Everyone wants AI… Well, in industry… And to run AI models, you need high-performance, high-performance hardware. For this, we have Intel’s H100 and Intel’s Gaudi2. To find out which solution is faster, Stability AI, the guys behind Stable Diffusion 3, an image-generating AI, are carrying out comparative tests.
Stable Diffusion 3: who’s got the fastest hardware, Intel or NVIDIA?
Of course, these results are based on the performance of entire servers. For example, two 16-accelerator nodes with 16 GPUs each are capable of generating 927 fps when equipped with Gaudi2 products. On the NVIDIA side, scores are not as good: with the same number of chips, H100s generate 595 images per second. With A100s, the result falls to 381.
Moving on to machines with 4096 gas pedals, the number of images generated per second rises to 12,654 on Gaudi2. With H100s, this number is still lower: 3,992, quite a difference. On a per-card basis, this represents a difference of 49.4 fps with Intel hardware versus 15.6 fps with H100.
Now, the software environment must also be taken into account. Until NVIDIA cards take advantage of TensorRT, Intel products perform better. However, with TensorRT, we learn that an A100 is 40% faster than a Gaudi2.
So, if the hardware is important, so is the API or Framework used.
