intermediateEnablement ~30 min

Configure MIG on an A100/H100

Partition a GPU into MIG instances and schedule onto them.

Scenario

You want to share one GPU across several small workloads with isolation.

Prerequisites
  • ·A MIG-capable GPU (A100/H100)
  • ·GPU Operator installed
Tasks
0 / 3 tasks
  1. bash
    kubectl label nodes $NODE nvidia.com/mig.config=all-1g.10gb --overwrite

    Expected · The operator reconfigures the GPU (the node briefly shows no GPUs).

  2. bash
    kubectl get node $NODE -o json | jq '.status.allocatable' | grep mig

    Expected · Entries like nvidia.com/mig-1g.10gb.

  3. bash
    # resources.limits:
#   nvidia.com/mig-1g.10gb: 1
Success criteria
  • ✓MIG instances appear as allocatable
  • ✓A pod schedules onto a MIG slice
Hints
  • ·Reconfiguring MIG drains the GPU — do it in a window.
  • ·Match the operator's MIG strategy (single vs mixed) to your profiles.