黄仁勋 COMPUTEX 主题演讲
文章语言：
简
繁
EN
Share
Minutes
原文
会议摘要
NVIDIA founder and CEO Huang Renxun detailed how AI and its infrastructure are reshaping the computer industry and global industrial landscape. NVIDIA has transformed from a chip company to a leader in the AI and robotics field, driving the AI revolution by introducing key technologies and products such as CUDA and DGX1. The AI data center, as an AI factory, generates high-value "tokens" through applying energy, becoming an important indicator of company productivity.
NVIDIA has made significant progress in areas such as accelerated computing, AI, libraries, and simulation science, with advancements like GForce, CUDA, professional libraries, and the Grace Hopper supercomputer driving progress in multiple fields. Additionally, NVIDIA has made recent advancements in communication, quantum computing, AI reasoning, and physical AI, collaborating with global partners to build AI infrastructure and ecosystems. By leveraging technological innovations such as envy link, Blackwell GPU, and AI data platforms, as well as collaborating with the storage industry to integrate AI models into the storytelling stack and AI platform, NVIDIA is driving large-scale data processing and decision support.
The emergence of AI Operations (AI Ops) heralds the future role of dedicated AI operations teams responsible for data management, model fine-tuning, and security, further strengthening AI's role in enterprise IT.
会议速览
Smart factories and Tokens usher in a new era of the future.In 2025, Tokens manufactured in smart factories will serve as the cornerstone of artificial intelligence, unlocking new possibilities in areas such as transforming images into scientific data, aiding in the exploration of the universe, overcoming real-world challenges, early detection of disease symptoms, studying human movements, and bridging the gap between the real and virtual worlds. These technologies not only improve the quality of life but also provide endless possibilities for future exploration.
NVIDIA CEO Huang Renxun announces innovative product and market strategies.The founder and CEO of NVIDIA introduced to the audience a series of exciting new products that the company will soon launch during his speech on May 19, 2025. These products will not only open up new markets, but also create new growth points. He emphasized the importance of developing the ecosystem together with valuable partners and friends, and mentioned that NVIDIA is at the core of the computer ecosystem, one of the most important industries in the world. He also teased some surprises that are about to be revealed.
The Journey of Nvidia and the Artificial Intelligence RevolutionSince 1993, Nvidia has transformed from a chip company to focus on creating new computing platforms. In 2006, Nvidia launched CUDA, which fundamentally changed the way computing is done. Ten years later, in 2016, the company realized that a new computing method had arrived, requiring a reinvention of every layer of the technology stack. In response, Nvidia invented the new system DGX-1, which was donated for the first time to the non-profit organization OpenAI, thus sparking the artificial intelligence revolution.
The road to innovation of artificial intelligence and data centers.With the development of artificial intelligence, the traditional way of software operation has been completely changed, and the architecture of data centers has undergone fundamental changes. New applications require a large number of processors to work together to respond to queries from millions of users, which has increased the importance of East-West traffic, especially for high-performance computing and large-scale distributed processing. Five years ago, in 2020, a company focused on such networks, Melanotan, was acquired to achieve the goal of transforming the entire data center into a computing unit. Today, the data center is seen as a single computing unit, rather than a single PC or server, and its operating system has also undergone a transformation. Over the past three years, this series of innovative ideas has gradually taken shape, reshaping the company's development direction.
NVIDIA is transforming into an AI infrastructure company.NVIDIA announced that it is no longer just a technology company, but is transitioning into an AI infrastructure company. It has revealed a roadmap for the next five years to help various industries around the world build AI infrastructure. This transformation is likened to the historical construction of electricity and internet infrastructure, indicating that AI will become a crucial part of future societal infrastructure. NVIDIA points out that the new AI data centers, known as AI factories, will generate significant value, and the company will measure its AI production capacity by the number of 'tokens' produced. This signals the comprehensive integration and importance of AI in various industries.
From Accelerated Computing to AI Factories: Outlook on Future Technology.Since its founding in 1993, the business opportunities of the company have grown from an initial three million dollars to today's trillion-dollar market, especially in the field of AI infrastructure. The company's core competitiveness lies in the integration of accelerated computing and AI, with a special emphasis on the importance of algorithms and libraries (such as cut x library). The upcoming presentation will cover the latest advances in simulation science and artificial intelligence, all of which are not artistic creations, but beautiful presentations based on technology.
