diff --git a/generator/test_case_generator.go b/generator/test_case_generator.go
index aee64a39..3b656f65 100644
--- a/generator/test_case_generator.go
+++ b/generator/test_case_generator.go
@@ -526,6 +526,14 @@ var testTypeToTestConfig = map[string][]testConfig{
 			ami:          "AL2023_x86_64_STANDARD",
 			k8sVersion:   "1.35",
 		},
+		{
+			testDir:      "./test/otel/multi_efa_dra",
+			terraformDir: "terraform/eks/daemon/otel-multi-efa-dra",
+			targets:      map[string]map[string]struct{}{"arc": {"amd64": {}}},
+			instanceType: "c6in.32xlarge",
+			ami:          "AL2023_x86_64_STANDARD",
+			k8sVersion:   "1.35",
+		},
 		{
 			testDir:      "./test/otel/neuron",
 			terraformDir: "terraform/eks/daemon/otel-neuron",
@@ -534,6 +542,22 @@ var testTypeToTestConfig = map[string][]testConfig{
 			ami:          "AL2023_x86_64_NEURON",
 			k8sVersion:   "1.35",
 		},
+		{
+			testDir:      "./test/otel/neuron_dra",
+			terraformDir: "terraform/eks/daemon/otel-neuron-dra",
+			targets:      map[string]map[string]struct{}{"arc": {"amd64": {}}},
+			instanceType: "trn1.2xlarge",
+			ami:          "AL2023_x86_64_NEURON",
+			k8sVersion:   "1.35",
+		},
+		{
+			testDir:      "./test/otel/gpu_dra",
+			terraformDir: "terraform/eks/daemon/otel-gpu-dra",
+			targets:      map[string]map[string]struct{}{"arc": {"amd64": {}}},
+			instanceType: "g4dn.12xlarge",
+			ami:          "AL2023_x86_64_NVIDIA",
+			k8sVersion:   "1.35",
+		},
 		{
 			testDir:      "./test/otel/performance",
 			terraformDir: "terraform/eks/daemon/otel-performance",
diff --git a/terraform/eks/daemon/otel-gpu-dra/main.tf b/terraform/eks/daemon/otel-gpu-dra/main.tf
new file mode 100644
index 00000000..c1bd650b
--- /dev/null
+++ b/terraform/eks/daemon/otel-gpu-dra/main.tf
@@ -0,0 +1,350 @@
+# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: MIT
+
+# GPU DRA-path integration test cluster. Mirrors otel-gpu but exposes GPUs via the
+# NVIDIA DRA driver (DeviceClass gpu.nvidia.com) instead of the device plugin, and
+# the burn workload claims a GPU via a ResourceClaimTemplate. One multi-GPU node
+# (g4dn.12xlarge = 4 T4) proves per-device DRA correlation (1 claimed GPU -> the
+# burn pod, the other 3 uncorrelated).
+
+module "common" {
+  source             = "../../../common"
+  cwagent_image_repo = var.cwagent_image_repo
+  cwagent_image_tag  = var.cwagent_image_tag
+}
+
+module "basic_components" {
+  source = "../../../basic_components"
+  region = var.region
+}
+
+locals {
+  aws_eks = "aws eks --region ${var.region}"
+}
+
+resource "aws_eks_cluster" "this" {
+  name     = "cwagent-eks-integ-${module.common.testing_id}"
+  role_arn = module.basic_components.role_arn
+  version  = var.k8s_version
+  vpc_config {
+    subnet_ids         = module.basic_components.public_subnet_ids
+    security_group_ids = [module.basic_components.security_group]
+  }
+}
+
+# --- IAM ---
+
+resource "aws_iam_role" "node_role" {
+  name = "cwagent-otel-gpu-dra-Worker-Role-${module.common.testing_id}"
+  assume_role_policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [{
+      Effect    = "Allow"
+      Principal = { Service = "ec2.amazonaws.com" }
+      Action    = "sts:AssumeRole"
+    }]
+  })
+}
+
+resource "aws_iam_role_policy_attachment" "node_AmazonEKSWorkerNodePolicy" {
+  policy_arn = "arn:aws:iam::aws:policy/AmazonEKSWorkerNodePolicy"
+  role       = aws_iam_role.node_role.name
+}
+
+resource "aws_iam_role_policy_attachment" "node_AmazonEKS_CNI_Policy" {
+  policy_arn = "arn:aws:iam::aws:policy/AmazonEKS_CNI_Policy"
+  role       = aws_iam_role.node_role.name
+}
+
+resource "aws_iam_role_policy_attachment" "node_AmazonEC2ContainerRegistryReadOnly" {
+  policy_arn = "arn:aws:iam::aws:policy/AmazonEC2ContainerRegistryReadOnly"
+  role       = aws_iam_role.node_role.name
+}
+
+resource "aws_iam_role" "pod_identity_role" {
+  name = "cwagent-otel-gpu-dra-pod-identity-${module.common.testing_id}"
+  assume_role_policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [{
+      Effect    = "Allow"
+      Principal = { Service = "pods.eks.amazonaws.com" }
+      Action    = ["sts:AssumeRole", "sts:TagSession"]
+    }]
+  })
+}
+
+resource "aws_iam_role_policy_attachment" "pod_identity_CloudWatchAgentServerPolicy" {
+  policy_arn = "arn:aws:iam::aws:policy/CloudWatchAgentServerPolicy"
+  role       = aws_iam_role.pod_identity_role.name
+}
+
+# --- Node Groups ---
+
+resource "aws_eks_node_group" "standard" {
+  cluster_name    = aws_eks_cluster.this.name
+  node_group_name = "standard-${module.common.testing_id}"
+  node_role_arn   = aws_iam_role.node_role.arn
+  subnet_ids      = module.basic_components.public_subnet_ids
+
+  scaling_config {
+    desired_size = 1
+    max_size     = 1
+    min_size     = 1
+  }
+
+  ami_type       = "AL2023_x86_64_STANDARD"
+  capacity_type  = "ON_DEMAND"
+  disk_size      = 20
+  instance_types = ["t3.medium"]
+
+  depends_on = [
+    aws_iam_role_policy_attachment.node_AmazonEC2ContainerRegistryReadOnly,
+    aws_iam_role_policy_attachment.node_AmazonEKS_CNI_Policy,
+    aws_iam_role_policy_attachment.node_AmazonEKSWorkerNodePolicy,
+  ]
+}
+
+# Multi-GPU node (g4dn.12xlarge, 4 GPUs). No taint: dedicated by being the only GPU
+# node; the DRA driver DaemonSet + burn pod land here via scheduling/nodeSelector.
+resource "aws_eks_node_group" "gpu_multi" {
+  cluster_name    = aws_eks_cluster.this.name
+  node_group_name = "gpu-multi-${module.common.testing_id}"
+  node_role_arn   = aws_iam_role.node_role.arn
+  subnet_ids      = module.basic_components.public_subnet_ids
+
+  scaling_config {
+    desired_size = 1
+    max_size     = 1
+    min_size     = 1
+  }
+
+  ami_type       = var.ami_type
+  capacity_type  = "ON_DEMAND"
+  disk_size      = 40
+  instance_types = [var.instance_type]
+
+  labels = {
+    "nvidia.com/gpu.present"         = "true"
+    "ci-test.example.com/node-color" = "green"
+  }
+
+  depends_on = [
+    aws_iam_role_policy_attachment.node_AmazonEC2ContainerRegistryReadOnly,
+    aws_iam_role_policy_attachment.node_AmazonEKS_CNI_Policy,
+    aws_iam_role_policy_attachment.node_AmazonEKSWorkerNodePolicy,
+  ]
+}
+
+# --- EKS Addon: Pod Identity agent ---
+
+resource "aws_eks_addon" "pod_identity_agent" {
+  depends_on   = [aws_eks_node_group.standard]
+  cluster_name = aws_eks_cluster.this.name
+  addon_name   = "eks-pod-identity-agent"
+}
+
+# --- Update kubeconfig ---
+
+resource "null_resource" "kubectl" {
+  depends_on = [aws_eks_cluster.this, aws_eks_node_group.standard, aws_eks_node_group.gpu_multi]
+  provisioner "local-exec" {
+    command = "${local.aws_eks} update-kubeconfig --name ${aws_eks_cluster.this.name}"
+  }
+}
+
+# --- NVIDIA DRA driver (Helm) — replaces the device plugin ---
+#
+# NVIDIA DRA driver install, tuned for a single-node whole-GPU EKS cluster. Verified
+# against chart 25.12.0:
+#   - resources.gpus.enabled=true turns on whole-GPU allocation (DeviceClass
+#     gpu.nvidia.com), but the chart hard-guards it behind gpuResourcesEnabledOverride
+#     =true (it refuses to co-exist with the standard GPU device plugin until KEP 5004
+#     is GA). We do not run the device plugin, so the override is safe and required —
+#     without it helm template fails a validation.yaml assertion.
