NVIDIA neutered H100 GPU graphics function: 18,000 cores cut to 512 cores
Last night at the GTC 2022 conference, NVIDIA officially released the H100 GPU accelerator card, also the much-anticipated Hopper new architecture, skipping 5nm directly on TSMC 4nm process, 80 billion transistors, power consumption up to 700W, very good and powerful performance in all aspects.
The full version has 8 groups of GPC (graphics processor clusters), 72 groups of TPC (texture processor clusters), 144 groups of SM (streaming multiprocessor units), and each group of SM has 128 FP32 CUDA cores, a total of 18,432.
Compared to the 10,000 CUDA cores of the current Ampere GPU architecture, Hoper's 18,000 CUDA cores are a big improvement, but you shouldn't expect the gaming performance of H100, because NVIDIA has cut it to the bottom this time, and the neutering is very powerful.
From NVIDIA's white paper, we can confirm that H100 has cut a lot of GPU-related functions. No matter PCIe 5.0 or SMX version of H100 core, only 2 groups of TPC units can support graphics operations, including vector, geometry and pixel rendering.
2 groups of TPC units is also 4 groups of SM units, a total of 512 CUDA cores are able to run the game, compared to the full 18,000 cores is insignificant, the performance is only equivalent to 1/36 of the full version of the H100 core, or about 3%, 97% of the game performance is gone.
NVIDIA explained that H100 is made for AI, HPC and data analysis, and is not designed for gaming.
Considering the performance improvements of H100 in AI and HPC, it is understandable that NVIDIA has cut a lot of gaming features in order to reduce the design difficulty, after all, the accelerator card will not be used to play games.
For gamers is the Ada Lovelace architecture, which was previously reported to have up to 18432 stream processors, but it will heavily neuter the compute unit and keep the complete graphics and light tracing unit, reversing the design concept with the H100.