Real-time computer graphics technology and AI revolutionIn 2025, a new graphics processor called the G Force RTX Fifty Series made a breakthrough in real-time computer graphics generation. This GPU was developed by a company called Emaciated and integrated into a new laptop called Emc Zed. Through artificial intelligence technology, the system only needs to calculate one-tenth of the pixels, with the rest predicted by AI, achieving high resolution and smooth frame rates in real-time graphics rendering. This technology, called OSS New Reminding, after ten years of research and development, completely revolutionized the field of computer graphics. The successful release of the G Force RTX Fifty Series marks a deep integration of AI with graphics processing technology, driving the development of the PC gaming industry.
Global Application and Development of Accelerated Computing with CUDA Technology.The promotion of accelerated computing and CUDA technology on a global scale is crucial. By widely installing CUDA GPUs, more developers are attracted to create powerful application libraries, thereby driving better applications and benefits for users, forming a virtuous cycle. Accelerated computing is different from general computing, requiring specific architectures to accelerate critical parts of applications, greatly improving operational efficiency. Through collaborative efforts with partners in various fields, CUDA technology has been applied in computer graphics, numerical computing, medical imaging, quantum computing, and other areas, providing strong acceleration support for 5G, AI, weather forecasting, and more. The widespread application of this technology has brought innovative changes to industries such as telecommunications and data centers.
The future development trend of quantum computing and AI.The conversation delved into the future prospects of quantum computing and artificial intelligence (AI). It was mentioned that quantum computing is still in its early stages of development, but there are already many application areas where it can be implemented. Looking ahead, it was discussed that all supercomputers will be equipped with quantum accelerators, forming modern computers that combine quantum processing units (QPUs), GPUs, and CPUs. In addition, the development of AI was discussed, from perceptual AI to generative AI, to agent AI with reasoning capabilities, ultimately leading to physical AI that can understand the physical world. Particularly in the field of physical AI, the use of generated videos to train scenarios such as self-driving cars was mentioned, demonstrating significant advances in AI technology.
Grace Blackwell系统：AI推理与高性能计算的未来In 2025, a high-performance computing system named Grace Blackwell was introduced, which aims to achieve fast reasoning and time scaling capabilities for AI, in order to handle more complex and faster AI tasks. The Grace Blackwell system not only achieves scalability, but also connects multiple computers together to work together, overcoming the limitations of semiconductor physics. Since the system went into full production at the end of last year, it has been widely used in various fields, with its performance continuously improving. The latest version, gb300, offers higher reasoning performance, memory, and network capabilities, significantly improving the overall system performance.
Grace Blackwall 计算系统革新与AI工厂建设Grace Blackwall system's computing nodes have been updated, and the new version GB300 system performance has improved by 1.5 times, especially in terms of inference performance. It was mentioned in the discussion that a 40PFLOPS system, equivalent to the performance of the 2018 SIA supercomputer, has replaced the original system of 18,000 GPUs with just one node. Through innovative technologies such as TF-C's CoaXl process and Envy Link high-speed switch, high-speed communication between GPUs has been achieved. This technological innovation has made it possible to build large-scale AI factories, such as the XAI Colosse factory, which covers four million square feet and has a total electronic equipment investment of forty to fifty billion dollars. These advancements highlight the rapid expansion of computational capabilities in the AI and video fields, foreshadowing the future development trends of AI factories.
Manufacturing and integration process of the Blackwell super chip.The manufacturing of the Blackwell chip starts with a blank silicon wafer, which goes through hundreds of processes and UV lithography steps to eventually build two hundred billion transistors on a twelve-inch wafer. After testing and screening, good chips are further processed, connected to HBM stacks and custom silicon interposers to form system-level packaging units. Subsequently, the components are subjected to high-temperature testing and integrated across multiple factories worldwide, ultimately manufacturing the Blackwell super chip incorporating 1.3 trillion transistors with the support of Taiwan's technology ecosystem. This process demonstrates the exquisite craftsmanship and precise collaboration of global partners.
