NVIDIA has revised the performance of the H100 accelerator upwards based on the GH100 “Hopper” GPU, an 80 billion transistor monster offered in SXM5 and PCI Express 5.0 format.
Originally NVIDIA had indicated for the SXM5 solution FP64 compute performance up to 30 TFLOPswhile the FP32 value reached 60 TFLOPs. Now, on its website, the company reports values respectively of 34 and 67 TFLOPs. As for the variant PCIeinstead, we pass from the previous data of 24 and 48 TFLOPs to the updated values of 26 and 51 TFLOPs.
The other values related to the capacities of the Tensor cores also rise, as well as the memory bandwidth of the variant SMX5with a leap from 3 to 3.35 TB / s. There is no doubt behind the improvements in computing power an increase in the operating frequency, estimated at about 200 MHz.
As we reported a few days ago, the production of H100 accelerators has finally entered the “volume” phase and this has evidently led NVIDIA and TSMC to evaluate the possibility of set clocks higher than originally supposed. It is not known whether the company has been too conservative in the past or whether production is doing better than expected.
The H100 GPU is built with process 4N, described as an NVIDIA-optimized version of TSMC’s 5nm (N5). The same technology is the basis of the Ada Lovelace GPUs on board the GeForce RTX 4000 video cards, arriving on the market starting October 20 with the GeForce RTX 4090.