Nvidia’s H100 Hopper Compute GPU Benchmarked in Video games, Discovered Missing
Though compute GPUs like Nvidia’s H100 formally belong to the class of graphics processing models, they’ll barely render graphics as they don’t have sufficient special-purpose {hardware}. Because it seems Nvidia’s H100, a card that prices over $30,000 performs worse than built-in GPUs in such benchmarks as 3DMark and Purple Useless Redemption 2, as found by Geekerwan.
Nvidia’s H100 card is predicated on the corporate’s GH100 processor with 14,592 CUDA cores that help a wide range of knowledge codecs used for AI and HPC workloads, together with FP64, TF32, FP32, FP16, INT8, and FP8. Against this, Nvidia’s shopper GPUs, resembling Nvidia’s AD102, solely correctly help FP32. In the meantime, GH100 solely has 24 raster working (ROPs) models and doesn’t have show engines or show outputs. Moreover, Nvidia doesn’t optimize Hopper drivers for gaming functions.
However apparently it’s nonetheless doable to make Nvidia’s H100 render graphics and even help ray tracing. Solely it renders graphics moderately slowly. One H100 board scores 2681 factors in 3DMark Time Spy, which is even slower than efficiency of AMD’s built-in Radeon 680M, which scores 2710.
However working video games on a card that prices over $30,000 doesn’t make a whole lot of sense and Nvidia actually didn’t design GH100 for rendering graphics. Whereas Nvidia’s GH100 has some graphics particular {hardware} inside, it’s not made to supply any substantial efficiency in video games, which is why it’s slower than AMD’s built-in Radeon 680M.
Though Nvidia’s flagship compute GPU shouldn’t be meant for graphics, it outperforms every thing in datacenter AI and HPC functions and that is precisely what it’s made for.
