Amazon EKS Auto Mode and Karpenter now support Dynamic Resource Allocation

Amazon EKS Dynamic Resource Allocation (DRA) for GPUs and EFA

Amazon EKS now supports Dynamic Resource Allocation (DRA) for NVIDIA GPUs and Elastic Fabric Adapter (EFA) network devices when using Amazon EKS Auto Mode and Karpenter. This allows you to define device attributes for your Kubernetes workloads, share GPU devices across Pods, and allocate topology-aligned GPU and EFA devices while dynamically scaling Amazon EC2 instances based on workload demand.

With DRA, you define ResourceClaim or ResourceClaimTemplate objects that describe what GPU or EFA devices can be allocated to workloads, by model, architecture, or other attributes. When running distributed training or inference, you can define PCIe-aligned GPU and EFA devices for optimal performance. Workloads then request those DRA objects in their configuration, and Karpenter automatically launches the lowest-cost instances that match.

Amazon EKS Auto Mode automatically prevents workloads using DRA and workloads using device plugins for the same device type from running on the same nodes to mitigate the risk of silent device oversubscription during migrations.

What to do

• Define ResourceClaim or ResourceClaimTemplate objects for your workloads.
• Request those DRA objects in your workload configuration.
• Ensure your Kubernetes version is 1.36 or newer for Amazon EKS Auto Mode.

DRA support for Amazon EKS Auto Mode and Karpenter is available in all AWS Regions, except the AWS GovCloud (US) Regions and the China Regions. To learn more, see the DRA documentation in the Amazon EKS User Guide and Karpenter documentation.

If you need further guidance on AWS, our experts are available at AWS@westloop.io. You may also reach us by submitting the Contact Us form.