Intel GPU (Operator + Plugin)¶

The Intel GPU stack enables iGPU passthrough for hardware-accelerated video transcoding in workloads such as Jellyfin and Tdarr.

Details¶

Field	Value
Charts	`intel-device-plugins-operator`, `intel-device-plugins-gpu`
Repository	https://intel.github.io/helm-charts
Version	0.35.0 (both)
Namespace	`intel-gpu-operator` (CreateNamespace=true)
Sync Waves	Operator: -2, Plugin: -1

Key Configuration¶

The stack is split into two ArgoCD Applications:

Intel Device Plugins Operator (sync wave -2)¶

Installs the operator that manages GPU device plugin lifecycle. Uses a retry policy of limit 30 with exponential backoff (10s to 5m) and SkipDryRunOnMissingResource=true to handle CRD creation timing.

Intel Device Plugins GPU (sync wave -1)¶

Deploys the GpuDevicePlugin custom resource that the operator reconciles into a DaemonSet.

sharedDevNum: 5 -- up to 5 pods can share the same GPU simultaneously.
nodeFeatureRule: false
ignoreDifferences: The .spec field of GpuDevicePlugin is excluded from ArgoCD diff detection because the operator actively manages the spec after initial creation.

Both applications share the same retry policy and SkipDryRunOnMissingResource setting.

Cluster Integration¶

Applications that require GPU access are configured with:

nodeSelector:
  gpu: intel
resources:
  limits:
    gpu.intel.com/i915: "1"
volumeMounts:
  - name: dev-dri
    mountPath: /dev/dri
volumes:
  - name: dev-dri
    hostPath:
      path: /dev/dri

PCI Passthrough

The physical iGPU is passed through to the Kubernetes node VMs at the Terraform/Proxmox layer via the pci_mappings field in terraform.tfvars, which references a Proxmox PCI device mapping. The Intel GPU plugin only handles the in-cluster device advertisement.

Upstream Documentation¶

https://github.com/intel/intel-device-plugins-for-kubernetes