+#   - resources.computeDomains.enabled=false drops the ComputeDomain controller, whose
+#     nodeAffinity requires node-role.kubernetes.io/control-plane. EKS has no such
+#     (customer-visible) nodes, so the controller would stay Pending and, with helm's
+#     default wait=true, time out the apply. We only need single-node whole-GPU
+#     correlation, so ComputeDomains (multi-node GPU/IMEX) is unnecessary.
+#   - kubeletPlugin.nodeSelector pins the plugin DaemonSet to the GPU node. The chart's
+#     default kubeletPlugin nodeAffinity already ORs in nvidia.com/gpu.present=true
+#     (which our nodegroup sets), so no affinity clearing is needed on this chart.
+# After apply, `kubectl get deviceclass gpu.nvidia.com` and a ResourceSlice for the
+# node should exist.
+resource "helm_release" "nvidia_dra_driver" {
+  depends_on = [aws_eks_node_group.gpu_multi, null_resource.kubectl]
+
+  name             = "nvidia-dra-driver-gpu"
+  repository       = var.nvidia_dra_repo
+  chart            = "nvidia-dra-driver-gpu"
+  namespace        = "nvidia-dra-driver-gpu"
+  create_namespace = true
+  version          = var.nvidia_dra_chart_version != "" ? var.nvidia_dra_chart_version : null
+
+  set = [
+    { name = "gpuResourcesEnabledOverride", value = "true" },
+    { name = "resources.gpus.enabled", value = "true" },
+    { name = "resources.computeDomains.enabled", value = "false" },
+    # nodeSelector map values must be strings; type=string stops the provider
+    # coercing "true" to a bool (which fails DaemonSet unmarshalling).
+    { name = "kubeletPlugin.nodeSelector.nvidia\\.com/gpu\\.present", value = "true", type = "string" },
+  ]
+}
+
+# --- multi-gpu-burn-dra Deployment: claims 1 GPU via DRA ---
+
+resource "null_resource" "gpu_burn_dra" {
+  depends_on = [helm_release.nvidia_dra_driver, null_resource.kubectl]
+  provisioner "local-exec" {
+    command = <<-EOT
+      cat <<'EOF' | kubectl apply -f -
+      apiVersion: resource.k8s.io/v1
+      kind: ResourceClaimTemplate
+      metadata:
+        name: gpu-1
+        namespace: default
+      spec:
+        spec:
+          devices:
+            requests:
+            - name: gpu
+              exactly:
+                deviceClassName: gpu.nvidia.com
+                count: 1
+      ---
+      apiVersion: apps/v1
+      kind: Deployment
+      metadata:
+        name: gpu-burn-dra
+        namespace: default
+      spec:
+        replicas: 1
+        revisionHistoryLimit: 2
+        progressDeadlineSeconds: 300
+        selector:
+          matchLabels:
+            app: gpu-burn-dra
+        template:
+          metadata:
+            labels:
+              app: gpu-burn-dra
+              ci-test.example.com/pod-color: magenta
+          spec:
+            nodeSelector:
+              node.kubernetes.io/instance-type: ${var.instance_type}
+            resourceClaims:
+            - name: gpu
+              resourceClaimTemplateName: gpu-1
+            containers:
+            - name: gpu-burn
+              image: chrstnhntschl/gpu_burn:latest
+              args: ["3600"]
+              resources:
+                claims:
+                - name: gpu
+      EOF
+    EOT
+  }
+}
+
+# --- Helm chart install (observability) ---
+
+data "external" "clone_helm_chart" {
+  program = ["bash", "-c", <<-EOT
+    rm -rf ./helm-charts
+    git clone -b ${var.helm_chart_branch} ${var.helm_chart_repo_url} ./helm-charts
+    echo '{"status":"ready"}'
+  EOT
+  ]
+}
+
+resource "helm_release" "aws_observability" {
+  name             = "amazon-cloudwatch-observability"
+  chart            = "./helm-charts/charts/amazon-cloudwatch-observability"
+  namespace        = "amazon-cloudwatch"
+  create_namespace = true
+  wait             = false
+  timeout          = 600
+
+  set = [
+    { name = "clusterName", value = aws_eks_cluster.this.name },
+    { name = "region", value = var.region },
+    { name = "otelContainerInsights.enabled", value = "true" },
+  ]
+
+  depends_on = [
+    aws_eks_addon.pod_identity_agent,
+    null_resource.kubectl,
+    data.external.clone_helm_chart,
+  ]
+}
+
+# --- Pod Identity association ---
+
+resource "aws_eks_pod_identity_association" "cloudwatch_agent" {
+  depends_on      = [helm_release.aws_observability]
+  cluster_name    = aws_eks_cluster.this.name
+  namespace       = "amazon-cloudwatch"
+  service_account = "cloudwatch-agent"
+  role_arn        = aws_iam_role.pod_identity_role.arn
+}
+
+# --- Patch agent image ---
+
+resource "null_resource" "update_image" {
+  depends_on = [helm_release.aws_observability, null_resource.kubectl]
+  triggers   = { timestamp = timestamp() }
+  provisioner "local-exec" {
+    command = <<-EOT
+      sleep 30
+      kubectl -n amazon-cloudwatch patch AmazonCloudWatchAgent cloudwatch-agent --type='json' \
+        -p='[{"op": "replace", "path": "/spec/image", "value": "${var.cwagent_image_repo}:${var.cwagent_image_tag}"}]'
+      kubectl -n amazon-cloudwatch patch AmazonCloudWatchAgent cloudwatch-agent-cluster-scraper --type='json' \
+        -p='[{"op": "replace", "path": "/spec/image", "value": "${var.cwagent_image_repo}:${var.cwagent_image_tag}"}]' 2>/dev/null || true
+      sleep 10
+    EOT
+  }
+}
+
+# --- Restart pods ---
+
+resource "null_resource" "restart_pods" {
+  depends_on = [aws_eks_pod_identity_association.cloudwatch_agent, null_resource.update_image]
+  triggers   = { timestamp = timestamp() }
+  provisioner "local-exec" {
+    command = <<-EOT
+      kubectl -n amazon-cloudwatch rollout restart daemonset/cloudwatch-agent
+      kubectl -n amazon-cloudwatch rollout restart deployment/cloudwatch-agent-cluster-scraper 2>/dev/null || true
+      kubectl -n amazon-cloudwatch rollout status daemonset/cloudwatch-agent --timeout=120s
+    EOT
+  }
+}
+
+# --- Test runner ---
+
+resource "null_resource" "validator" {
+  depends_on = [null_resource.restart_pods, null_resource.gpu_burn_dra]
+  triggers   = { always_run = timestamp() }
+  provisioner "local-exec" {
+    command = <<-EOT
+      echo "Running OTEL GPU DRA cluster integration tests"
+      cd ../../../..
+
+      echo "Waiting for dcgm-exporter pods to be ready..."
+      for i in $(seq 1 30); do
+        READY=$(kubectl get pods -n amazon-cloudwatch -l app.kubernetes.io/name=dcgm-exporter -o jsonpath='{.items[*].status.phase}' 2>/dev/null | tr ' ' '\n' | grep -c Running || true)
+        if [ "$READY" -ge 1 ] 2>/dev/null; then break; fi
+        sleep 20
+      done
+
+      echo "Waiting 6 minutes for GPU metrics to propagate (Zeus 5-min staleness window)..."
