AI Infrastructureeks

DRA on EKS Auto Mode and Karpenter: Choosing Between Device Plugins and Dynamic Resource Allocation for GPUs

P&C

Prompt & Circuit Editorial

November 16, 2026 · 12 Min Read

DRA on EKS Auto Mode and Karpenter: Choosing Between Device Plugins and Dynamic Resource Allocation for GPUs

Why this matters

On November 9, 2026, AWS announced that Amazon EKS supports Dynamic Resource Allocation (DRA) for NVIDIA GPUs and Elastic Fabric Adapter (EFA) network devices when you use EKS Auto Mode and Karpenter [1]. For architects running distributed training or inference on Kubernetes, this moves GPU requests from "give me N GPUs" to "give me devices with these attributes, aligned like this", while node scaling stays automatic [1]. The catch is that the two compute options do not support the same things, and several limits are documented. This guide walks through them.

Device plugins versus DRA

Device plugins advertise hardware as extended resources such as nvidia.com/gpu, which you request in container resource requests and limits [2]. EKS documentation lists their limitations: devices are requested as opaque integer counts with no attribute-based filtering, there is no device sharing between containers or Pods, and there is no expressive topology-aware allocation across device types [2].

DRA, generally available in Kubernetes 1.34, addresses these. Drivers publish device attributes through ResourceSlice objects, and you request devices with ResourceClaim and ResourceClaimTemplate objects that reference DeviceClass categories [2]. According to the EKS guide, this enables attribute-based selection with Common Expression Language (CEL) expressions, and topology-aware allocation that co-locates devices on the same PCIe switch or NUMA domain [2]. The announcement describes the practical result: you can define PCIe-aligned GPU and EFA devices for distributed training or inference, and Karpenter launches the lowest-cost instances that match [1].

EKS recommends DRA drivers for new deployments on Kubernetes 1.34 and later, and states that device plugins remain fully supported for all Kubernetes versions [2][4].

EKS Auto Mode and Karpenter compared

The two options differ in who owns the drivers, which Kubernetes versions qualify, and what sharing is available.

AspectEKS Auto ModeKarpenter
DRA driversNVIDIA DRA driver and EFA DRA driver (DRANET) are built in [1]You install and manage the drivers [2]
Kubernetes version1.36 or later [2]Announcement says all Kubernetes versions [1]; the DRA driver itself is recommended from 1.34 [2]
Dynamic provisioning statusDocumented as supported for NVIDIA GPUs and EFA [2]Supported as of Karpenter 1.15, documented as alpha and off by default [2]
GPU time-slicing and MIG through DRANot available [2]Time-slicing available; MIG needs static capacity, managed node groups or self-managed nodes [2][3]
DeviceClass objectsYou create them yourself [2]Created by the driver chart when you install it with Helm [2]

Note the difference in maturity. The announcement presents both as supported [1], while the EKS User Guide labels Karpenter's dynamic provisioning for DRA as alpha and off by default [2]. Treat that as a signal to test before committing production capacity to it.

How Auto Mode keeps DRA and device plugins apart

Running a DRA driver and a device plugin for the same device on one node can cause silent oversubscription, because each mechanism allocates the device independently [2]. EKS Auto Mode prevents this: it chooses the device manager per device type when it launches a node, based on the workloads it launches the node for [1][2].

Two consequences matter for migration:

  • A node keeps its role for its whole lifetime, even when no Pods run on it, and you cannot change the role in place [2].
  • To move a workload from device plugins to DRA you must replace the nodes, and with a Capacity Reservation or Capacity Block this means scaling down before scaling up [2].

Auto Mode applies the labels eks.amazonaws.com/device-manager-nvidia and eks.amazonaws.com/device-manager-efa, set to dra or device-plugin, and you can set them on a NodePool to choose the device manager explicitly [2]. For static capacity provisioning, setting the label is required for DRA, and a static NodePool without it uses the device plugin [2].

On Karpenter, the separation is your job: you are responsible for keeping DRA drivers and device plugins off the same node, and the documentation suggests labels and node selectors to keep DRA Pods on DRA nodes [2].

