A100 VS A40 performance issue

truatpasteurdotfr commented 1 year ago

Hi

I am wondering if you can give some ideas about the performance difference between A100:40G and A40:48GB ? "toy system" used as benchmark: https://rest.uniprot.org/uniprotkb/F0NHT7.fasta

runnning on the A100 (40GB), and A40 (48GB): omegafold --subbatch_size 768 --weights_file ~/APPASCRATCH/OmegaFold/model.pt ~/F0NHT7.fasta ~/APPASCRATCH/OmegaFold-48G omegafold --subbatch_size 768 --weights_file ~/APPASCRATCH/OmegaFold/model.pt ~/F0NHT7.fasta ~/APPASCRATCH/OmegaFold-40G on the A100(80GB): omegafold --weights_file ~/APPASCRATCH/OmegaFold/model.pt ~/F0NHT7.fasta ~/APPASCRATCH/OmegaFold-80G

the A100 are A100-SXM4-40GB or A100-SXM4-80GB while the A40 is PCI-e based.

INFO:root:Loading weights from /pool/omegafold/weights_files/20220921/model.pt
INFO:root:Constructing OmegaFold
INFO:root:Reading /pool/omegafold/F0NHT7.fasta
INFO:root:Predicting 1th chain in /pool/omegafold/F0NHT7.fasta
INFO:root:1229 residues in this chain.
INFO:root:Finished prediction in 2200.10 seconds.

while the A100-40G is completing the task in ~990 seconds and the A100-80G finishes in ~840 seconds.

RuiWang1998 commented 1 year ago

Hi,

Sorry we do not have access to A40 GPUs so we do not think we could answer this.

truatpasteurdotfr commented 1 year ago

Is omegafold using double precision or just single precision on the gpu? that might explain the huge penalty on A40 VS A100.

HeliXonProtein / OmegaFold

A100 VS A40 performance issue #40