Use Multi-Instance GPU (MIG)
This guide walks you through configuring Multi-Instance GPU (MIG) on VKS to partition an NVIDIA H100 GPU into multiple isolated MIG instances — each with dedicated VRAM and Streaming Multiprocessors, ensuring complete workload isolation.

Prerequisites
A running VKS Cluster with a node group using an NVIDIA H100 GPU (or Ampere/Hopper or newer — MIG requires Ampere architecture or later).
kubectlandhelminstalled on your local machine.yqversion ≥ 4 installed for safe YAML merging (see Step 2).kubectlconnected to the cluster. Verify withkubectl get nodes.
Step 1: Install GPU Operator with mixed MIG Strategy
MIG on VKS requires the NVIDIA GPU Operator. VKS nodes do not include a pre-installed NVIDIA driver, so driver.enabled=true is mandatory.
The mixed strategy allows a node to run some GPUs in MIG mode and others as full (non-MIG) GPUs simultaneously — for example, GPU 0 partitioned into MIG instances while GPU 1 remains a full GPU.
Step 1.1: Add the NVIDIA Helm repo and install GPU Operator
Step 1.2: Verify pod status
Wait 5–10 minutes for the GPU Operator to deploy the driver container to the node, then run:
All pods must be Running or Completed:
gpu-feature-discovery
Running
nvidia-container-toolkit
Running
nvidia-cuda-validator
Completed
nvidia-dcgm-exporter
Running
nvidia-device-plugin-daemonset
Running
nvidia-driver-daemonset
Running
nvidia-mig-manager
Running
nvidia-operator-validator
Running

Step 2: Create a Custom MIG ConfigMap
GPU Operator overwrites default-mig-parted-config on every reconcile. You must create a new ConfigMap with a different name and point migManager.config.name to it.
Step 2.1: Install yq if not already available
Step 2.2: Export the default ConfigMap
Step 2.3: Merge the new profile using yq
In this example:
GPU 0 is enabled for MIG with profile
3g.40gb— split into 2 instances, each with 40 GB VRAM and 60 Streaming Multiprocessors.GPU 1 remains non-MIG (full 80 GB).
Step 2.4: Verify the YAML structure before applying
Step 2.5: Create the ConfigMap and upgrade GPU Operator
Step 3: Apply MIG Config to the Node
Label the node with nvidia.com/mig.config to activate the MIG profile:
Follow MIG Manager logs to confirm the config is applied:
Expected output:

Step 4: Verify Node Resources
Check that the node exposes the correct MIG resources:
Expected result (2x MIG 3g.40gb + 1 full GPU):
Verify node labels set by MIG Manager:
nvidia.com/mig-3g.40gb.count
2
Number of MIG instances
nvidia.com/mig-3g.40gb.memory
40448 MiB
VRAM per instance
nvidia.com/mig-3g.40gb.multiprocessors
60
Streaming Multiprocessors per instance
nvidia.com/mig.config.state
success
Config apply status
nvidia.com/mig.strategy
mixed
Active MIG strategy

Step 5: Deploy a Workload Using MIG Instances
Declare nvidia.com/mig-3g.40gb in resources.limits — the Kubernetes scheduler automatically assigns an available MIG instance to the Pod with no additional configuration.
Deploy 2 pods simultaneously to verify isolation:
Each Pod receives a different MIG UUID — confirming complete isolation:

(Optional) Step 6: Clean Up
Result
After completing these steps, your VKS node exposes both MIG instances and a full GPU simultaneously:
GPU 0 — MIG (2x 3g.40gb)
nvidia.com/mig-3g.40gb
2
GPU 1 — non-MIG (full 80 GB)
nvidia.com/gpu
1
Each 3g.40gb MIG instance provides: 40 GB VRAM · 60 SM · full isolation.
Monitor GPU resources
Autoscale GPU Nodegroup
Compare GPU sharing modes
Last updated