+      sleep 360
+
+      go test -tags integration -timeout 1h -v ${var.test_dir} \
+        -eksClusterName=${aws_eks_cluster.this.name} \
+        -computeType=EKS \
+        -eksDeploymentStrategy=DAEMON \
+        -region=${var.region}
+    EOT
+  }
+}
diff --git a/terraform/eks/daemon/otel-gpu-dra/providers.tf b/terraform/eks/daemon/otel-gpu-dra/providers.tf
new file mode 100644
index 00000000..a562a7f1
--- /dev/null
+++ b/terraform/eks/daemon/otel-gpu-dra/providers.tf
@@ -0,0 +1,28 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+provider "aws" {
+  region = var.region
+}
+
+provider "kubernetes" {
+  host                   = aws_eks_cluster.this.endpoint
+  cluster_ca_certificate = base64decode(aws_eks_cluster.this.certificate_authority.0.data)
+  exec {
+    api_version = "client.authentication.k8s.io/v1beta1"
+    command     = "aws"
+    args        = ["eks", "get-token", "--cluster-name", aws_eks_cluster.this.name]
+  }
+}
+
+provider "helm" {
+  kubernetes = {
+    host                   = aws_eks_cluster.this.endpoint
+    cluster_ca_certificate = base64decode(aws_eks_cluster.this.certificate_authority.0.data)
+    exec = {
+      api_version = "client.authentication.k8s.io/v1beta1"
+      command     = "aws"
+      args        = ["eks", "get-token", "--cluster-name", aws_eks_cluster.this.name]
+    }
+  }
+}
diff --git a/terraform/eks/daemon/otel-gpu-dra/variables.tf b/terraform/eks/daemon/otel-gpu-dra/variables.tf
new file mode 100644
index 00000000..1c689d17
--- /dev/null
+++ b/terraform/eks/daemon/otel-gpu-dra/variables.tf
@@ -0,0 +1,66 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+variable "region" {
+  type    = string
+  default = "us-west-2"
+}
+
+variable "test_dir" {
+  type    = string
+  default = "./test/otel/gpu_dra"
+}
+
+variable "cwagent_image_repo" {
+  type    = string
+  default = "public.ecr.aws/cloudwatch-agent/cloudwatch-agent"
+}
+
+variable "cwagent_image_tag" {
+  type    = string
+  default = "latest"
+}
+
+# The chart must render the DRA correlation config (dra_device_types incl. the
+# gpu-dra entry keyed on the gpu.nvidia.com driver) and grant the agent
+# ServiceAccount resource.k8s.io RBAC. Override to a fork/branch until that lands
+# on main (helm-charts PR #356).
+variable "helm_chart_branch" {
+  type    = string
+  default = "main"
+}
+
+variable "helm_chart_repo_url" {
+  type    = string
+  default = "https://github.com/aws-observability/helm-charts.git"
+}
+
+# The awsdevicepodcorrelation processor uses the GA resource.k8s.io/v1 DRA API
+# (available since Kubernetes 1.34), so the cluster must be 1.34+.
+variable "k8s_version" {
+  type    = string
+  default = "1.35"
+}
+
+variable "ami_type" {
+  type    = string
+  default = "AL2023_x86_64_NVIDIA"
+}
+
+# Multi-GPU node: g4dn.12xlarge = 4 T4 GPUs.
+variable "instance_type" {
+  type    = string
+  default = "g4dn.12xlarge"
+}
+
+# NVIDIA DRA driver Helm chart (DeviceClass gpu.nvidia.com).
+# Repo: https://helm.ngc.nvidia.com/nvidia, chart: nvidia-dra-driver-gpu.
+variable "nvidia_dra_repo" {
+  type    = string
+  default = "https://helm.ngc.nvidia.com/nvidia"
+}
+
+variable "nvidia_dra_chart_version" {
+  type    = string
+  default = ""
+}
diff --git a/terraform/eks/daemon/otel-multi-efa-dra/.gitignore b/terraform/eks/daemon/otel-multi-efa-dra/.gitignore
new file mode 100644
index 00000000..7a0bdf43
--- /dev/null
+++ b/terraform/eks/daemon/otel-multi-efa-dra/.gitignore
@@ -0,0 +1,3 @@
+*.tfstate*
+.terraform/
+helm-charts/
diff --git a/terraform/eks/daemon/otel-multi-efa-dra/cloudwatch-pod-identity.tf b/terraform/eks/daemon/otel-multi-efa-dra/cloudwatch-pod-identity.tf
new file mode 100644
index 00000000..a09bf5a0
--- /dev/null
+++ b/terraform/eks/daemon/otel-multi-efa-dra/cloudwatch-pod-identity.tf
@@ -0,0 +1,60 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+# IAM role for CloudWatch Observability addon
+resource "aws_iam_role" "cloudwatch_observability" {
+  name = "cloudwatch-observability-${module.common.testing_id}"
+
+  assume_role_policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [
+      {
+        Effect = "Allow"
+        Principal = {
+          Service = "pods.eks.amazonaws.com"
+        }
+        Action = [
+          "sts:AssumeRole",
+          "sts:TagSession"
+        ]
+      }
+    ]
+  })
+
+  tags = {
+    Name  = "cloudwatch-observability-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Attach CloudWatch policies
+resource "aws_iam_role_policy_attachment" "cloudwatch_observability_server" {
+  policy_arn = "arn:aws:iam::aws:policy/CloudWatchAgentServerPolicy"
+  role       = aws_iam_role.cloudwatch_observability.name
+}
+
+# Pod Identity association for CloudWatch agent
+resource "aws_eks_pod_identity_association" "cloudwatch_agent" {
+  cluster_name    = module.eks.cluster_name
+  namespace       = "amazon-cloudwatch"
+  service_account = "cloudwatch-agent"
+  role_arn        = aws_iam_role.cloudwatch_observability.arn
+
+  tags = {
+    Name  = "cloudwatch-agent-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Pod Identity association for Fluent Bit
+resource "aws_eks_pod_identity_association" "fluent_bit" {
+  cluster_name    = module.eks.cluster_name
+  namespace       = "amazon-cloudwatch"
+  service_account = "fluent-bit"
+  role_arn        = aws_iam_role.cloudwatch_observability.arn
+
+  tags = {
+    Name  = "fluent-bit-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
diff --git a/terraform/eks/daemon/otel-multi-efa-dra/main.tf b/terraform/eks/daemon/otel-multi-efa-dra/main.tf
new file mode 100644
index 00000000..798f5ea5
--- /dev/null
+++ b/terraform/eks/daemon/otel-multi-efa-dra/main.tf
@@ -0,0 +1,244 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+locals {
+  aws_eks = "aws eks --region ${var.region}"
+}
+
+module "common" {
+  source             = "../../../common"
+  cwagent_image_repo = var.cwagent_image_repo
+  cwagent_image_tag  = var.cwagent_image_tag
+}
+
+module "eks" {
+  source  = "terraform-aws-modules/eks/aws"
+  version = "~> 21.0"
+
+  name               = "cwagent-eks-integ-${module.common.testing_id}"
+  kubernetes_version = var.k8s_version
+
+  # Full role name (<=64); the module's "<name>-cluster-" name_prefix would exceed the 38-char cap.
+  iam_role_use_name_prefix = false
+
+  vpc_id     = aws_vpc.efa_test_vpc.id
+  subnet_ids = aws_subnet.efa_test_public_subnet[*].id
+
+  endpoint_public_access                   = true
+  enable_cluster_creator_admin_permissions = true
+
+  eks_managed_node_groups = {
+    standard = {
+      ami_type       = "AL2023_x86_64_STANDARD"
+      instance_types = ["t3.medium"]
+      min_size       = 1
+      max_size       = 1
+      desired_size   = 1
+      subnet_ids     = aws_subnet.efa_test_public_subnet[*].id
+    }
+
+    multi_efa = {
+      # enable_efa_support wires EFA network interfaces into the launch template.
+      # It does NOT install the EFA k8s device plugin (that is a separate helm
+      # release on the device-plugin path); here EFA is exposed via DRA instead.
+      enable_efa_support = true
+      ami_type           = var.ami_type
+      instance_types     = [var.instance_type]
+      min_size           = 1
+      max_size           = 1
+      desired_size       = 1
+      subnet_ids         = aws_subnet.efa_test_private_subnet[*].id
+
+      labels = {
+        "ci-test.example.com/multi-efa-dra" = "true"
+        "ci-test.example.com/node-color"    = "yellow"
+      }
+    }
+  }
+
+  addons = {
+    coredns                = {}
+    eks-pod-identity-agent = { before_compute = true }
+    kube-proxy             = {}
+    vpc-cni                = { before_compute = true }
+  }
+
+  tags = { Owner = "default" }
+}
+
+# --- DRA driver (dranet) ---
+# Replaces the EFA device plugin: exposes the node's EFA interfaces as DRA
+# devices under the DeviceClass efa.networking.k8s.aws (driver name "dra.net")
+# and publishes a ResourceSlice carrying the dra.net/rdmaDevice attribute that
+# the processor keys on. dranet and the EFA device plugin must not both manage
+# EFA on the same node, so the device plugin is intentionally not installed here.
