Stability AI, the developer of the Stable Diffusion model, has conducted some important tests for its AI. It published a performance report testing Stable Diffusion 3 with different hardware accelerators. The benchmark pitted Intel Gaudi2 and NVIDIA H100 chips against each other.
Intel Gaudi2 vs NVIDIA H100 vs A100
Stability AI tested Stable Diffusion 3 on Intel Gaudi2 96GB, NVIDIA H100 80GB Hopper GPU and A100 80GB Ampere GPU. In a test of 16 images per device, Gaudi2 was able to produce images 56 percent faster than the H100 GPU.
In terms of numbers, Gaudi2 reportedly reached a speed of 927 images/second and H100 595 images/second. In addition, the A100 ranked third with 381 images/second.
As the cluster sizes increased, the performance gap widened. A cluster of 256 Gaudi2 chips achieved 12,654 images/second (49.4 images/second per device). In comparison, the A100 cluster achieved 3,992 images/second (15.6 images/second per device).
However, Stability AI said the A100 chips were better in performance thanks to TensorRT technology. There is a 40 percent performance difference compared to Gaudi2.
Stability AI plans to use the Intel Gaudi2 chip on the Stability Cloud platform in the coming period. Performance tests also show that NVIDIA, which leads the industry, is facing a formidable competitor.
So how did you find these tests? You can share your opinions with us in the Comments section below
{{user}} {{datetime}}
{{text}}