The team at Stability.AI has reported on the performance testing results of artificial intelligence model training chips, comparing the Intel Gaudi 2, NVIDIA H100, and NVIDIA A100 chips. It was discovered that the Intel chip outperformed the others significantly when running with similar training parameters. Gaudi 2 was able to train up to 927 images per second compared to H100 at 595 images per second, a 1.5x increase. Additionally, Gaudi 2 boasts a higher RAM capacity of 96GB, allowing for further parameter adjustments to speed up the process.
However, when testing the Stable Diffusion model with TensorRT, performance on the A100 chip proved to be 40% better than the Gaudi 2. Yet, running the Stable Beluga 70B LLM model still showed a 28% performance advantage on the A100 over the Gaudi 2.
Stability.AI concluded that the Gaudi chip is a pivotal choice for the market due to its cost-effectiveness, faster delivery, and superior performance-to-cost ratio.
TLDR: Stability.AI’s testing revealed that the Intel Gaudi 2 chip outperforms NVIDIA chips in image training speed but falls behind in other model testing scenarios. They recommend the Gaudi chip for its affordability, fast delivery, and overall performance.
Leave a Comment