+
+resource "helm_release" "aws_dranet" {
+  name       = "aws-dranet"
+  repository = "https://aws.github.io/eks-charts"
+  chart      = "aws-dranet"
+  version    = var.dranet_version
+  namespace  = "kube-system"
+  wait       = true
+
+  depends_on = [module.eks]
+}
+
+# --- Helm chart install ---
+
+resource "null_resource" "kubectl" {
+  depends_on = [module.eks]
+  provisioner "local-exec" {
+    command = "${local.aws_eks} update-kubeconfig --name ${module.eks.cluster_name}"
+  }
+}
+
+data "external" "clone_helm_chart" {
+  program = ["bash", "-c", <<-EOT
+    rm -rf ./helm-charts
+    git clone -b ${var.helm_chart_branch} ${var.helm_chart_repo_url} ./helm-charts
+    echo '{"status":"ready"}'
+  EOT
+  ]
+}
+
+resource "helm_release" "aws_observability" {
+  name             = "amazon-cloudwatch-observability"
+  chart            = "./helm-charts/charts/amazon-cloudwatch-observability"
+  namespace        = "amazon-cloudwatch"
+  create_namespace = true
+  wait             = false
+  timeout          = 600
+
+  set = [
+    { name = "clusterName", value = module.eks.cluster_name },
+    { name = "region", value = var.region },
+    { name = "otelContainerInsights.enabled", value = "true" },
+  ]
+
+  depends_on = [
+    module.eks,
+    null_resource.kubectl,
+    data.external.clone_helm_chart,
+  ]
+}
+
+# --- Patch agent image ---
+
+resource "null_resource" "update_image" {
+  depends_on = [helm_release.aws_observability, null_resource.kubectl]
+  triggers   = { timestamp = timestamp() }
+  provisioner "local-exec" {
+    command = <<-EOT
+      sleep 30
+      kubectl -n amazon-cloudwatch patch AmazonCloudWatchAgent cloudwatch-agent --type='json' \
+        -p='[{"op": "replace", "path": "/spec/image", "value": "${var.cwagent_image_repo}:${var.cwagent_image_tag}"}]'
+      kubectl -n amazon-cloudwatch patch AmazonCloudWatchAgent cloudwatch-agent-cluster-scraper --type='json' \
+        -p='[{"op": "replace", "path": "/spec/image", "value": "${var.cwagent_image_repo}:${var.cwagent_image_tag}"}]' 2>/dev/null || true
+      sleep 10
+    EOT
+  }
+}
+
+# --- Restart pods ---
+
+resource "null_resource" "restart_pods" {
+  depends_on = [null_resource.update_image]
+  triggers   = { timestamp = timestamp() }
+  provisioner "local-exec" {
+    command = <<-EOT
+      kubectl -n amazon-cloudwatch rollout restart daemonset/cloudwatch-agent
+      kubectl -n amazon-cloudwatch rollout restart deployment/cloudwatch-agent-cluster-scraper 2>/dev/null || true
+      kubectl -n amazon-cloudwatch rollout status daemonset/cloudwatch-agent --timeout=120s
+    EOT
+  }
+}
+
+# --- efaburn workload (DRA) ---
+# efaburn claims 1 EFA via a ResourceClaimTemplate, so on a 2-EFA node exactly
+# one device is claimed and the other stays unclaimed — exercising both the
+# claimed->pod and unclaimed->no-pod correlation paths.
+
+resource "null_resource" "efaburn" {
+  depends_on = [module.eks, null_resource.kubectl, helm_release.aws_dranet]
+  provisioner "local-exec" {
+    command = <<-EOT
+      cat <<'EOF' | kubectl apply -f -
+      apiVersion: resource.k8s.io/v1
+      kind: ResourceClaimTemplate
+      metadata:
+        name: efa-claim-template
+        namespace: default
+      spec:
+        spec:
+          devices:
+            requests:
+            - name: efa-device
+              exactly:
+                deviceClassName: efa.networking.k8s.aws
+                count: 1
+      ---
+      apiVersion: apps/v1
+      kind: Deployment
+      metadata:
+        name: efaburn
+        namespace: default
+      spec:
+        replicas: 1
+        selector:
+          matchLabels:
+            app: efaburn
+        template:
+          metadata:
+            labels:
+              app: efaburn
+              ci-test.example.com/pod-color: teal
+          spec:
+            nodeSelector:
+              ci-test.example.com/multi-efa-dra: "true"
+            resourceClaims:
+            - name: efa-device
+              resourceClaimTemplateName: efa-claim-template
+            containers:
+            - name: efaburn
+              image: ${var.efaburn_image}
+              resources:
+                claims:
+                - name: efa-device
+                requests:
+                  memory: 8000Mi
+              securityContext:
+                allowPrivilegeEscalation: false
+                runAsNonRoot: true
+                runAsUser: 1000
+      EOF
+    EOT
+  }
+}
+
+# --- Test runner ---
+
+resource "null_resource" "validator" {
+  depends_on = [
+    null_resource.restart_pods,
+    null_resource.efaburn,
+  ]
+
+  triggers = { always_run = timestamp() }
+
+  provisioner "local-exec" {
+    command = <<-EOT
+      echo "Running OTEL Multi-EFA DRA cluster integration tests"
+      cd ../../../..
+
+      echo "Waiting 3 minutes for metrics to propagate..."
+      sleep 180
+
+      go test -tags integration -timeout 1h -v ${var.test_dir} \
+        -eksClusterName=${module.eks.cluster_name} \
+        -computeType=EKS \
+        -eksDeploymentStrategy=DAEMON \
+        -region=${var.region}
+    EOT
+  }
+}
diff --git a/terraform/eks/daemon/otel-multi-efa-dra/providers.tf b/terraform/eks/daemon/otel-multi-efa-dra/providers.tf
new file mode 100644
index 00000000..581e0b6b
--- /dev/null
+++ b/terraform/eks/daemon/otel-multi-efa-dra/providers.tf
@@ -0,0 +1,47 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+terraform {
+  required_providers {
+    aws = {
+      source  = "hashicorp/aws"
+      version = "!= 6.22.0"
+    }
+    helm = {
+      source  = "hashicorp/helm"
+      version = "~> 3.0"
+    }
+    kubernetes = {
+      source  = "hashicorp/kubernetes"
+      version = "~> 2.0"
+    }
+  }
+}
+
+provider "aws" {
+  region = var.region
+}
+
+provider "helm" {
+  kubernetes = {
+    host                   = module.eks.cluster_endpoint
+    cluster_ca_certificate = base64decode(module.eks.cluster_certificate_authority_data)
+
+    exec = {
+      api_version = "client.authentication.k8s.io/v1beta1"
+      command     = "aws"
+      args        = ["eks", "get-token", "--cluster-name", module.eks.cluster_name]
+    }
+  }
+}
+
+provider "kubernetes" {
+  host                   = module.eks.cluster_endpoint
+  cluster_ca_certificate = base64decode(module.eks.cluster_certificate_authority_data)
+
+  exec {
+    api_version = "client.authentication.k8s.io/v1beta1"
+    command     = "aws"
+    args        = ["eks", "get-token", "--cluster-name", module.eks.cluster_name]
+  }
+}
diff --git a/terraform/eks/daemon/otel-multi-efa-dra/variables.tf b/terraform/eks/daemon/otel-multi-efa-dra/variables.tf
new file mode 100644
index 00000000..6351f17f
--- /dev/null
+++ b/terraform/eks/daemon/otel-multi-efa-dra/variables.tf
@@ -0,0 +1,68 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+variable "region" {
+  type    = string
+  default = "us-west-2"
+}
+
+variable "test_dir" {
+  type    = string
+  default = "./test/otel/multi_efa_dra"
+}
+
+variable "cwagent_image_repo" {
+  type    = string
+  default = "public.ecr.aws/cloudwatch-agent/cloudwatch-agent"
+}
+
+variable "cwagent_image_tag" {
+  type    = string
+  default = "latest"
+}
+
+# The chart must render the DRA correlation config (dra_device_types keyed on the
+# dra.net driver + the dra.net/rdmaDevice ResourceSlice attribute) and grant the
+# agent ServiceAccount get/list/watch on resource.k8s.io resourceclaims and
+# resourceslices. Until that lands on main, point helm_chart_repo_url/branch at
+# the fork/branch that carries it (helm-charts PR #356).
