Configuring Kubernetes

Table of Contents

Intel iGPU Split Passthrough - This article is part of a series.

Part 3: Adding the GPU to the Kubernetes Nodes

Part 4: This Article

You have everything in place to make Kubernetes GPU-aware! For that you’ll be using the Intel Device Plugins for Kubernetes.

Configuration
#

Step 1: Label GPU-Enabled Nodes
#

First, label each GPU-equipped node so the Intel Device Plugin Operator knows where to work its magic (replace <node-name> with your own):

kubectl label nodes <node-name> intel.feature.node.kubernetes.io/gpu=true

Step 2: Install Intel Device Plugin Components
#

NOTE
Update (July 2025): When I first documented this process, I was living the manual life – helm install this, kubectl apply that. Since then, I’ve migrated to FluxCD for GitOps-based deployments. For those interested in the automated approach, my home-ops repository shows how to deploy Intel Device Plugins declaratively.

Grab the Intel Helm charts and update:

helm repo add intel https://intel.github.io/helm-charts/
helm repo update

Deploy the Device Plugin Operator:

helm install --namespace=kube-system intel-device-plugins-operator intel/intel-device-plugins-operator

Step 3: Configure the GPU Device Plugin
#

Create values.yaml to define your sharing strategy:

name: i915
sharedDevNum: 1         # Maximum pods per GPU
nodeFeatureRule: false  # Disable automatic node feature discovery

Deploy the GPU plugin with your created config:

helm install --namespace=kube-system intel-device-plugins-gpu intel/intel-device-plugins-gpu -f values.yaml

Using GPU Resources
#

Resource Requests
#

You can now configure pods to request GPU resources through Helm charts or Deployment manifests:

resources:
  limits:
    gpu.intel.com/i915: "1"
  requests:
    gpu.intel.com/i915: "1"

Node Selection
#

Got a mixed cluster? Force GPU workloads to GPU enabled nodes:

nodeSelector:
  intel.feature.node.kubernetes.io/gpu: "true"

Final Check
#

When you deployed your first application with a resource request for a GPU, verify within the container by executing:

ls /dev/dri

Sweet success looks (again) like this:

by-path  card0  renderD128

Congratulations! You’ve just pulled off the triple axel of virtualization: GPU passthrough from host to VM to container. Your media stack is now hardware-accelerated, your streams are smooth, and somewhere, a CPU is breathing a sigh of relief. Happy transcoding!