On the evening of June 2nd Beijing time, on the eve of the opening of the Computex Taipei International Computer Show, NVIDIA founder and CEO Huang Renxun gave a keynote speech titled "Unleashing the New Industrial Revolution". He spent two hours elaborating on his understanding of "GPU accelerating AI computing revolution" and "AI factory".
Facing a total audience of 6500, he publicly showcased the company's strongest performance product, the GB200 motherboard, on stage. This motherboard is equipped with two Blackwell B200 GPU chips and one Grauce CPU chip, combined into one through its interconnection technology. The product made its debut at the Nvidia GTC conference in March this year. According to Huang Renxun, Blackwell products have been fully put into production and an enhanced version of the Blackwell Ultra AI chip will continue to be launched next year.
Huang Renxun believes that as the CPU expansion speed slows down, it will eventually come to a basic stop. However, the amount of data that needs to be processed continues to grow exponentially, ultimately leading to computational inflation and an increase in computational costs.
The GPU will be the key to changing all of this. Compared to Nvidia's Pascal GPU released in 2016, Blackwell GPU's computing power (corresponding to artificial intelligence floating-point computing power) has increased by 1000 times, almost surpassing Moore's Law's growth during its optimal period.
While computing power is improving, costs are constantly decreasing. Previously, using Pascal products to train a GPT-4 model would consume up to 1000 gigawatt hours of energy (1 gigawatt hour equals 1 million kilowatt hours), while using Blackwell products only required 3 gigawatt hours.
According to Huang Renxun's consistent belief that "the more you buy, the more you save", the combination of GPU and CPU achieves up to 100 times faster computing while only increasing power consumption by three times. When converted, the average unit performance is 25 times higher than using CPU alone, significantly achieving energy-saving and cost reducing effects.
During the speech, Huang Renxun personally moved a larger DGX server. This server is connected to 8 Blackwell GPUs and uses air cooling technology for heat dissipation. At present, this product is used by AI servers that are being purchased globally and is also a core device in major data centers.
According to Huang Renxun's vision, as artificial intelligence models continue to expand, the requirements for GPU performance will also become higher. Traditional data centers will evolve into "AI factories", and the cornerstone of factories is super GPUs.
Super GPUs mainly follow the path of interconnection and clustering, connecting more and more GPUs through Nvidia's NVLink interconnection technology to form a single, ultra large scale GPU. In the MGX data center solution showcased on site, a "super GPU" will accommodate a total of 72 GPUs, improving performance by 9 times but only increasing power by 10 times, capable of accommodating AI large language models with parameters reaching billions of dollars. Huang Renxun believes that this super GPU is the true form of GPU.
According to its introduction, the "AI factory" composed of numerous BlackWell system clusters contains 32000 GPUs, with an area larger than a basketball court.
In this speech, Huang Renxun also emphasized the annual upgrade roadmap of Nvidia's chip products. The company will adhere to the route of "iterating the new architecture once a year" in the future, breaking Moore's Law, which used to be updated every two years. The new architecture iterated every year will also cover the entire data center GPU product line.
He also publicly announced for the first time the names of the next generation GPU and CPU architecture. The next generation CPU is named "Vera" and the GPU is named "Rubin", inspired by Vera Rubin, a pioneer in dark matter research in the universe and an American female astronomer.
As early as a month ago, there were rumors in the industry that Nvidia's next generation R series AI chips may adopt TSMC's 3-nanometer process technology, use CoWoS-L packaging technology, and be equipped with the new generation HBM4 high bandwidth memory technology. It is expected to achieve large-scale production in the fourth quarter of 2025.
In addition, Huang Renxun also introduced the implementation of a series of new products and services by Nvidia in this speech, including NIM cloud native microservices that help enterprises deploy AI models, a new Ethernet Spectrum X specifically designed for AI, and Nvidia's robot technology platform Isaac, among others.
The CEO also specifically mentioned the combination of artificial intelligence and robots, believing that the application of robot technology goes far beyond humanoid robots. "Mechanization will become the norm, factories will fully realize automation, robots will work together to manufacture a series of mechanized products."