NVIDIA announces the establishment of the first AI supercomputer in Taiwan.NVIDIA has announced a partnership with the Taiwan government and industry partners to jointly establish the first AI supercomputer in Taiwan, aimed at advancing Taiwan's AI infrastructure and ecosystem. In addition, NVIDIA has also launched Envy Link Fusion, a new solution that allows enterprises and research institutions to customize AI infrastructure based on their needs, including integrating custom CPUs and GPUs. Meanwhile, NVIDIA's DGX Spark supercomputer has entered full production stage and is expected to be available in the market soon.
Reshaping computation: AI native computing and digital employeesBy 2025, with the rapid development of AI technology, the enterprise IT sector is undergoing a major transformation. AI-native computers, designed specifically for modern AI applications, no longer need to run traditional IT software but focus on efficiently running AI applications. At the same time, AI is changing the three layers of computing, storage, and networking architecture to adapt to new enterprise IT requirements. In the future, businesses will use digital employees, AI agents, to perform a variety of tasks to address the global labor shortage issue. These digital employees will become a new type of human resource for enterprises, and the IT department will be responsible for managing and optimizing these AI agents to improve work efficiency and productivity. To achieve this vision, it is necessary to fundamentally redesign computing systems to meet the dual requirements of modern AI and enterprise IT.
NVIDIA发布全新RTX Pro企业级服务器In 2025, NVIDIA launched a new RTX Pro enterprise server, which can not only run traditional high-performance visualization and classic applications, but also has powerful enterprise AI processing capabilities, supporting various forms of AI processing such as text, graphics, and video. The server uses the innovative CX A chip, providing up to 800Gbps of network bandwidth, significantly improving the inference speed and efficiency of AI models. Compared to other top servers on the market, the RTX Pro shows higher performance and lower latency, especially demonstrating huge advantages in AI inference and large-scale data processing. In addition, NVIDIA also introduced its AI data platform, aimed at handling unstructured data queries to meet AI's semantic and meaningful query needs.
Future Prospects of AI Ops and Robotics TechnologyIn 2025, with the continuous development of AI technology, a new software layer AI Ops is proposed to provide AI support to enterprise IT departments through operations such as data processing, model fine-tuning, evaluation, and protection. Multiple partners such as CD Crowd Strike, Data IQ, Data Robots, etc. will participate in building the AI Operations ecosystem to provide integration and deployment of AI models for enterprises. In addition, through collaboration with DeepMind, Disney Research, etc., a high-level physics engine called Newton has been developed to accelerate the learning of robots in virtual worlds, achieving more realistic and efficient robot simulations. The platform is not only suitable for virtual worlds but also supports applications in real-world robots and autonomous driving vehicle systems. Currently, version 1.5 of the Isaac Group platform has been open-sourced and available for global download and use.
Breakthroughs and developments in robotics technology and AI in the industrial field.With the rapid development of robotics technology and artificial intelligence (AI), the industrial sector is undergoing significant changes. In the face of a global labor shortage, robots, especially general-purpose robots that can adapt to various environments, have become key in addressing the challenges. By using AI to amplify human demonstration systems, researchers are tackling the challenge of robot learning data, greatly improving the efficiency and adaptability of robots. In addition, through digital twin technology and large-scale synthetic data generation, robots can learn and optimize their behavior in virtual environments, preparing for real-world applications. This technological advancement not only accelerates the maturity of robotics technology but also heralds the arrival of a new industrial era dominated by robots and AI.
NVIDIA announced the establishment of a large new office building in Taiwan.On May 19, 2025, NVIDIA announced the establishment of a large new office building in Taiwan called "NVIDIA conStellation" to address the growing business needs and engineering team. The construction has received support from the local mayor, but further approval from the public is required. In addition, NVIDIA emphasized the opportunities for creating new industries such as AI factories, enterprise agents, and robots with partners, demonstrating a huge expectation and confidence in the future.
要点回答
Q：What are the potential applications of tokens in various fields?
