Nvidia’s recent GPU technology conference showcased unprecedented advances in AI computing with the unveiling of the world’s most powerful chip, the GB200.
The newly introduced B200 GPU has an impressive capacity of 20 FP4 petaflops, thanks to a staggering 208 billion transistors packed into the chip Notably, the GB200 exhibits a remarkable jump in performance over its predecessor the H100, which exhibits remarkable efficiency in LLM mathematical work, and 25 concurrent power consumption -The properties are remarkably reduced In the GPT-3 LLM design, the GB200 is seven times more active than H100.
To illustrate shows impressive capabilities, consider that previously 8,000 hopper GPUs for about 15 MW were required to train an image with 1.8 billion parameters. However, only Blackwell GPU 2000 can do the same work with only 4 MW.
To further increase performance, Nvidia has developed a new network switch chip with 50 billion transistors. This chip facilitates seamless communication between 576 GPUs, providing a two-way bandwidth of 1.8 TB/s potential. This enhancement solves a major problem with previous systems, where 60% of the time was spent communicating and only 40% of the time was spent doing actual computation on 16 GPUs.
Nvidia offers comprehensive solutions for enterprises, exemplified by the GB200 NVL72, which offers 36 CPUs and 72 GPUs in a single water-cooled rack Furthermore, the DGX Superpod with DGX GB200 integrates eight such systems, and consists of 288 CPUs, 576 GPUs, and an impressive configuration of 240TB of memory there
Leading technology companies such as Oracle, Amazon, Google, and Microsoft already plan to integrate NVL72 racks into their cloud services, underscoring industry recognition of NVIDIA’s advances in AI computing.
The GPU architecture underpinning the Blackwell B200 GPU is poised to become the cornerstone of the upcoming RTX 5000 series, further solidifying Nvidia’s position at the forefront of AI and GPU technology.