How does MIG work in the context of Kubernetes and which GPUs support it?

MIG (Multi-Instance GPU) is a technology by NVIDIA that allows a single physical GPU to be partitioned into multiple isolated logical GPUs (instances). In the context of Kubernetes, this enables running multiple containers with dedicated parts of a GPU, providing resource isolation and increasing GPU utilization efficiency.

In Kubernetes, the NVIDIA Device Plugin is typically used to work with MIG, which detects and exports each MIG instance as a separate device. This allows pods to request specific MIG instances via resources in their manifests.

Supported GPU architectures include NVIDIA Ampere and newer, such as A100, A30, A40, and others that have hardware support for MIG. Older GPU generations like Turing or Volta do not support MIG.

Example usage in a Pod spec:

resources:
  limits:
    nvidia.com/mig-1g.5gb: 1

Here, mig-1g.5gb is a type of MIG instance indicating the allocation of a part of the GPU with specific resources.