Although compute GPUs like Nvidia's H100 formally belong to the category of graphics processing units, they can barely render graphics as they do not have enough special-purpose hardware. As it turns out Nvidia's H100, a card that costs over $30,000 performs worse than integrated GPUs in such benchmarks as 3DMark and Red Dead Redemption 2, as discovered by Geekerwan.
Nvidia's H100 card is based on the company's GH100 processor with 14,592 CUDA cores that support a variety of data formats used for AI and HPC workloads, including FP64, TF32, FP32, FP16, INT8, and FP8. By contrast, Nvidia's consumer GPUs, such as Nvidia's AD102, only properly support FP32. Meanwhile, GH100 only has 24 raster operating (ROPs) units and does not have display engines or display outputs. Furthermore, Nvidia does not optimize Hopper drivers for gaming applications.
But apparently it is still possible to make Nvidia's H100 render graphics and even support ray tracing. Only it renders graphics rather slowly. One H100 board scores 2681 points in 3DMark Time Spy, which is even slower than performance of AMD's integrated Radeon 680M, which scores 2710.
But running games on a card that costs over $30,000 does not make a lot of sense and Nvidia certainly did not design GH100 for rendering graphics. While Nvidia's GH100 has some graphics specific hardware inside, it is not made to offer any substantial performance in games, which is why it is slower than AMD's integrated Radeon 680M.
Although Nvidia's flagship compute GPU is not meant for graphics, it outperforms everything in datacenter AI and HPC applications and this is exactly what it is made for.
