MIG have layerd structure GPU => GI => CI
p.s. GPU Instance (GI) and Compute Instance (CI)
Is there any document on multiple CI usage?
I read MIG user guide.
It writes how to use multi CI on Black Sholes.
But effective use case is not written.
Previously GPU can share by multiple processes (ex. Docker)
It is helpful any document on multiple CI usage compared to MPS and usual Docker (Streams).
MIG user guide
One place where MIG may have benefits over MPS is in a multi-client/tenant situation. For example with MIG you can guarantee that each of 7 clients gets exactly the same amount of GPU memory available to them. Can’t do that with MPS. Or Docker. Or Streams. This allows for more predictable behavior if you want to share a GPU among multiple users.
Another benefit of MIG over MPS is in the area of error containment. With MIG, an error in a particular instance does not get reflected to other instances.
It will be harder to find obvious benefits for MIG if you are comparing it only with the case where the same workload could be efficiently performed on a non-fractionalized GPU. I don’t think you’ll find claims that MIG mode is faster than non-MIG mode in this case, and that seems to be what you are looking for.
Thank you, I know GI use case.
I am wondering CI use case.
Since If GI to CI relation ship is 1 to 1, then it should remove ci definition command, it is helpful for user.