NVIDIA CEO: “We created processors for the generative AI era”

At the largest GTC conference to date, NVIDIA founder and CEO Jensen Huang brought about the releases of NVIDIA Blackwell, NIM microservices, Omniverse Cloud API, etc.

Generative AI promises to revolutionize every industry it touches—and mastering the technology is key to meeting the challenge.

NVIDIA founder and CEO Jensen Huang launched the new Blackwell computing platform and outlined the major advances that improvements in computing power can bring to software, services, robotics, medical technology, and more.

"Accelerated computing has reached a tipping point, and general-purpose computing has run out of steam," Huang told his keynote audience at Silicon Valley's cavernous SAP Center arena. GTC gathered more than 11,000 attendees onsite and tens of thousands of online viewers.

"We need a completely new way of computing - so that we can continue to scale, continue to reduce the cost of computing, and continue to do more and more calculations while ensuring sustainability. Compared with general-purpose computing, accelerated computing makes Every industry can use significant speedups.”

Standing in front of a 40-foot-tall 8K screen the size of a tennis court, Huang spoke to a packed crowd of CEOs, developers, AI enthusiasts and entrepreneurs.

In order to promote the goal of large-scale upgrading of the world's AI infrastructure, Huang Renxun released the NVIDIA Blackwell platform, which has the ability to generate trillion-parameter large language models in real time.

Jen-Hsun Huang introduced NVIDIA inference microservices, NVIDIA NIM. This is a new way of packaging and delivering software that connects developers to hundreds of millions of GPUs to deploy a variety of customized AI.

Huang Renxun also introduced the Omniverse Cloud API, which can provide advanced simulation capabilities and introduce AI into the physical world.

At the end of the speech, Huang Renxun gave a wonderful demonstration, explained the partnership ecology with some large enterprises, and also introduced in detail more than twenty releases to elaborate on his vision, thus completing the grand launch of GTC 2024. period.

The GTC conference has been held for 15 years. It was originally held in a local hotel banquet room and has now developed into the most important AI conference in the world. This conference is the first time in the past five years that it has resumed offline.

This year's conference features more than 900 sessions, including an expert panel on Transformer moderated by Jen-Hsun Huang, which will feature conversations with eight of the first pioneers to develop the technology. In addition, there will be more than 300 presentations and more than 20 technical workshops.

GTC is an AI event covering almost all fields. In the opening speech, Refik Anadol, the world's leading AI artist, gave a stunning performance, showing a huge real-time AI data sculpture, with undulating vortices of green, blue, yellow and red on the screen, tumbling, intertwining and... Scattered.

Huang began his speech by explaining that the rise of multimodal AI—the ability to handle diverse data types that are handled by different models—gives AI greater adaptability and capabilities. By adding parameters, these models can handle more complex analysis tasks.

But it also means the need for computing power rises significantly. As these collaborative, multimodal systems become more complex (with trillions of parameters), the need for advanced computing infrastructure increases.

“We need a bigger model,” Huang said. “We’re going to train it using multimodal data, not just text on the Internet. We’re going to train it using text and images, graphs and charts, just like we Just like learning from television, it also requires learning from massive amounts of video.”

Next generation accelerated computing

Huang Renxun said: "We need a bigger GPU." The Blackwell platform is built to address this challenge. He took a Blackwell chip out of his pocket and held it up next to the Hopper chip, which looked smaller.

The new architecture is named after David Harold Blackwell, a mathematician at the University of California, Berkeley. He specializes in game theory and statistics and is the first black scholar elected to the National Academy of Sciences. The new architecture surpasses the NVIDIA Hopper architecture launched two years ago.

Blackwell's FP8 performance in single-chip training is 2.5 times that of its previous generation architecture, and its FP4 performance in inference is 5 times that of its previous generation architecture. It features fifth-generation NVLink interconnect, is twice as fast as Hopper, and is scalable to 576 GPUs.

The NVIDIA GB200 Grace Blackwell super chip connects two Blackwell NVIDIA B200 Tensor Core GPUs to the NVIDIA Grace CPU via the 900GB/s ultra-low-power NVLink chip-to-chip interconnect.

Huang Renxun held up a circuit board with the system and said: "This computer is the first of its kind to be able to accommodate so much calculation in such a small space. Because its memory is coherent, it feels like a One big happy family working together on an app.”

For maximum AI performance, GB200-based systems can connect with NVIDIA Quantum-X800 InfiniBand and Spectrum-X800 Ethernet platforms announced today, which provide advanced networking at speeds up to 800Gb/s.

