Intel Gaudi2 vs NVIDIA H100: Unveiling the Fastest AI Chip

MetaversePlanet March 12, 2024

0 1 minute read

Stability AI, the developer behind the Stable Diffusion model, has carried out significant tests for its artificial intelligence.

Stable Diffusion published a performance report comparing its model on various hardware accelerators. In this comparative analysis, the Intel Gaudi2 and NVIDIA H100 artificial intelligence chips went head-to-head.

Intel Gaudi2 vs NVIDIA H100 vs A100

Stability AI conducted tests on Stable Diffusion 3 using Intel Gaudi2 96GB, NVIDIA H100 80GB Hopper GPU, and A100 80GB Ampere GPU. In a benchmark of producing 16 images per device, Gaudi2 outperformed the H100 GPU by 56 percent in speed.

Specifically, the Gaudi2 achieved a remarkable speed of 927 images per second, while the H100 recorded 595 images per second, with the A100 following at 381 images per second.

As the cluster sizes increased, the disparity in performance became more pronounced. A cluster of 256 Gaudi2 chips managed 12,654 images per second, equating to 49.4 images per device. On the other hand, an A100 cluster reached 3,992 images per second, or 15.6 images per second per device.

However, Stability AI noted that the A100 chips outperformed in efficiency thanks to TensorRT technology, with a 40 percent performance edge over Gaudi2.

Stability AI intends to incorporate the Intel Gaudi2 chip into its Stability Cloud platform shortly. These performance tests highlight that NVIDIA, a leader in the industry, now has a serious contender in Intel.

What are your thoughts on these tests? Feel free to share your opinions in the comments section below.