+variable "helm_chart_branch" {
+  type    = string
+  default = "main"
+}
+
+# Repository to clone the observability Helm chart from. Override to a fork when
+# validating chart changes that are not yet merged upstream.
+variable "helm_chart_repo_url" {
+  type    = string
+  default = "https://github.com/aws-observability/helm-charts.git"
+}
+
+# The awsdevicepodcorrelation processor uses the GA resource.k8s.io/v1 DRA API
+# (available since Kubernetes 1.34), so the cluster must be 1.34+.
+variable "k8s_version" {
+  type    = string
+  default = "1.35"
+}
+
+variable "ami_type" {
+  type    = string
+  default = "AL2023_x86_64_STANDARD"
+}
+
+variable "instance_type" {
+  type    = string
+  default = "c6in.32xlarge"
+}
+
+# dranet Helm chart version (eks/aws-dranet, from https://aws.github.io/eks-charts).
+# This is the CHART version (1.0.0); the app version it ships is v1.2.0-eksbuild.2.
+variable "dranet_version" {
+  type    = string
+  default = "1.0.0"
+}
+
+variable "efaburn_image" {
+  type    = string
+  default = "506463145083.dkr.ecr.us-west-2.amazonaws.com/efaburn:latest"
+}
diff --git a/terraform/eks/daemon/otel-multi-efa-dra/vpc-endpoints.tf b/terraform/eks/daemon/otel-multi-efa-dra/vpc-endpoints.tf
new file mode 100644
index 00000000..6a136138
--- /dev/null
+++ b/terraform/eks/daemon/otel-multi-efa-dra/vpc-endpoints.tf
@@ -0,0 +1,82 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+# VPC Endpoints for private subnet access
+resource "aws_vpc_endpoint" "ecr_dkr" {
+  vpc_id              = aws_vpc.efa_test_vpc.id
+  service_name        = "com.amazonaws.${var.region}.ecr.dkr"
+  vpc_endpoint_type   = "Interface"
+  subnet_ids          = aws_subnet.efa_test_private_subnet[*].id
+  security_group_ids  = [aws_security_group.vpc_endpoints.id]
+  private_dns_enabled = true
+
+  tags = {
+    Name  = "efa-test-ecr-dkr-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+resource "aws_vpc_endpoint" "ecr_api" {
+  vpc_id              = aws_vpc.efa_test_vpc.id
+  service_name        = "com.amazonaws.${var.region}.ecr.api"
+  vpc_endpoint_type   = "Interface"
+  subnet_ids          = aws_subnet.efa_test_private_subnet[*].id
+  security_group_ids  = [aws_security_group.vpc_endpoints.id]
+  private_dns_enabled = true
+
+  tags = {
+    Name  = "efa-test-ecr-api-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+resource "aws_vpc_endpoint" "s3" {
+  vpc_id            = aws_vpc.efa_test_vpc.id
+  service_name      = "com.amazonaws.${var.region}.s3"
+  vpc_endpoint_type = "Gateway"
+  route_table_ids   = [aws_route_table.efa_test_private_rt.id]
+
+  tags = {
+    Name  = "efa-test-s3-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+resource "aws_vpc_endpoint" "eks" {
+  vpc_id              = aws_vpc.efa_test_vpc.id
+  service_name        = "com.amazonaws.${var.region}.eks"
+  vpc_endpoint_type   = "Interface"
+  subnet_ids          = aws_subnet.efa_test_private_subnet[*].id
+  security_group_ids  = [aws_security_group.vpc_endpoints.id]
+  private_dns_enabled = true
+
+  tags = {
+    Name  = "efa-test-eks-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Security group for VPC endpoints
+resource "aws_security_group" "vpc_endpoints" {
+  name_prefix = "efa-test-vpc-endpoints-${module.common.testing_id}"
+  vpc_id      = aws_vpc.efa_test_vpc.id
+
+  ingress {
+    from_port   = 443
+    to_port     = 443
+    protocol    = "tcp"
+    cidr_blocks = [aws_vpc.efa_test_vpc.cidr_block]
+  }
+
+  egress {
+    from_port   = 0
+    to_port     = 0
+    protocol    = "-1"
+    cidr_blocks = ["0.0.0.0/0"]
+  }
+
+  tags = {
+    Name  = "efa-test-vpc-endpoints-sg-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
diff --git a/terraform/eks/daemon/otel-multi-efa-dra/vpc.tf b/terraform/eks/daemon/otel-multi-efa-dra/vpc.tf
new file mode 100644
index 00000000..1db6495f
--- /dev/null
+++ b/terraform/eks/daemon/otel-multi-efa-dra/vpc.tf
@@ -0,0 +1,152 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+# VPC for EFA integration test
+resource "aws_vpc" "efa_test_vpc" {
+  cidr_block           = "10.0.0.0/16"
+  enable_dns_hostnames = true
+  enable_dns_support   = true
+
+  tags = {
+    Name  = "efa-test-vpc-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Internet Gateway
+resource "aws_internet_gateway" "efa_test_igw" {
+  vpc_id = aws_vpc.efa_test_vpc.id
+
+  tags = {
+    Name  = "efa-test-igw-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Public Subnets
+resource "aws_subnet" "efa_test_public_subnet" {
+  count = 2
+
+  vpc_id                  = aws_vpc.efa_test_vpc.id
+  cidr_block              = "10.0.${count.index + 1}.0/24"
+  availability_zone       = data.aws_availability_zones.available.names[count.index]
+  map_public_ip_on_launch = true
+
+  tags = {
+    Name  = "efa-test-public-subnet-${count.index + 1}-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Route Table for Public Subnets
+resource "aws_route_table" "efa_test_public_rt" {
+  vpc_id = aws_vpc.efa_test_vpc.id
+
+  route {
+    cidr_block = "0.0.0.0/0"
+    gateway_id = aws_internet_gateway.efa_test_igw.id
+  }
+
+  tags = {
+    Name  = "efa-test-public-rt-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Associate Route Table with Public Subnets
+resource "aws_route_table_association" "efa_test_public_rta" {
+  count = length(aws_subnet.efa_test_public_subnet)
+
+  subnet_id      = aws_subnet.efa_test_public_subnet[count.index].id
+  route_table_id = aws_route_table.efa_test_public_rt.id
+}
+
+# Security Group
+resource "aws_security_group" "efa_test_sg" {
+  name_prefix = "efa-test-sg-${module.common.testing_id}"
+  vpc_id      = aws_vpc.efa_test_vpc.id
+
+  ingress {
+    from_port = 0
+    to_port   = 65535
+    protocol  = "tcp"
+    self      = true
+  }
+
+  egress {
+    from_port   = 0
+    to_port     = 0
+    protocol    = "-1"
+    cidr_blocks = ["0.0.0.0/0"]
+  }
+
+  tags = {
+    Name  = "efa-test-sg-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Private Subnets
+resource "aws_subnet" "efa_test_private_subnet" {
+  count = 2
+
+  vpc_id            = aws_vpc.efa_test_vpc.id
+  cidr_block        = "10.0.${count.index + 10}.0/24"
+  availability_zone = data.aws_availability_zones.available.names[count.index]
+
+  tags = {
+    Name  = "efa-test-private-subnet-${count.index + 1}-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Elastic IP for NAT Gateway
+resource "aws_eip" "efa_test_nat_eip" {
+  domain = "vpc"
+
+  tags = {
+    Name  = "efa-test-nat-eip-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# NAT Gateway
+resource "aws_nat_gateway" "efa_test_nat" {
+  allocation_id = aws_eip.efa_test_nat_eip.id
+  subnet_id     = aws_subnet.efa_test_public_subnet[0].id
+
+  tags = {
+    Name  = "efa-test-nat-${module.common.testing_id}"
+    Owner = "default"
+  }
+
+  depends_on = [aws_internet_gateway.efa_test_igw]
+}
+
+# Route Table for Private Subnets
+resource "aws_route_table" "efa_test_private_rt" {
+  vpc_id = aws_vpc.efa_test_vpc.id
+
+  route {
+    cidr_block     = "0.0.0.0/0"
+    nat_gateway_id = aws_nat_gateway.efa_test_nat.id
+  }
+
+  tags = {
+    Name  = "efa-test-private-rt-${module.common.testing_id}"
+    Owner = "default"
+  }
+}
+
+# Associate Route Table with Private Subnets
+resource "aws_route_table_association" "efa_test_private_rta" {
+  count = length(aws_subnet.efa_test_private_subnet)
+
+  subnet_id      = aws_subnet.efa_test_private_subnet[count.index].id
+  route_table_id = aws_route_table.efa_test_private_rt.id
+}
+
+# Data source for availability zones
+data "aws_availability_zones" "available" {
+  state = "available"
+}
diff --git a/terraform/eks/daemon/otel-neuron-dra/main.tf b/terraform/eks/daemon/otel-neuron-dra/main.tf
new file mode 100644
index 00000000..fe3c2325
--- /dev/null
+++ b/terraform/eks/daemon/otel-neuron-dra/main.tf
@@ -0,0 +1,419 @@
+# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: MIT
+
+# Neuron DRA-path integration test cluster. Mirrors otel-neuron but exposes Neuron
+# devices via the AWS Neuron DRA driver (DeviceClass neuron.aws.com) instead of the
+# device plugin, and the burn workload claims a device via a ResourceClaimTemplate.
