Nvidia has already introduced the Blackwell architecture and accelerators based on it. At the same time, not all features and features were revealed to us at the time of the announcement. Now, Nvidia senior vice president and GPU architect Jonah Albe and vice president of hyperscaling and HPC Ian Buck have filled in some of the gaps.

A very important point that Nvidia did not explain during the announcement is that Blackwell is a completely new architecture, very different from Hopper. By the way, this may mean that RTX 50 game adapters will be very different from RTX 40.

As we mentioned before, Blackwell has a second-generation Transformer Engine that provides support for FP4 and FP6 compute, which is not available in other Nvidia adapters. This is necessary for certain scenarios of working with artificial intelligence and has not been implemented before because such a format was not needed. However, performance in double-precision mode (FP64) improved only by 32% compared to Hopper. This is because such a mode is not required for AI-related calculations.

At first, Nvidia did not directly name the new GPU during the announcement, which caused some confusion. Apparently the full GPU is still called B200. This chip is based on the GB200 Grace Blackwell Superchip, and the TDP of the GPU alone reaches 1200W! As a result, the GB200 Grace Blackwell Superchip can consume up to 2.7 kW of power.

There is also the B200 version used by the DGX and HGX platforms. The TDP of such a GPU is 1000 W and provides almost 90% of the performance of the full version. At the same time, its differences from the full GPU are still unknown. And then there’s the Blackwell B100, a 700W TDP model that offers around 70% of the performance of the full B200. Interestingly, Nvidia says that it may introduce a single-chip Blackwell GPU in the future instead of the current two Blackwell GPUs. That is, with exactly half the productivity.