NVIDIA Hopper H100 is one of the most powerful graphics cards in the world, however its architecture is fully optimized for supercomputing in scientific studies and artificial intelligence applications, therefore it is not a good option for gamers who would be willing to spend more than US $30K for the GPU.
This Monday (19), enthusiast and content producer geekerwan released a video that shows a variety of tests with the Hopper H100 GPU and compared its results with the GeForce RTX 4090, the brand’s most powerful gaming graphics card. The experiments included artificial intelligence processing and in-game benchmarks.
The first test used Stable Diffusion, a model of deep learning for text-to-image transformation similar to Open AI’s Dall-E. The Hopper H100 took 2.82 seconds to generate an image — fast, but slower than the GeForce RTX 4090. The performance difference is due to Xformers’ lack of support for the Hopper GPU.
With that, the enthusiast used the Donut, a model that extracts the maximum benefits from the Hopper H100 GPU acceleration engines. In this test, the video card for supercomputing proved to be 30% faster than the GeForce RTX 4090 and the RTX 6000 Ada.
For the tests, the Hopper H100 was tested with four different configurations to analyze the scalability of the GPU. Interestingly, a dual-GPU rig showed 43% more performance than NVIDIA’s gaming graphics cards, while a quad-GPU configuration performed comparable to a single-GPU performance. Check out:
Moving on to model training with VITS, which is very demanding on the memory available in the hardware, the H100 performed 23% more than the GeForce RTX 4090. The enthusiast noted that the massive 80GB capacity of the Hopper GPU contributed to a stress-free process. latency compared to the gaming model, with only 24 GB.
Finally, the Hopper H100 performed training on Meta’s LLaMA model, handling a total of 65 billion parameters — almost eleven times more than the GeForce RTX 4090.
On a curiosity level, geekerwan tested the Hopper H100’s capabilities in PC games. Using 3DMark Time Spy and Red Dead Redemption 2 to analyze its gaming performance, the enthusiast found that the supercomputing video card loses to the Radeon 680M, a GPU integrated into AMD’s new processors.
As mentioned, the Hopper H100 doesn’t have any optimizations for games. Its architecture and software are completely designed with a focus on specific calculations made by supercomputers in the field of medical research and artificial intelligence training.
The Radeon 680M, on the other hand, benefits from extensive support for the most popular games on the market, but could never match the performance and efficiency of the Hopper H100 in heavier workloads such as AI.
NVIDIA is one of the leading companies in the supercomputing space, but the competition is increasingly dedicated to threatening the rank of the “green team”. AMD, for example, recently introduced the Instinct MI300 accelerators with CDNA 3 architecture and options that include up to 192 GB of HBM3 memory.
