to revolutionize robotics involves AI's ability to perceive, reason, and generate videos and words, which is now applicable to cars. The revolution in robotics is anticipated to be similar to the developments in the automotive industry, where AI will be able to generate local motion and articulation abilities for robots.
Q:What are the challenges faced by traditional robots and what solution does Nvidia propose?
A:Traditional robots are hard to program, and only large companies can afford to install, teach, and operate them. Nvidia proposes a solution by offering robots that can learn from simple teaching using user-friendly toolkits, similar to the ones used for Nemo.
Q:What is the core technology introduced by Nvidia for robotics and how does it operate?
A:The core technology introduced by Nvidia for robotics is the Thor processor, a robotic computer in a dev kit form. The kit includes sensors and a Thor chip, which operates within an operating system designed for robotics. It uses transformer models to process sensor data and generate flight or paths for motor controls for articulation of arms, fingers, and legs. The training of these robots happens in a digital twin world within Nvidia's Omniverse, which simulates and trains robots following the laws of physics.
Q:How is Nvidia collaborating with other companies to advance AI and robotics?
A:Nvidia is collaborating with Disney Research and DeepMind to create the world's most sophisticated physics simulation. This partnership aims to advance AI and robotics by enabling more realistic and varied training scenarios in a virtual world that obey the laws of physics.
Q:What is the significance of Greg, the robot, in demonstrating AI advancements?
A:Greg, the robot, demonstrates AI advancements by learning to walk in a virtual world that simulates various terrains. This learning translates to the physical world, where Greg can now perform complex tasks such as jumping and dancing. The significance lies in showing how AI can learn from virtual experiences and apply that learning to real-world situations.
Q:How is the new wave of AI impacting inference workloads and AI factories?
A:The new wave of AI is causing inference workloads to explode and grow exponentially. The use of prompts and tokens generated from AI has increased significantly, and as a result, there is a need for specialized computers like Blackwell designed for thinking and reasoning. These Blackwell chips will be used in new types of data centers, referred to as AI factories, which will generate tokens, becoming the primary resource.

NVIDIA Corp.
Follow