On March 19th, at the GTC developer conference last night, NVIDIA officially unveiled the next-generation Blackwell GPU architecture designed for artificial intelligence models, along with the new generation AI accelerator card based on this architecture called GB200. NVIDIA’s CEO, Huang Renxun, pointed out: “Blackwell is not just a chip, but a platform.”
(Image Source: GTC 2024)
In other words, GB200 is actually an AI acceleration platform composed of two Blackwell B200 GPUs and a Grace CPU. Each B200 GPU contains 208 billion transistors. Compared to the H100, the computing power of GB200 has increased by 6 times, and when handling tasks in specific multimodal areas, its computing power can reach up to 30 times that of the H100.
(Image Source: GTC 2024)
Although the increase in computing power is important, it is not the main highlight of this chip update. What is more crucial is the significant decrease in power consumption and cost. The previous generation H100 AI accelerator card had a peak power consumption of 700 watts, and training a model with 18 trillion parameters required 8000 Hopper GPUs and 15 megawatts of power.
Compared to the H100, the new generation GB200 AI accelerator card requires only 1/25 of the cost and power consumption, meaning that the same task can be accomplished with 2000 Blackwell GPUs at only 4 megawatts of power.
In simple terms, the performance of GB200 is 7 times that of the H100, with lower costs and power consumption.
(Image Source: GTC 2024)
Experts had previously predicted that with the widespread deployment of H100, its total power consumption would rival that of a major U.S. city or even surpass some small European countries. The significant performance improvement and efficiency enhancement of GB200 mean that data centers and AI applications will be able to run more complex and demanding computing tasks at lower costs, thereby accelerating the development and application of AI technology.
Undoubtedly, from smartphones over a decade ago to AI today, with technology advancing rapidly over these years, AI has experienced explosive growth, leading to higher computational requirements. NVIDIA’s progress in the field of AI has already outpaced many competitors, and the introduction of GB200 is likely to compel other companies to expedite the development and release of their own high-performance AI accelerator cards, intensifying market competition.
(Image Source: GTC 2024)
Moreover, with NVIDIA also launching an innovative software called NIM and a hardware and software platform for manufacturing humanoid robots, deploying artificial intelligence will become even easier. This provides yet another reason for customers to continue using NVIDIA chips in increasingly competitive fields.
(Image Source: GTC 2024)
In the long run, this will drive technological innovation, offer users more and better choices, and propel the entire industry forward.
In conclusion, the significant performance difference between GB200 and its predecessors may prompt data centers and enterprise customers to upgrade their hardware to take advantage of the benefits of new technology. This also implies that in fields requiring extensive data processing and complex computations, the application of AI technology across various industries will further accelerate.