Oracle Adds AMD GPU Operator to OKE for Broader GPU Lifecycle Management

Oracle Kubernetes Engine now offers the AMD GPU Operator as an optional add-on for enhanced clusters. The change gives platform teams an OKE-managed route to coordinate the drivers, discovery, scheduling, monitoring, validation and configuration needed around AMD Instinct GPU workloads.

More than a device-plugin layer

OKE previously supported AMD Instinct hardware through an AMD GPU Plugin add-on that managed the AMD Device Plugin for Kubernetes. The operator expands that foundation: rather than leaving teams to separately install and operate the operator in each cluster, its lifecycle can be managed through OKE.

Preparing nodes before workloads land

The operator uses Node Feature Discovery to detect AMD Instinct hardware and publish each node’s capabilities as Kubernetes labels. Kernel Module Management and the Controller Manager support installing, upgrading and removing GPU driver kernel modules according to the chosen configuration, helping ready worker nodes before GPU workloads are scheduled.

How workloads find capacity

• The AMD GPU Device Plugin registers AMD Instinct GPUs as Kubernetes resources, including amd.com/gpu.
• Application teams request those resources in pod specifications, allowing Kubernetes to place work on nodes with available GPU capacity.
• Node selectors, taints, tolerations and GPU-specific labels can reserve or target nodes for particular hardware requirements.

Health checks and fleet-level configuration

The Device Metrics Exporter produces Prometheus-format data on GPU utilization, temperature and health. The operator can use health information with the device plugin so unhealthy GPUs are not offered as schedulable capacity.

For supported AMD Instinct environments, the Device Config Manager applies GPU-partitioning profiles stored in Kubernetes ConfigMaps to labeled nodes. The operator also includes a test runner for validation, diagnostics and benchmarking; depending on the selected test tooling, scenarios can include stress, PCIe bandwidth, memory and burn-in tests. It can run pre-start checks as init containers for GPU workload pods.

Managed does not mean automatic compatibility

The add-on is limited to enhanced clusters. Oracle directs teams to check supported combinations of OKE Kubernetes versions, AMD GPU worker-node images, hardware, drivers and add-on versions before enabling it. Administrators can select a supported add-on version, choose automatic updates or retain the deployed version, and apply supported configuration arguments.

Teams must also size operator-managed components for their environment. Oracle says requirements vary with GPU-node count, GPUs per node, monitoring frequency and workload intensity. The new add-on centralizes more of the software control plane, but it does not remove the need to validate the cluster stack and allocate operational resources for it.