Documented limits

The EKS guide lists these constraints for DRA with EKS Auto Mode [2]:

  • Only NVIDIA GPUs and EFA devices are supported. Use the Neuron device plugin for AWS Trainium and Inferentia [2].
  • The g7e instance family is not supported for dynamic provisioning, because there is no DRA device metadata for it. Use the NVIDIA device plugin for g7e GPUs [2][4].
  • ComputeDomains for Multi-Node NVLink require Karpenter static capacity provisioning, managed node groups or self-managed nodes [2].
  • EFA nodes need a routable subnet. Auto Mode attaches an EFA interface on every EFA-capable network card by default, so Amazon EC2 does not assign a public IP address, and in a public subnet the node joins the cluster but cannot pull images [2].
  • Do not combine dynamic capacity provisioning, static network interface configuration and topology-aligned GPU and EFA allocation. The behavior is undefined, and the Pod can remain Pending with no error at provisioning time [2].

The Karpenter section adds more [2]:

  • Karpenter evaluates claims only against instance types it ships metadata for, so support for a new instance type requires a new Karpenter release [2].
  • settings.enableDRA assumes both the NVIDIA and EFA DRA drivers run on your nodes. If one does not, Karpenter launches nodes for claims that can never be satisfied, so enable only the gate for the driver you run [2].
  • A Pod that requests nvidia.com/gpu as an extended resource is not treated as a DRA Pod, even when a DRA driver serves that resource [2].
  • Karpenter's device metadata describes whole GPUs only, so it does not launch nodes for claims against the MIG DeviceClass [2].
  • Upgrading an existing release with helm upgrade --reuse-values fails when you set settings.enableDRA, because earlier chart versions default settings.ignoreDRARequests to true. The guide says to use --reset-then-reuse-values, which requires Helm 3.14 or later [2].

Regional availability is also bounded: the announcement says DRA support is available in all AWS Regions except the AWS GovCloud (US) Regions and the China Regions [1].

Sharing GPUs

Sharing is where the choice of compute option bites. The EKS guide describes MIG as a hardware feature that partitions one physical GPU into isolated instances, and time-slicing as software time-multiplexing without isolation [3]. Time-slicing is available with Karpenter, managed node groups and self-managed nodes, but not in EKS Auto Mode [3].

The guide explains that time-slicing and MIG through the NVIDIA DRA driver are not available in Auto Mode because these capabilities are alpha in the upstream driver [2]. It also recommends static configuration with the DRA driver or the device plugin if you use MIG or time-slicing [4]. In Auto Mode, one supported sharing route is a ResourceClaim that several Pods reference, which allocates the same device to each of them [2].

Decision guide

SituationDocumented direction
New cluster on Kubernetes 1.36 or later, whole-GPU or GPU plus EFA workloadsEKS Auto Mode with DRA; drivers are built in [1][2]
Need MIG or time-slicingKarpenter (time-slicing), or static configuration; MIG needs static capacity or node groups [2][3][4]
g7e instancesNVIDIA device plugin [2][4]
Trainium or InferentiaNeuron device plugin on Auto Mode [2]
Kubernetes older than 1.34Device plugins [2]

Architect's checklist

  • Confirm your cluster's Kubernetes version: 1.36 or later for DRA on Auto Mode, 1.34 or later for the recommended DRA path elsewhere [2].
  • Check that your Region is not in GovCloud (US) or China [1].
  • List the GPU instance families you plan to use and verify none is g7e [2].
  • Decide whether you need MIG, time-slicing or ComputeDomains, and pick the compute option accordingly [2][3].
  • On Auto Mode, create the DeviceClass objects yourself before workloads claim devices, since a claim that references a missing class is never allocated [2].
  • Plan node replacement, and capacity scale-down before scale-up for reservations, if you migrate from device plugins [2].
  • Place EFA workloads on private subnets with a NAT gateway or the PrivateLink endpoints they need [2].
  • On Karpenter, treat dynamic DRA provisioning as alpha, enable only the gates for the drivers you run, and keep DRA and device plugin nodes separate [2].

Sources

Numbers in square brackets refer to these sources, checked on 2026-11-16.

  1. Amazon EKS Auto Mode and Karpenter now support Dynamic Resource Allocation - AWS
  2. Manage hardware devices on Amazon EKS - Amazon EKS
  3. Manage NVIDIA GPUs on Amazon EKS - Amazon EKS
  4. Use the NVIDIA DRA driver or device plugin on Amazon EKS - Amazon EKS
Tags:ekskarpentergpukubernetesdra
Share Article

Responses (0)

Sign in to join the conversation and share your thoughts.

P&C

Prompt & Circuit Editorial

Contributing Editor

Articles on AI infrastructure for cloud architects, written from official vendor sources and checked claim by claim before publishing.