NVIDIA updated its website with a list of H100 accelerator characteristics on the Hopper architecture. The company provided more accurate data on the performance of these solutions, which is likely to be linked to the increase in the tactical frequency of graphic processors.
So the NVIDIA H100 SXM output in single accuracy operations from 30 to 34 Tflops respectively. In turn, the PCIe version H100 added the same items from 24 to 26 Tflops and from 48 to 51 Tflops. NVIDIA also updated data on accelerator performance in operations INT8, FP8, FP16, BFLOT16.
As the VideoCardz portal indicates, the new NVIDIA H100 accelerator performance data may indicate that their graphic processors will receive a higher tactical frequency than previously planned. For example, in the SXM version, the frequency of the graphic chip with 16896 CUDA kernels could grow from 1,775 to at least 1982 MHz. Since the graphic processors of these accelerators are made from the TSMC N4 process, there should be a reserve for the acceleration.
It will be recalled that the NVIDIA H100 accelerations are intended for the data processing centre segment. They will be produced both in the PCI Express 5.0 extension maps with a TDP of 350 W and in the SXM5 form factor with the declared TDP 700 W. As part of the GTC 2022 conference, the head of the company Jensen Huang confirmed that the supplies of NVIDIA H100, as well as the NVIDIa DGX workstations, would start in the first quarter of 2023.