Working on VGPUs in EKS, but facing some issues on my head.

[josephlee518]

Hello.

I've looking some 'virtualized gpu' framework things. but I cannot solve some questions.

1. If I've allocated some "virtualized gpus" in the instances. How can I solve CUDA Core Fraction?
   • I've heard that Ampere's MIG Support Slicing NVIDIA CUDA Cores. but I did not heard that feature in Volta Instances.
   • So, Someone is Running that requires GPU resource. and other need to Wait for pre task is finished?
2. If that task is not for the inference things (kinda like train some lightweight model) how can I correctly tells them one of the container can use n amount of cuda cores and memory?

Thanks.