AMD has released GPU Operator v1.5.0, a tool designed to improve Kubernetes-based GPU deployments by adding support for Dynamic Resource Allocation (DRA) and automated GPU node recovery. The update enables more flexible resource management and improves infrastructure control for teams using AMD GPUs. These features are part of broader efforts to align AMD's GPU infrastructure with Kubernetes' newer resource allocation models. According to AMD, the release expands deployment flexibility through independent Kernel Module Management (KMM) installation and configurable kubelet socket paths. It also includes enhancements to the Device Metrics Exporter, such as new ECC metrics and per-process Compute Unit occupancy telemetry. Source: amd

The new DRA support changes how GPUs are represented and allocated within Kubernetes. Instead of statically advertising devices through kubelet, DRA introduces Kubernetes-native resource allocation objects like DeviceClass and ResourceSlice, enabling scheduler-driven allocation workflows. This allows multiple containers within a pod to coordinate access to the same GPU resources, which is particularly useful for AI workflows requiring shared accelerator access. DRA is configured through the DeviceConfig CR, which allows users to enable or disable the feature and specify the driver image. The operator automatically deploys the DRA driver, creates required DeviceClass objects, and detects the highest supported DRA API version available on the cluster. Source: amd

The release also introduces Auto Node Remediation (ANR), which automates recovery workflows for unhealthy GPU worker nodes. This feature integrates Device Metrics Exporter, Node Problem Detector, and Argo Workflows to create a real recovery loop. ANR behavior is configured through the DeviceConfig CR, allowing users to set limits on simultaneous remediation events and control recovery policies. Remediation workflows can trigger auto-recovery, scheduler avoidance, or cluster-wide alerts based on GPU-related conditions such as inband RAS errors or repeated ECC error accumulation. Source: amd