On Tuesday morning Beijing time, the most important agenda item at NVIDIA GTC 2024 conference, the keynote speech by founder and CEO Huang Renxun, had just ended. As expected by the market, the global capital market has just seen new computing chips/servers, as well as a bunch of AI software applications.
As usual, as of the time of publication, Nvidia has released 40 press releases since Huang Renxun began speaking. This article will focus on summarizing some key developments this morning for investors to refer to.
Larger GPU - Blackwell architecture as scheduled
Although the entire market knew beforehand that a new flagship computing power GPU would be introduced today, Huang Renxun did not explain the name clearly in his speech - only stating the launch of a larger and stronger Blackwell architecture GPU, which caused media chaos at one point. But according to the data on the official website, today Lao Huang should be holding a B200 chip, and the website also lists the existence of a B100 chip in the Blackwell architecture. Nvidia has not disclosed the selling price, only stating that it will ship to its partners within the year.
Nvidia has disclosed that the new B200 chip has 208 billion transistors and is manufactured using TSMC's customized 4NP process. It is worth mentioning that this chip connects two dies into a unified GPU, and the communication speed between dies can reach 10TB/s. As expected, this chip uses 192GB of HBM3E memory.
The GB200 Grace Blackwell superchip is a combination of two B200 chips (four dies) and a Grace CPU. Compared to H100, the performance of the large language model has been improved by 30 times, while the energy consumption is only one 25th.
Lao Huang gave an example in his speech that to train a GPT model with 1.8 trillion parameters, it would require 8000 Hopper GPUs, consume 15 megawatts of electricity, and run continuously for 90 days. But if using the GB200 Blackwell GPU, only 2000 cards are needed, and running for 90 days also consumes only a quarter of the electricity. Not only training, but also the cost of generating tokens will be significantly reduced.
In conjunction with this new set of chips, Nvidia has also launched the fifth generation of new NVLink chips, as well as a series of products such as GB200 NVL72 servers, X800 series network switches, and the next-generation artificial intelligence supercomputer NVIDIA DGX SuperPOD.
New way to develop software: NIM microservices
After discussing the hardware updates, Huang Renxun also devoted the remaining time to the software ecosystem. In addition to the digital twin of Earth's climate and pharmaceutical development AI, Nvidia has also launched a series of "microservices" in AI Enterprise 5.0, including simplifying NIM for enterprises to deploy AI models into production environments.
Huang Renxun said, "In the future, companies will no longer need to write software, but will assemble AI models, present tasks to them, provide examples of work products, review plans, and intermediate results."
Nvidia stated that NIM microservices simplify the deployment process of AI models by packaging algorithms, optimizing systems and operations, and adding industry standard APIs. This allows developers to integrate NIM into existing applications and infrastructure without the need for extensive customization or expertise.
Digital twin support for Vision Pro
Nvidia also announced on Monday that Omniverse Cloud now allows developers to stream their industrial scenes from content creation applications to Nvidia's Graphics Delivery Network (GDN), allowing advanced 3D experiences to be transmitted to Apple Vision Pro.
This new workflow combines the high-resolution display of Apple Vision Pro with Nvidia's cloud rendering to provide a spatial computing experience with only devices and Internet connections.
There are also many scattered official announcements
In the semiconductor field, Nvidia announced that TSMC and Synopsys will invest Nvidia's computing lithography platform CuLitho in the production of advanced chips.
In the telecommunications field, Huang Renxun announced a research cloud called NVIDIA 6G, which is a platform driven by generative artificial intelligence and Omniverse technology, aimed at promoting the development of the next generation of communication.
In the field of transportation, BYD, the world's largest electric vehicle company, will adopt NVIDIA's centralized in vehicle computing platform DRIVE Thor to develop the next generation of electric vehicles. In addition, BYD will also use Nvidia's infrastructure for autonomous driving model training, as well as Nvidia Isaac to design/simulate intelligent factory robots.
Robots are also the final stage of the entire speech. Huang Renxun announced multiple software programs to assist in the development of robot technology. This includes the Isaac Perceptor software development toolkit, which involves multi camera visual mileage measurement, 3D reconstruction, and depth perception. There is also Isaac Manipulator - a library for robot arm perception, path planning, and kinematic control. Finally, he also announced a project called GR00T, which is a universal foundational model for humanoid robots aimed at driving the company's breakthroughs in robotics technology and embodied intelligence.
Accompanied by a pair of Disney robots Orange and Green using Nvidia Jetson chips, the entire press conference came to an end.