A：Tokens can be used to explore the universe, help overcome current challenges on Earth, advance physical定律, assist in discovering diseases, optimize human actions, bring欢乐 to the world, and make美好生活触手可及。
Q：What are the new products and markets that NVIDIA's CEO, Jenson Wong, plans to discuss?
A：Jenson Wong plans to discuss new products and markets that open new markets for NVIDIA, as well as how NVIDIA will develop the ecosystem with great partners.
Q：How has NVIDIA's focus shifted over the years?
A：NVIDIA started out as a chip company aiming to create a new computing platform and introduced CUDA in 2006, which revolutionized computing. Over time, they realized a new computing approach required a reinvention of the technology stack, leading to the invention of new systems like the DGX one, which started the AI revolution.
Q：What are the key elements of the new computing approach that NVIDIA realized?
A：The new computing approach requires a reinvention of every layer of the technology stack, including new processors, software, and systems. This new way of doing software, now called artificial intelligence, necessitates many processors working together to serve queries for millions of people, fundamentally different from traditional data centers.
Q：What are the two types of networks that NVIDIA recognized were essential for high performance computing?
A：NVIDIA recognized two kinds of networks as essential for high performance computing: northbound and east-west networks. The northbound network is for managing control planes and connecting to the outside world, while the east-west network is for computers to talk to each other to solve problems, emphasizing the importance of efficient communication between processing units.
Q：How did NVIDIA transform a data center into a computing unit?
A：NVIDIA converted an entire data center into one computing unit by utilizing the entire data center as a unit of computing, running one job, and changing the operating system to Nvidia's CUDA, thus creating a significant shift in the way computing is organized and managed.
Q：What does NVIDIA's detailed roadmap for the next five years signify about the company's transformation?
A：NVIDIA's detailed roadmap for the next five years indicates its transformation into an essential infrastructure company. This reveals the company's strategic vision and the intention to build data centers around the world, as infrastructure is critical for the future and requires alignment with global infrastructure planning.
Q：How does NVIDIA view AI in relation to the internet and electricity?
A：NVIDIA views AI as the new type of infrastructure comparable to electricity and the internet. Just as the internet and electricity became essential social and business infrastructures, NVIDIA anticipates that AI will be integrated into every facet of life, requiring AI infrastructure in every company, region, industry, and country.
Q：What are the upcoming developments in AI infrastructure that NVIDIA CEO, Jenson Wong, promises?
A：Jenson Wong promises that the upcoming developments in AI infrastructure will fundamentally change how data centers are built and utilized. He suggests that these AI data centers will become essential factories, producing valuable tokens that are integral to various industries.
Q：How does AI integrate into NVIDIA's current and future products?
A：AI integrates into NVIDIA's current and future products by revolutionizing computer graphics with the introduction of AI-accelerated GPUs. It is also applied to areas such as physics simulations, 5G radio signal processing, and is poised to bring AI to other industries by enabling the development of AI factories or data centers.
Q：What is CUDA and how does it relate to AI and accelerated computing?
A：CUDA is a parallel computing platform and programming model created by NVIDIA. It enables developers to harness the power of NVIDIA GPUs for general-purpose computing, significantly accelerating applications in various fields, which includes AI and other compute-intensive tasks.
Q：How has the library ecosystem revolutionized computing and IT?
A：The library ecosystem revolutionized computing and IT by providing the deep learning and necessary libraries for training and inference, thus starting this revolution.
Q：What are the applications of Coup TA co dss and cool space for sparse structure simulators?
A：Coup TA co dss and cool space for sparse structure simulators are used for computational fluid dynamics, element analysis, and are important for the EDA and semiconductor industry.
Q：What is the significance of the fully accelerated radio access network stack?
A：The fully accelerated radio access network stack is significant as it achieves incredible performance for data rates per MHz and stands on par with state-of-the-art ASICs, enabling the introduction of AI for 5G.
Q：How does the new AI capability compare to traditional AI?
A：The new AI capability is more advanced, going beyond understanding patterns to generating from text to text, text to image, text to video, and more, showcasing the AI's ability to reason and solve problems it has not seen before.
Q：What is the concept of physical AI and how is it relevant to the next era of AI?
A：Physical AI refers to AI that understands physical properties such as inertia and friction, enabling it to reason about the physical world and solve problems related to it, which is crucial for the next era of AI.
