Intel GPU Base Operator v0.4.0

Release 0.4.0

Minor release for Intel GPU Base Operator.

Please note that while the operator continues to develop, the ClusterPolicy CRD will receive changes, sometimes breaking ones. Helm does not update the CRDs automatically, thus it's left for the cluster admin to update them. Easiest way to achieve this is to remove the previous chart, remove the CRDs and then install the new version. Eventually the CRD will stabilize and CRD update needs become scarce.

To delete CRDs:
kubectl delete crds clusterpolicies.intel.com gpufirmwareupdates.intel.com

Breaking CRD changes:

• GPUFirmwareUpdate
  • AMC's credentials removed
  • FW file regex tightened

Changes from 0.3.0 release:

• Support for OoT X KMD (#79) by @abyrne55
  • ClusterPolicy's kernelModule object
  • Documentation in OOTDRIVER.md
• ClusterPolicy components support affinity.
• New B-series labels in NFD
• Support for GPU recovery via a new GPURecoveryPlan
  • Documentation in RECOVERY.md
• Operator's webhook disabling is no longer possible
  • Webhooks are critical for operator's functionality
• Updated Intel dependencies
  • Intel GPU DRA driver 0.12.0
  • Intel GPU Plugin 0.37.0
  • Intel XPUMD 2.2.0
• Various Golang dependency updates

Fixes:

• Kueue queue name changes are handled better (#63)
• DRA: DeviceClass: filter out driverless devices without binding (#123)

Known issues/limitations:

• KMM support supports only in-cluster build in Ubuntu26.04.
  • In-cluster build depends on a registry with write access
  • An example sample is included in config/samples directory.
• GPURecoveryPlan functionality depends on specific functionality in DRA, xpumd and xpu-smi containers. The current versions lack the support thus GPU Recovery is not operational on 0.4.0
• DRA 0.12.0 repots itself as 0.11.0 in the container logs even though the container is correct.

Basic install with Intel GPU DRA Driver:

helm install --create-namespace --namespace "intel-gpu-base-operator" --version 0.4.0 gpu-operator oci://ghcr.io/intel/intel-gpu-base-operator-chart --wait
helm install --namespace "intel-gpu-base-operator" --version 0.4.0 gpu-policy oci://ghcr.io/intel/intel-gpu-base-operator-policy-chart --set resourceRegistration=dra

Basic install with Intel GPU Device Plugin:

helm install --create-namespace --namespace "intel-gpu-base-operator" --version 0.4.0 gpu-operator oci://ghcr.io/intel/intel-gpu-base-operator-chart --wait
helm install --namespace "intel-gpu-base-operator" --version 0.4.0 gpu-policy oci://ghcr.io/intel/intel-gpu-base-operator-policy-chart --set resourceRegistration=dp

Collateral images:

docker.io/intel/intel-gpu-plugin:0.37.0@sha256:7001200071db10587fc31184eedda4f5cf291efba737adf85a8c80539f7fa5e1
ghcr.io/intel/intel-gpu-resource-driver:v0.12.0@sha256:5be06e1cf810c6091cec17ed78a775be64323f2084038c8840f283f643ce4207
ghcr.io/intel/xpumanager/xpumd:v2.2.0@sha256:fdefec8e5d93f76a1c7b4dda56001b7f3c2c15eae29bfbbc8971b900e6e74ae8

What's Changed

• Add Openshift SCC handling by @tkatila in #36
• Dependabot updates by @pfl in #42
• helm: scope metrics-reader ClusterRole name to release by @abyrne55 in #47
• Fixes from 0.3.0 rc0 by @tkatila in #46
• Helm fix name collisions by @tkatila in #49
• readme: update install instructions for 0.3.0 by @tkatila in #56
• Controller refactoring by @tkatila in #67
• Fix kueue updates and improve reconciler by @tkatila in #63
• Kueue combine code and add test by @tkatila in #76
• Update kueue to v0.19 by @pfl in #78
• Add optional KMM integration for OoT kernel module loading by @abyrne55 in #79
• controller: set n/a status if no DSs are created by @tkatila in #87
• Node affinity by @pfl in #89
• xpumd: update to v2.1.0 by @tkatila in #90
• KMM: nfd support and docs by @tkatila in #96
• FW update: drop credentials from AMC update by @tkatila in #98
• controller: drain: extract node taint and pod drain primitives by @tkatila in #99
• controller: fix requeueReconcileErr handling if it ever contains a reason by @tkatila in #100
• fwupdate: make content image verification generic by @tkatila in #101
• Recovery: API and docs by @tkatila in #102
• Misc fixes vol9 by @tkatila in #108
• recovery: add the GPURecoveryPlan webhook by @tkatila in #107
• recovery: detect GPUs needing recovery from ResourceSlices by @tkatila in #109
• Recovery: approvals and reset job creation by @tkatila in #110
• Recovery: drain handling for GPU reset by @tkatila in #111
• recovery: allow only one recovery per node at a time by @tkatila in #115
• Recovery: reflash job and timeouts by @tkatila in #116
• recovery: add the OpenShift SCC for recovery Jobs by @tkatila in #117
• Xpumd restart functionality by @tkatila in #119
• recovery: align xpu-smi use in reset and reflash Jobs by @tkatila in #120
• recovery: remove automatic retry and improve Job failure detection by @tkatila in #122
• recovery: add the kubectl-gpurecovery plugin by @tkatila in #121
• DRA: add cel expression for driver by @tkatila in #123
• recovery: split tests into small files by @tkatila in #124
• Recovery: allow sbr as default, update docs and add missing flag to slot reset by @tkatila in #125
• Doc updates vol2 by @tkatila in #130
• misc: nfr: add detection for class 1200 devices by @tkatila in #129
• Fixes for 0.4.0 release by @tkatila in #131
• Deps update q3 2026 by @tkatila in #135

New Contributors

• @pfl made their first contribution in #42
• @abyrne55 made their first contribution in #47

Full Changelog: v0.3.0...v0.4.0