+# One Trainium node (trn1.2xlarge = 1 device x 2 cores) proves per-device/per-core
+# DRA correlation: the claimed device's 2 cores attribute to the burn pod, and to
+# no other pod. Trainium (not Inferentia): the Neuron DRA driver image 1.2.0 supports
+# Trainium only and rejects inf1/inf2 at device discovery ("unsupported instance
+# type"); trn1.2xlarge is the smallest/most-available Trainium instance.
+
+module "common" {
+  source             = "../../../common"
+  cwagent_image_repo = var.cwagent_image_repo
+  cwagent_image_tag  = var.cwagent_image_tag
+}
+
+module "basic_components" {
+  source = "../../../basic_components"
+  region = var.region
+}
+
+locals {
+  aws_eks = "aws eks --region ${var.region}"
+}
+
+resource "aws_eks_cluster" "this" {
+  name     = "cwagent-eks-integ-${module.common.testing_id}"
+  role_arn = module.basic_components.role_arn
+  version  = var.k8s_version
+  vpc_config {
+    subnet_ids         = module.basic_components.public_subnet_ids
+    security_group_ids = [module.basic_components.security_group]
+  }
+}
+
+# --- IAM ---
+
+resource "aws_iam_role" "node_role" {
+  name = "cwagent-otel-neuron-dra-Worker-Role-${module.common.testing_id}"
+  assume_role_policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [{
+      Effect    = "Allow"
+      Principal = { Service = "ec2.amazonaws.com" }
+      Action    = "sts:AssumeRole"
+    }]
+  })
+}
+
+resource "aws_iam_role_policy_attachment" "node_AmazonEKSWorkerNodePolicy" {
+  policy_arn = "arn:aws:iam::aws:policy/AmazonEKSWorkerNodePolicy"
+  role       = aws_iam_role.node_role.name
+}
+
+resource "aws_iam_role_policy_attachment" "node_AmazonEKS_CNI_Policy" {
+  policy_arn = "arn:aws:iam::aws:policy/AmazonEKS_CNI_Policy"
+  role       = aws_iam_role.node_role.name
+}
+
+resource "aws_iam_role_policy_attachment" "node_AmazonEC2ContainerRegistryReadOnly" {
+  policy_arn = "arn:aws:iam::aws:policy/AmazonEC2ContainerRegistryReadOnly"
+  role       = aws_iam_role.node_role.name
+}
+
+resource "aws_iam_role" "pod_identity_role" {
+  name = "cwagent-otel-neuron-dra-pod-identity-${module.common.testing_id}"
+  assume_role_policy = jsonencode({
+    Version = "2012-10-17"
+    Statement = [{
+      Effect    = "Allow"
+      Principal = { Service = "pods.eks.amazonaws.com" }
+      Action    = ["sts:AssumeRole", "sts:TagSession"]
+    }]
+  })
+}
+
+resource "aws_iam_role_policy_attachment" "pod_identity_CloudWatchAgentServerPolicy" {
+  policy_arn = "arn:aws:iam::aws:policy/CloudWatchAgentServerPolicy"
+  role       = aws_iam_role.pod_identity_role.name
+}
+
+# --- Node Groups ---
+
+resource "aws_eks_node_group" "standard" {
+  cluster_name    = aws_eks_cluster.this.name
+  node_group_name = "standard-${module.common.testing_id}"
+  node_role_arn   = aws_iam_role.node_role.arn
+  subnet_ids      = module.basic_components.public_subnet_ids
+
+  scaling_config {
+    desired_size = 1
+    max_size     = 1
+    min_size     = 1
+  }
+
+  ami_type       = "AL2023_x86_64_STANDARD"
+  capacity_type  = "ON_DEMAND"
+  disk_size      = 20
+  instance_types = ["t3.medium"]
+
+  depends_on = [
+    aws_iam_role_policy_attachment.node_AmazonEC2ContainerRegistryReadOnly,
+    aws_iam_role_policy_attachment.node_AmazonEKS_CNI_Policy,
+    aws_iam_role_policy_attachment.node_AmazonEKSWorkerNodePolicy,
+  ]
+}
+
+# Neuron (Trainium) node (trn1.2xlarge). No taint: the node is dedicated by
+# being the only Neuron node, and the DRA driver DaemonSet + burn pod land here via
+# scheduling/nodeSelector. Avoiding a taint keeps the Neuron DRA driver DaemonSet
+# (whose default tolerations are not assumed) schedulable.
+# Pinned to usw2-az4: trn1.2xlarge is offered only in usw2-az1/usw2-az4 in us-west-2.
+data "aws_subnet" "az" {
+  for_each = toset(module.basic_components.public_subnet_ids)
+  id       = each.value
+}
+
+locals {
+  neuron_az_id   = "usw2-az4"
+  neuron_subnets = [for s in data.aws_subnet.az : s.id if s.availability_zone_id == local.neuron_az_id]
+}
+
+resource "aws_eks_node_group" "neuron_multi" {
+  cluster_name    = aws_eks_cluster.this.name
+  node_group_name = "neuron-multi-${module.common.testing_id}"
+  node_role_arn   = aws_iam_role.node_role.arn
+  subnet_ids      = local.neuron_subnets
+
+  scaling_config {
+    desired_size = 1
+    max_size     = 1
+    min_size     = 1
+  }
+
+  ami_type       = var.ami_type
+  capacity_type  = "ON_DEMAND"
+  disk_size      = 100
+  instance_types = [var.instance_type]
+
+  labels = {
+    "aws.amazon.com/neuron.present"  = "true"
+    "ci-test.example.com/node-color" = "purple"
+  }
+
+  depends_on = [
+    aws_iam_role_policy_attachment.node_AmazonEC2ContainerRegistryReadOnly,
+    aws_iam_role_policy_attachment.node_AmazonEKS_CNI_Policy,
+    aws_iam_role_policy_attachment.node_AmazonEKSWorkerNodePolicy,
+  ]
+}
+
+# --- EKS Addon: Pod Identity agent ---
+
+resource "aws_eks_addon" "pod_identity_agent" {
+  depends_on   = [aws_eks_node_group.standard]
+  cluster_name = aws_eks_cluster.this.name
+  addon_name   = "eks-pod-identity-agent"
+}
+
+# --- Update kubeconfig ---
+
+resource "null_resource" "kubectl" {
+  depends_on = [aws_eks_cluster.this, aws_eks_node_group.standard, aws_eks_node_group.neuron_multi]
+  provisioner "local-exec" {
+    command = "${local.aws_eks} update-kubeconfig --name ${aws_eks_cluster.this.name}"
+  }
+}
+
+# --- Neuron DRA driver (Helm) — replaces the device plugin ---
+# DeviceClass neuron.aws.com, driver neuron.aws.com. devicePlugin disabled so it
+# does not compete with the DRA driver for the same devices.
+
+resource "helm_release" "neuron_dra_driver" {
+  depends_on = [aws_eks_node_group.neuron_multi, null_resource.kubectl]
+
+  name       = "neuron-helm-chart"
+  repository = "oci://public.ecr.aws/neuron"
+  chart      = "neuron-helm-chart"
+  # The chart renders its own Namespace object for neuron-dra-driver (via
+  # draDriver.namespaceOverride, no disable toggle) AND places the driver
+  # DaemonSet/SA/RBAC there. So we install the release into the pre-existing
+  # kube-system namespace (Helm needs the release namespace to exist to store the
+  # release secret) with create_namespace=false, and let the chart create and own
+  # neuron-dra-driver. Installing into neuron-dra-driver directly deadlocks: with
+  # create_namespace=true the provider's namespace collides with the chart's
+  # ("already exists"); with false Helm fails "namespace not found".