Q：What is the purpose of Grace blackwall and how does it enhance AI capabilities?
A：Grace blackwall is a new computer system designed to support fast thinking and reasoning in AI, with the ability to scale up to create larger systems and scale out to connect multiple computers together, significantly enhancing AI capabilities.
Q：What is the difference between one-shot AI and thinking AI?
A：One-shot AI is when an AI generates a response based on data it has learned from, while thinking AI involves reasoning and thinking about different options and their benefits before producing an answer, signifying an ability to think and reason like a human.
Q：What are the characteristics of the new AI infrastructure being built?
A：The new AI infrastructure includes high power density, incredible memory and networking bandwidth, the ability to scale into large systems, and is referred to as an AI colosseum factory. It utilizes an entire rack with one hundred and twenty kilowatts of power, has one gigabyte of bandwidth, and is built using AI-specific technology.
Q：Why are the new data centers referred to as 'AI factories'?
A：The new data centers are referred to as 'AI factories' because they are designed to support large-scale AI operations, with a focus on efficiency and capacity for extensive computing power.
Q：What is the scale and financial investment of the AI factory mentioned?
A：The AI factory mentioned covers four million square feet and is likely to cost between sixty to eighty billion dollars. The computing part of IT, which is the electronics, is estimated to be forty to fifty billion dollars worth of investment.
Q：What complexities are involved in the technology used in AI factories?
A：The technology used in AI factories is extremely complicated, involving hundreds of chip processing steps, ultra-Violet lithography, and the assembly of multiple components including GPUs and interposers. The process requires careful integration and high precision, with robots working around the clock and a multitude of partners contributing to the effort.
Q：What is the purpose of the AI supercomputer being built in Taiwan?
A：The AI supercomputer being built in Taiwan is meant to provide a local infrastructure and ecosystem for AI research and development, as well as support scientific work in robotics and other fields that rely on advanced AI capabilities.
Q：What is Envylink Fusion and what does it enable?
A：Envylink Fusion is a product that enables the creation of custom AI infrastructure. It integrates NVIDIA GPUs, CPU, networking, and other necessary components to build powerful AI supercomputers. It allows for flexibility in integrating different types of accelerators and CPUs into the system and connects to existing AI ecosystems.
Q：Who will be able to use Envylink Fusion and how does it support their needs?
A：Envylink Fusion can be used by anyone who wants to build data centers, including NVIDIA, other companies that use GPUs, and individuals who might want to integrate specific ASICs or CPUs into their AI infrastructure. It supports their needs by providing a flexible and open ecosystem for integrating various computing elements into large-scale AI systems.
Q：What are the specifications and benefits of the DG X Spark computer?
A：The DG X Spark computer has a computational capacity of one petaflop and 128 GB of HBM memory. It also features 128 GB of LPDDR5x memory. The benefits of this computer include allowing developers, students, and researchers to conduct AI work locally without relying on cloud services, and it is expected to be ready for use in a few weeks.
Q：What is the DG X Station capable of handling in terms of AI models?
A：The DG X Station is capable of handling a one trillion parameter AI model.
Q：What is the primary purpose of the new generation of AI native computers?
A：The primary purpose of the new generation of AI native computers is to run the modern AI-native applications, which don't necessarily have to be compatible with traditional IT software or operate through a hypervisor.
Q：How are AI native computers expected to impact traditional enterprise computing?
A：AI native computers are expected to reinvent computing for traditional enterprise computing by introducing a new capability called 'agented AI,' which can perform tasks like digital marketing, research, and customer service, among others.
Q：What is 'agented AI' and how does it relate to traditional jobs?
A：'Agented AI' refers to digital agents that can work alongside humans to perform various tasks, such as digital marketing campaign managers, researchers, software engineers, customer service, and chip designers. They are essentially digital versions of traditional jobs.
Q：Why is there a shortage of labor by 2030 and how might digital agents help?
A：There is a projected shortage of labor by 2030 due to various factors limiting the world's ability to grow. Digital agents could help by assisting in developing better code and working more productively, effectively acting as digital employees.
Q：What are the key components of the new RTX Pro Enterprise server?
