Sometimes the difference is three times
Nvidia dominates the AI accelerator market, but this does not mean that its solutions are the most productive in all modes. For example, tests show that Intel Gaudi 2 can be many times faster than Nvidia H100.
The new tests were published not by Intel at all, but by Stability AI, which makes them more interesting. The company decided to compare Intel's new product with Nvidia H100 and A100 in terms of AI training speed in the Stable Diffusion 3 model.
As you can see, Gaudi 2 can be twice as fast as the H100 and three times as fast as the A100.
At the same time, in output tests with the Stable Diffusion 3 8B parameter model, the Gaudi 2 accelerator provides output speeds similar to Nvidia A100 with basic PyTorch. At the same time, thanks to TensorRT optimization, A100 accelerators create images 40% faster than Gaudi 2.
In tests of the Stable Beluga 2.5 70B model without any optimizations, the Intel solution showed 28% higher performance than the A100.
Thus, the new Intel accelerator has considerable potential. Of course, for many clients, versatility, ecosystem and much more are important, but still Intel Gaudi 2 can definitely be an alternative to Nvidia accelerators. Let us remember that yesterday the head of Nvidia said that even if competitors give away their accelerators for free, it will not be cheap enough compared to the solutions of Nvidia itself.