+  namespace        = "kube-system"
+  create_namespace = false
+  version          = var.neuron_helm_chart_version != "" ? var.neuron_helm_chart_version : null
+
+  set = [
+    { name = "devicePlugin.enabled", value = "false" },
+    { name = "npd.enabled", value = "false" },
+    { name = "scheduler.enabled", value = "false" },
+    { name = "draDriver.enabled", value = "true" },
+  ]
+}
+
+# --- neuron-burn-dra Deployment: claims 1 whole Neuron device via DRA ---
+
+resource "null_resource" "neuron_burn_dra" {
+  depends_on = [helm_release.neuron_dra_driver, null_resource.kubectl]
+  # Re-apply the manifest (kubectl apply is idempotent) if the instance type
+  # changes, so the burn pod's node-type nodeSelector tracks var.instance_type.
+  triggers = {
+    instance_type = var.instance_type
+    cluster       = aws_eks_cluster.this.name
+  }
+  provisioner "local-exec" {
+    command = <<-EOT
+      cat <<'EOF' | kubectl apply -f -
+      apiVersion: resource.k8s.io/v1
+      kind: ResourceClaimTemplate
+      metadata:
+        name: neuron-1
+        namespace: default
+      spec:
+        spec:
+          devices:
+            requests:
+            - name: neuron
+              exactly:
+                deviceClassName: neuron.aws.com
+                count: 1
+      ---
+      apiVersion: apps/v1
+      kind: Deployment
+      metadata:
+        name: neuron-burn-dra
+        namespace: default
+      spec:
+        replicas: 1
+        revisionHistoryLimit: 2
+        progressDeadlineSeconds: 600
+        selector:
+          matchLabels:
+            app: neuron-burn-dra
+        template:
+          metadata:
+            labels:
+              app: neuron-burn-dra
+              neuron-test: "true"
+              ci-test.example.com/pod-color: violet
+          spec:
+            nodeSelector:
+              node.kubernetes.io/instance-type: ${var.instance_type}
+            resourceClaims:
+            - name: neuron
+              resourceClaimTemplateName: neuron-1
+            containers:
+            - name: neuron-burn
+              image: public.ecr.aws/neuron/pytorch-inference-neuronx:2.1.2-neuronx-py310-sdk2.20.2-ubuntu20.04
+              command: ["python3", "-c"]
+              args:
+              - |
+                import torch, torch_neuronx, time
+                print("Compiling neuron trace (this takes a minute)...")
+                x = torch.randn(256, 256)
+                model = torch.nn.Linear(256, 256, bias=False)
+                traced = torch_neuronx.trace(model, x)
+                print("Trace compiled. Starting burn loop...")
+                iteration = 0
+                while True:
+                    start = time.time()
+                    for _ in range(1000):
+                        _ = traced(x)
+                    elapsed = time.time() - start
+                    iteration += 1
+                    print(f"Iteration {iteration}: 1000 inferences in {elapsed:.2f}s")
+              resources:
+                claims:
+                - name: neuron
+                requests:
+                  cpu: "1"
+                  memory: 4Gi
+      EOF
+    EOT
+  }
+}
+
+# --- Helm chart install (observability) ---
+
+data "external" "clone_helm_chart" {
+  program = ["bash", "-c", <<-EOT
+    rm -rf ./helm-charts
+    git clone -b ${var.helm_chart_branch} ${var.helm_chart_repo_url} ./helm-charts
+    echo '{"status":"ready"}'
+  EOT
+  ]
+}
+
+resource "helm_release" "aws_observability" {
+  name             = "amazon-cloudwatch-observability"
+  chart            = "./helm-charts/charts/amazon-cloudwatch-observability"
+  namespace        = "amazon-cloudwatch"
+  create_namespace = true
+  wait             = false
+  timeout          = 600
+
+  set = [
+    { name = "clusterName", value = aws_eks_cluster.this.name },
+    { name = "region", value = var.region },
+    { name = "otelContainerInsights.enabled", value = "true" },
+  ]
+
+  depends_on = [
+    aws_eks_addon.pod_identity_agent,
+    null_resource.kubectl,
+    data.external.clone_helm_chart,
+  ]
+}
+
+# --- Pod Identity association ---
+
+resource "aws_eks_pod_identity_association" "cloudwatch_agent" {
+  depends_on      = [helm_release.aws_observability]
+  cluster_name    = aws_eks_cluster.this.name
+  namespace       = "amazon-cloudwatch"
+  service_account = "cloudwatch-agent"
+  role_arn        = aws_iam_role.pod_identity_role.arn
+}
+
+# --- Patch agent image ---
+
+resource "null_resource" "update_image" {
+  depends_on = [helm_release.aws_observability, null_resource.kubectl]
+  triggers   = { timestamp = timestamp() }
+  provisioner "local-exec" {
+    command = <<-EOT
+      sleep 30
+      kubectl -n amazon-cloudwatch patch AmazonCloudWatchAgent cloudwatch-agent --type='json' \
+        -p='[{"op": "replace", "path": "/spec/image", "value": "${var.cwagent_image_repo}:${var.cwagent_image_tag}"}]'
+      kubectl -n amazon-cloudwatch patch AmazonCloudWatchAgent cloudwatch-agent-cluster-scraper --type='json' \
+        -p='[{"op": "replace", "path": "/spec/image", "value": "${var.cwagent_image_repo}:${var.cwagent_image_tag}"}]' 2>/dev/null || true
+      sleep 10
+    EOT
+  }
+}
+
+# --- Restart pods ---
+
+resource "null_resource" "restart_pods" {
+  depends_on = [aws_eks_pod_identity_association.cloudwatch_agent, null_resource.update_image]
+  triggers   = { timestamp = timestamp() }
+  provisioner "local-exec" {
+    command = <<-EOT
+      kubectl -n amazon-cloudwatch rollout restart daemonset/cloudwatch-agent
+      kubectl -n amazon-cloudwatch rollout restart deployment/cloudwatch-agent-cluster-scraper 2>/dev/null || true
+      kubectl -n amazon-cloudwatch rollout status daemonset/cloudwatch-agent --timeout=120s
+    EOT
+  }
+}
+
+# --- Wait for neuron-monitor pods ---
+
+resource "null_resource" "wait_neuron_monitor" {
+  depends_on = [null_resource.restart_pods, null_resource.neuron_burn_dra]
+  triggers   = { timestamp = timestamp() }
+  provisioner "local-exec" {
+    command = <<-EOT
+      echo "Waiting for neuron-monitor pods to be ready..."
+      for i in $(seq 1 30); do
+        READY=$(kubectl -n amazon-cloudwatch get pods -l app.kubernetes.io/name=neuron-monitor --no-headers 2>/dev/null | grep -c "Running" || true)
+        if [ "$READY" -ge 1 ]; then
+          echo "neuron-monitor ready ($READY running)"
+          break
+        fi
+        echo "Attempt $i: $READY neuron-monitor pods running, waiting..."
+        sleep 10
+      done
+    EOT
+  }
+}
+
+# --- Test runner ---
+
+resource "null_resource" "validator" {
+  depends_on = [null_resource.wait_neuron_monitor]
+  triggers   = { always_run = timestamp() }
+  provisioner "local-exec" {
+    command = <<-EOT
+      echo "Running OTEL Neuron DRA cluster integration tests"
+      cd ../../../..
+
+      echo "Waiting for Neuron runtime to initialize (image pull + trace compile)..."
+      READY=0
+      for i in $(seq 1 90); do
+        READY=$(kubectl logs -n default -l app=neuron-burn-dra --tail=5 2>/dev/null | grep -c "^Iteration " || true)
+        if [ "$READY" -gt 0 ]; then
+          echo "Neuron burn loop active (after $((i*10))s)"
+          break
+        fi
+        if [ $((i % 6)) -eq 0 ]; then
+          echo "--- Attempt $i ($((i*10))s): iter_lines=$READY ---"
+          kubectl get pods -n default -l app=neuron-burn-dra -o wide 2>&1 | head || true
+        fi
+        sleep 10
+      done
+      if [ "$READY" -eq 0 ]; then
+        echo "ERROR: neuron-burn-dra loop not active after 15 minutes"
+        kubectl get pods -n default -l app=neuron-burn-dra -o wide 2>&1 || true
+        kubectl describe pod -n default -l app=neuron-burn-dra 2>&1 | tail -40 || true
+        kubectl get resourceclaims -n default 2>&1 || true
+        exit 1
+      fi
+
+      echo "Waiting 6 minutes for metrics to propagate (Zeus 5-min staleness window)..."
