Hi all
I’m thinking about upgrading my server to A100 GPU cards. They’re pretty expensive, but I was under the impression that multiple cryosparc jobs could be run on a single A100 card? Is that true? Anyone have A100s that they run cryosparc on that can comment? Is it worth the $$$$?
Thanks
Yes, I have tested cryoSPARC on A100 GPUs. Everything works. not sure what you meant by running multiple jobs on a single card though…
Yes, they work great, but may not be the most economical investment when compared to other available GPUs. I have also used RTX2080Tis, RTX3090s, A40s and A6000s and queued multiple jobs to all without issue (as long as you don’t surpass the RAM available). Note that you end up trading bandwidth for performance (speed), but that can be advantageous for CryoSparc workflows.
Submitting multiple jobs to the same GPU is tedious as you need to manually assign the GPU you want in the job submission prompt. It can start to become difficult to keep track of how many jobs are running on what GPU if you have many. Multi-instance GPU (MIG) has not worked as far as I know, but would be a more stable way to run multiple jobs on a single GPU, since it prevents crashing all running jobs when you exceed the capacity of the GPU.
TL;DR - It works, but I think you get more bang for your buck elsewhere. A100s are expensive for non-enterprise CryoSparc-specific applications.
I was told by my IT department that A100 can be split into multiple partitions so that they are like multiple virtual gpus. I was also told that there’s no way to buy server modes with 3080 Ti cards anymore, that nvidia has ‘disallowed’ it somehow.
The plan was to buy fewer a100 cards in order to slot more cryosparc jobs on each card…that’s not going to work?
It hasn’t worked for us. You should look into further posts on the forum like this one:
You can manually queue multiple jobs to one GPU but it will be poorly controlled.