This is a completely new architecture
Nvidia has already presented the Blackwell architecture and accelerators based on it. At the same time, during the announcement, not all features and characteristics were revealed to us. Now, Nvidia Senior Vice President and GPU Architect Jonah Albe & Vice President of Hyperscale and HPC Ian Buck have filled in some of the gaps.
A very important point that Nvidia did not explain at the time of the announcement is that Blackwell — this is a completely new architecture that is very different from Hopper. By the way, this may mean that the RTX 50 gaming adapters will be very different from the RTX 40.
As we have already noted, Blackwell has introduced the second generation Transformer Engine, which provides support for FP4 and FP6 calculations, which are simply not available in other Nvidia adapters. This is needed for certain scenarios of working with AI, and previously this format was not required, so it was not implemented. But performance in double precision mode (FP64) increased relative to Hopper by only 32%. Simply because for calculations related to AI, such a mode is not particularly needed.
Initially, during the announcement, Nvidia did not directly name the name of the new GPU, which caused some confusion. Apparently, the full GPU is still called B200. This chip powers 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 a B200 version used by the DGX and HGX platforms. This GPU has a TDP of 1000 W and provides almost 90% of the performance of the full version. However, the differences from a full GPU are still unknown.
And then there's the Blackwell B100, which is a 700W TDP variant offering about 70% of the performance of the full B200.
Interestingly, Nvidia says that in the future it may release a Blackwell GPU with one die instead of the current two. That is, the productivity is exactly half as much.