(GP)GPU batched jobs

Dear All

Suppose I have several users using a (GP)GPU. Can I configure (the API driver) in which each user run exclusively in the (GP)GPU?

So when all the users submit a process (with CUDA compiled) at same time the processes are batched and ran exclusively each at a time.


Luis Gonçalves

The GPU offers a default compute mode which generally allows this with the caveat that all users are actually sharing the GPU, and if each user attempts to allocate most of the memory, it won’t work. To have non-shared batched exclusive usage, you should use a job scheduler or similar tool.