June 16, 2024

NVIDIA Roadmap: NVIDIA offers the Blackwell Ultra, Rubin, and Rubin Ultra devices

At a keynote at Computex in Taipei, NVIDIA provided an outlook on its future data center roadmap. The Blackwell GPU with Blackwell architecture introduced this year is complemented by the Blackwell Ultra. This is likely a B200 accelerator with a Blackwell GPU, but the HBM3e has 12 stacks instead of eight stacks as planned in the first step. This increases the memory capacity from 192 to 244 GB. The memory bandwidth remains the same. The B100 and B200 accelerators make up the Blackwell platform.

And now the next generation has a name we’ve already heard: Ruby. There will be GPU Ruby which will be based on HBM4. In addition, there will be up to eight accelerators directly connected to each other.

There will also be a Rubin Ultra for the Rubin platform. HBM4 will also be used here. In addition, Rubin Ultra will go from the current typical system of eight accelerators on the motherboard to 12 accelerators.

While Hopper and Blackwell rely on the Grace CPU in their superchips, a new GPU called Vera will be used for Robin. GB200 (Grace Hopper) will be followed by VR200 (Vera Rubin).

See also  Capcom says Pragmata is making 'steady' strides