Contrary to rumors that the NVIDIA H100 on the Hopper architecture base will use TSMC N5, NVIDIA announced that its last accelerator will use a special TSMC N4 process. Rumors were erroneous and about the number of transistors. This monolithic graphics processor consists of 80 billion transistors. The chip has up to 16896 CUDA FP32 cores for the SXM and 14592 variant for the PCIE-based model.
NVIDIA H100 has HBM3 memory with a 3 TB / s bandwidth, which is 1.5 times more than that of A100. This new generation accelerator is equipped with 80 GB of memory with high bandwidth. However, the technology will depend on the option: the SXM model has HBM3 with a rating of 3 TB / s, while the H100 graphics processor based on the PCIE has HBM2E with a 2 TB / s rating.
As for performance, NVIDIA declares that computing power is 3 times higher in FP64, TF32, FP16 and 6 times higher in FP8 than A100.
The accelerator will use the PCIE GEN5 or SXM form factor. The latter will have a TDP 700 W, exactly 300 W more than that of A100.
On March 21, NVIDIA also launched two GRACE processor platforms based on ARM. One of them is equipped with a GRACE processor in combination with the Hopper graphics processor with 600 GB of memory, and the other is equipped with two grace processors with a total number of kernels 144. It will be equipped with the LPDDR5X memory. Both Grace Superchips will be available in the first half of 2023.