A：The key components of the new RTX Pro Enterprise server include a new motherboard with a switched network and Cx A chips, which are the most advanced networking chips in the world, allowing communication between GPUs at high speeds. The server supports a wide range of workloads, including those that require high through-put and low latency.
Q：How does the performance of the new AI models compare to previous models?
A：The new AI models have much higher performance compared to previous models. For example, Deep Segar One has four times the performance of the state-of-the-art model and is considered a significant breakthrough for the AI industry.
Q：What is the NVIDIA AI Data Platform and what are its benefits for businesses?
A：The NVIDIA AI Data Platform is a new type of storage platform that embeds meaning in unstructured data, making it suitable for AI querying. It is supported by a new query system called IQ, which accelerates retrieval and improves query results. This platform enables businesses to create specialized AI agents for decision-making and data analysis.
Q：What is AI ops and how does it differ from traditional operational roles in businesses?
A：AI ops is a new layer of software that parallels traditional operational roles like supply chain or HR, but for AI. It will manage, improve, and evaluate a family of AI agents within a company, optimizing the use and performance of these AI tools in the enterprise environment.
Q：What are the primary functions of AI models integrated into enterprise IT?
A：The primary functions of AI models integrated into enterprise IT include fine-tuning models, evaluating models, and securing models. These AI models are designed to enhance the capabilities of IT systems and to be deployed across various enterprise operations.
Q：Who are the companies that are working with AI Operations to create and integrate AI models into IT systems?
A：Companies like CrowdStrike, Data IQ, and Data Robots are working with AI Operations to create and integrate AI models into IT systems. Additionally, NVIDIA works with various partners, including those who create AI agents or 'robots,' and has partnered with DeepMind, Google DeepMind, and Disney Research to build an advanced physics engine for simulating robots.
Q：What is the significance of the new physics engine developed by NVIDIA and its partners?
A：The new physics engine developed by NVIDIA and its partners is significant because it is the world's most advanced and will be open source in July. This physics engine has the ability to deal with rigid and soft body simulation with high fidelity, which is essential for training AI robots in a virtual world that simulates the physical world.
Q：How does NVIDIA's simulation and AI integration platform work?
A：NVIDIA's simulation and AI integration platform works by using a combination of a simulation environment, a computer training system, and an ASIC runtime. The platform starts with a simulation environment called 'Isaac', which is used to train AI models. The trained models are then deployed into a human or robotic system, like self-driving cars or robots. NVIDIA provides the computing power through GPUs, a specialized operating system called 'Isaac OS', and pre-trained models to enable these systems to learn and operate effectively.
Q：What challenges does the lack of large-scale data pose to robot makers and how can AI amplify human demonstration systems in this context?
A：The lack of large-scale data poses challenges to robot makers as human demonstrations aren't scalable due to the limited number of hours available to developers. AI can amplify human demonstration systems by generalizing from a single demonstration and by using AI to expand the amount of data collected from human demonstrations, thus training more advanced robot models with minimal manual captures.
Q：How does AI facilitate the training of robots without requiring extensive manual data collection?
A：AI facilitates the training of robots by using AI to synthesize large amounts of data from human demonstrations. This process includes fine-tuning a general model with human demonstration data recorded by teleoperators. AI can then generate future world states (dreams) from images and new instructions, selecting the best ones for training. This helps robots learn a variety of new actions with minimal manual data capture.
Q：What is the potential economic impact of AI and robotics on the manufacturing industry?
A：The potential economic impact of AI and robotics on the manufacturing industry is enormous, with an estimated five trillion dollars of new plants being planned globally over the next three years. This is due to the reshoring and reindustrialization of manufacturing processes worldwide, and the necessity to build smarter, more efficient plants that utilize digital twins and AI technology.
Q：What is the role of digital twins in preparing for a robotic future?
A：Digital twins play a crucial role in preparing for a robotic future by creating a virtual representation of a physical entity or system. This includes digital twins of robots, equipment, and entire factories. They help in planning, testing, and training robots, and are integral to the development of AI for robotics, allowing for efficient and cost-effective deployment of robots in various environments.
NVIDIA Corp.
Follow
