🧀 BigCheese.ai


Intel Gaudi2 chips outperform Nvidia H100 on diffusion transformers


In this latest blog 'Behind the Compute', Stability AI benchmarks compute solutions comparing Intel Gaudi 2 accelerators with Nvidia's A100 and H100. The tests involve training Stable Diffusion 3 and Stable Beluga 2.5 70B models. Gaudi 2 shows 1.5 times training speed of H100 and 3x more images processed per second compared to A100. While A100 excels in optimized inferences, expectations are high for Gaudi 2's ongoing optimization to meet or surpass A100 performance. The findings highlight the need for powerful, efficient, and cost-effective computing alternatives like Gaudi 2 in the AI industry.

  • Gaudi 2 is 1.5x faster than H100 in training speed
  • Gaudi 2 processes 3x more images/sec than A100
  • Stable Diffusion 3 and Stable Beluga 2.5 70B were benchmarked
  • Anticipation of Gaudi 2's optimization to improve performance
  • Stability AI highlights the importance of computing alternatives