Question about inference speed on H100 GPU

[yunhzou]

Hi,
I am curious to learn the inference speed of AF3 when deployed on H100 GPUs. Could this be released? Thank you!

[Augustin-Zidek]

Hi,

the inference speed on H100 should be at least as fast as on A100 (see https://github.com/google-deepmind/alphafold3/blob/main/docs/performance.md).

I recommend running a benchmark on your machine with the following number of tokens: 1024, 2048, 3072, 4096, 5120 and comparing these to the A100 numbers we report.