H100 for final training run

[ghost]

Are we allowed to use the Nvidia H100 for our final training run? I am planning to do +-160k weight updates and my model is quite heavy, so being able to use the H100 will probably significantly speed up the process. Or will this burn through the compute credits?

[TimJaspers0801]

Yes, if you believe it's necessary, you're allowed to proceed. However, please make sure to use the H100s efficiently. Remember that each H100 node has fewer CPU cores, so if you don't adjust accordingly, you might end up being charged for two H100s. Please review your setup carefully. While H100s are generally more cost-effective than A100 GPUs, this is only true when they are used optimally.

[ghost]

Thanks, I will keep that in mind!

[ghost]

Apparently training on the h100 requires another version of CUDA. So i'll just stick to A100.