"In this way we can save a lot of energy, network bandwidth and time." Huang Renxun said, "The future will be generative, which is why this is a completely new industry. Our computing methods are fundamentally different, so NVIDIA is generative A new processor has been created for the AI ​​era.”

To expand Blackwell's scale, NVIDIA built a new chip called NVLink Switch. Each chip can connect four NVLinks at 1.8 TB per second, eliminating traffic congestion by reducing intra-network traffic.

NVIDIA Switch and GB200 are what Jen-Hsun Huang calls "one giant GPU" and are key components of NVIDIA GB200 NVL72. NVIDIA GB200 NVL72 is a multi-node, liquid-cooled, rack-scale system that leverages Blackwell to provide powerful computing for trillion-parameter models, achieving 720 petaflops of AI training performance and 1.4 exaflops of AI inference performance in a single rack.

The machine contains 600,000 parts and weighs 3,000 pounds. When introducing this machine, Huang Renxun said: "At this moment, there are maybe only three exaflop machines on the earth. And this is 1 exaflop in a single rack AI systems.”

In addition, NVIDIA also launched a new generation of more powerful AI supercomputers - NVIDIA DGX SuperPOD powered by NVIDIA GB200 Grace Blackwell superchip, which can be used to process trillion-parameter models with sustained uptime to achieve ultra-large scale Generative AI training and inference workloads.

Featuring a new high-efficiency liquid-cooled rack-scale architecture, the new DGX SuperPOD is built on the NVIDIA DG GB200 system, delivering 11.5 exaflops of AI supercomputing power and 240 TB of fast memory at FP4 precision, and is expandable with additional racks.

“In the future, data centers will become AI factories,” Huang said. “The mission of an AI factory is to create revenue and intelligence.”

Blackwell has been welcomed by all walks of life and has received support from multiple industry leaders, including Alphabet and Google CEO Sundar Pichai, Amazon CEO Andy Jassy, ​​Dell CEO Michael Dell, Google DeepMind CEO Officer Demis Hassabis, Meta CEO Mark Zuckerberg, Microsoft CEO Satya Nadella, OpenAI CEO Sam Altman, Oracle Chairman Larry Ellison and Tesla and xAI CEO Elon Musk.

Blackwell is being adopted by major global cloud service providers, leading AI companies, systems and server vendors, as well as regional cloud service providers and telecommunications companies around the world.

"The entire industry is preparing for Blackwell," Huang said.

A new way to create software

Huang Renxun said that generative AI has changed the way applications are written.

He explained that companies of the future will focus on assembling AI models, giving them tasks, giving examples of working products, and reviewing plans and intermediate results rather than writing software.

The NVIDIA NIM package is built on NVIDIA's accelerated computing libraries and generative AI models.

"How do we build software in the future? It's unlikely that you're going to write it from scratch, and it's unlikely that you're going to write a bunch of Python code or anything like that," Huang said. "It's likely that you're going to just assemble an AI team."

These microservices support industry-standard APIs, are easy to connect, work on NVIDIA's vast CUDA installation base, are re-optimized for new GPUs, and are constantly scanned for security vulnerabilities and threats.

Huang Renxun said that customers can use off-the-shelf NIM microservices, or NVIDIA can build exclusive AI and AI assistants for them, providing specialized training for the model expertise required by specific companies to help you create valuable new services.

"The enterprise IT industry is sitting on a 'gold mine,'" Huang said. "They have all these amazing tools (and data) they've created over the years. If they can turn this 'gold mine' into an AI assistant , can provide users with more possibilities.”

Leading technology companies are already taking action. Jen-Hsun Huang detailed how NVIDIA helps Cohesity, NetApp, SAP, ServiceNow and Snowflake build AI assistants and virtual assistants. All walks of life are also joining the ranks.

In telecommunications, Huang announced the launch of the NVIDIA 6G Research Cloud, a generative platform powered by AI and Omniverse designed to drive the next era of communications. It is built using NVIDIA's Sionna neural radio framework, NVIDIA Aerial CUDA accelerated radio access network and NVIDIA Aerial Omniverse Digital Twin for 6G.

In the field of semiconductor design and manufacturing, Huang Jensen announced that NVIDIA is collaborating with TSMC and Synopsys to put its breakthrough computational lithography platform cuLitho into production. The platform will accelerate the most computationally intensive workloads in semiconductor manufacturing by 40-60 times.