NVIDIA has unveiled the GB200 Grace Blackwell superchip.
Powered by six new technologies, the platform works to enable organizations to build and run real-time generative artificial intelligence, NVIDIA said Monday.
The first of the six technologies underpinning the superchip are Blackwell-architecture GPUs called the B200. Manufactured using TSMC’s 4NP process, a B200 GPU comprises two-reticle limit GPU dies connected by a chip-to-chip link with a bandwidth of 10 Terabytes per second.
The second key technology is the fifth-generation NVIDIA NVLink, which links a pair of B200 GPUs with the NVIDIA Grace CPU. Combined, these three processing units make up a GB200 Grace Blackwell superchip.
The four other technologies that underpin the GB200 are a second-generation Transformer Engine; a dedicated reliability, availability and serviceability engine; advanced confidential computing capabilities that work to secure AI models; and a dedicated decompression engine.
Blackwell is expected to be adopted by numerous organizations, including Google, Microsoft, Meta, OpenAI, Amazon Web Services and Oracle.