NVIDIA officially released the specifications of the Hopper H100 GPU, and the performance is more powerful than expected. Earlier this year, NVIDIA announced the Hopper H100 GPU for AI data centers, capable of up to 60 TFLOPs and 30 TFLOPs FP64.
However, as the release date approaches, NVIDIA has updated its graphics card specifications again, indicating another improvement in performance. One reason the performance numbers are boosted is that when the chips go into production, GPU manufacturers can finalize the numbers based on actual clock speeds. NVIDIA is likely to use conservative clock frequencies to provide preliminary performance figures, and as production is in full swing, the company finds that the chip can deliver better clock frequencies.
At GTC last month, NVIDIA confirmed that their Hopper H100 GPUs are in full production, with partners launching the first wave of products this October. It also confirmed that the global rollout of Hopper will consist of three phases, the first of which will be pre-orders for NVIDIA DGX H100 systems, as well as free labs directly from NVIDIA to customers, including systems such as Dell’s Power Edge servers, which are now available at Nvidia Launchpad.
The NVIDIA Hopper GH100 GPU consists of a huge 144 SM (Streaming Multiprocessor) chip layout with a total of 8 GPCs. These GPCs have a total of 9 TPCs, and each TPC is further composed of 2 SM units. This gives us 18 SMs per GPC, and in the full 8 GPC configuration, we have 144. Each SM consists of up to 128 FP32 units, which should give us a total of 18,432 CUDA cores.