+      sleep 360
+
+      go test -tags integration -timeout 1h -v ${var.test_dir} \
+        -eksClusterName=${aws_eks_cluster.this.name} \
+        -computeType=EKS \
+        -eksDeploymentStrategy=DAEMON \
+        -region=${var.region}
+    EOT
+  }
+}
diff --git a/terraform/eks/daemon/otel-neuron-dra/providers.tf b/terraform/eks/daemon/otel-neuron-dra/providers.tf
new file mode 100644
index 00000000..a562a7f1
--- /dev/null
+++ b/terraform/eks/daemon/otel-neuron-dra/providers.tf
@@ -0,0 +1,28 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+provider "aws" {
+  region = var.region
+}
+
+provider "kubernetes" {
+  host                   = aws_eks_cluster.this.endpoint
+  cluster_ca_certificate = base64decode(aws_eks_cluster.this.certificate_authority.0.data)
+  exec {
+    api_version = "client.authentication.k8s.io/v1beta1"
+    command     = "aws"
+    args        = ["eks", "get-token", "--cluster-name", aws_eks_cluster.this.name]
+  }
+}
+
+provider "helm" {
+  kubernetes = {
+    host                   = aws_eks_cluster.this.endpoint
+    cluster_ca_certificate = base64decode(aws_eks_cluster.this.certificate_authority.0.data)
+    exec = {
+      api_version = "client.authentication.k8s.io/v1beta1"
+      command     = "aws"
+      args        = ["eks", "get-token", "--cluster-name", aws_eks_cluster.this.name]
+    }
+  }
+}
diff --git a/terraform/eks/daemon/otel-neuron-dra/variables.tf b/terraform/eks/daemon/otel-neuron-dra/variables.tf
new file mode 100644
index 00000000..bc7bdad1
--- /dev/null
+++ b/terraform/eks/daemon/otel-neuron-dra/variables.tf
@@ -0,0 +1,66 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+variable "region" {
+  type    = string
+  default = "us-west-2"
+}
+
+variable "test_dir" {
+  type    = string
+  default = "./test/otel/neuron_dra"
+}
+
+variable "cwagent_image_repo" {
+  type    = string
+  default = "public.ecr.aws/cloudwatch-agent/cloudwatch-agent"
+}
+
+variable "cwagent_image_tag" {
+  type    = string
+  default = "latest"
+}
+
+# The chart must render the DRA correlation config (dra_device_types incl. the
+# neuron-dra entry keyed on the neuron.aws.com driver) and grant the agent
+# ServiceAccount resource.k8s.io RBAC. Override to a fork/branch until that lands
+# on main (helm-charts PR #356).
+variable "helm_chart_branch" {
+  type    = string
+  default = "main"
+}
+
+variable "helm_chart_repo_url" {
+  type    = string
+  default = "https://github.com/aws-observability/helm-charts.git"
+}
+
+# The awsdevicepodcorrelation processor uses the GA resource.k8s.io/v1 DRA API
+# (available since Kubernetes 1.34), so the cluster must be 1.34+.
+variable "k8s_version" {
+  type    = string
+  default = "1.35"
+}
+
+variable "ami_type" {
+  type    = string
+  default = "AL2023_x86_64_NEURON"
+}
+
+# Neuron (Trainium) node: trn1.2xlarge = 1 device × 2 cores.
+# Must be a Trainium type: the Neuron DRA driver (driver image 1.2.0) supports
+# Trainium only and rejects Inferentia (inf1/inf2) at device discovery. trn1.2xlarge
+# is the smallest/most-available Trainium instance, enough to exercise DRA
+# per-device correlation (1 claimed device -> both cores -> burn pod).
+variable "instance_type" {
+  type    = string
+  default = "trn1.2xlarge"
+}
+
+# Neuron DRA driver Helm chart (OCI). draDriver.enabled=true installs the DRA
+# driver (DeviceClass neuron.aws.com); devicePlugin must be disabled (the two
+# cannot coexist on a node).
+variable "neuron_helm_chart_version" {
+  type    = string
+  default = ""
+}
diff --git a/test/otel/gpu_dra/gpu_dra_test.go b/test/otel/gpu_dra/gpu_dra_test.go
new file mode 100644
index 00000000..5fd88859
--- /dev/null
+++ b/test/otel/gpu_dra/gpu_dra_test.go
@@ -0,0 +1,170 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+// Package gpu_dra validates per-device GPU pod correlation on the Dynamic Resource
+// Allocation (DRA) path — the counterpart of the device-plugin `gpu` package's
+// multi_gpu tests. GPUs are allocated via the NVIDIA DRA driver (DeviceClass
+// gpu.nvidia.com, driver gpu.nvidia.com) through a ResourceClaimTemplate instead of
+// the nvidia.com/gpu device-plugin resource. The emitted DCGM metrics are identical,
+// so the assertions mirror the device-plugin multi_gpu_test.
+//
+// Cluster topology:
+//   - 1x g4dn.12xlarge = 4 T4 GPUs
+//   - GPUs exposed via the NVIDIA DRA driver, not the device plugin
+//   - A burn workload (gpu-burn-dra) claims 1 GPU via DRA
+package gpu_dra
+
+import (
+	"context"
+	"fmt"
+	"strconv"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+const (
+	multiGpuInstanceType  = "g4dn.12xlarge"
+	expectedMultiGPUCount = 4
+
+	// burnPodPrefix is the DRA burn workload that claims 1 GPU via a ResourceClaimTemplate.
+	burnPodPrefix        = "gpu-burn-dra"
+	expectedClaimedGPUs  = 1
+	expectedUncorrelated = expectedMultiGPUCount - expectedClaimedGPUs
+)
+
+// TestGPUDRADeviceCount validates the node exposes the expected number of GPUs.
+func TestGPUDRADeviceCount(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "DCGM_FI_DEV_GPU_UTIL")
+	require.NoError(t, err, "querying DCGM_FI_DEV_GPU_UTIL")
+	require.NotEmpty(t, results, "DCGM_FI_DEV_GPU_UTIL not available")
+
+	multi := filterByHostType(results, multiGpuInstanceType)
+	require.True(t, len(multi) > 0, "No DCGM_FI_DEV_GPU_UTIL results from %s node", multiGpuInstanceType)
+
+	gpus := uniqueDatapointValues(multi, "gpu")
+	require.Equal(t, expectedMultiGPUCount, len(gpus),
+		"Expected %d distinct gpu on %s, got %d: %v", expectedMultiGPUCount, multiGpuInstanceType, len(gpus), gpus)
+}
+
+// TestGPUDRAConsecutiveIndices validates GPU indices are 0..N-1.
+func TestGPUDRAConsecutiveIndices(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "DCGM_FI_DEV_GPU_UTIL")
+	require.NoError(t, err, "querying DCGM_FI_DEV_GPU_UTIL")
+	multi := filterByHostType(results, multiGpuInstanceType)
+	require.True(t, len(multi) > 0, "No results from %s node", multiGpuInstanceType)
+
+	gpus := uniqueDatapointValues(multi, "gpu")
+	require.Equal(t, expectedMultiGPUCount, len(gpus), "Expected %d GPU indices, got %d: %v", expectedMultiGPUCount, len(gpus), gpus)
+	for i := 0; i < expectedMultiGPUCount; i++ {
+		expected := strconv.Itoa(i)
+		found := false
+		for _, g := range gpus {
+			if g == expected {
+				found = true
+				break
+			}
+		}
+		require.True(t, found, "Expected GPU index %d, got indices: %v", i, gpus)
+	}
+}
+
+// TestGPUDRAAllMetricsPerDevice validates every DCGM metric reports for all 4 GPUs.
+func TestGPUDRAAllMetricsPerDevice(t *testing.T) {
+	t.Parallel()
+	for _, metricName := range dcgmMetricNamesList {
+		metricName := metricName
+		t.Run(metricName, func(t *testing.T) {
+			t.Parallel()
+			results, err := queryCache.Get(context.Background(), metricName)
+			require.NoError(t, err, "querying %s", metricName)
+			multi := filterByHostType(results, multiGpuInstanceType)
+			require.True(t, len(multi) > 0, "No %s results from %s node", metricName, multiGpuInstanceType)
+
+			gpus := uniqueDatapointValues(multi, "gpu")
+			require.Equal(t, expectedMultiGPUCount, len(gpus),
+				"%s: expected %d GPUs, got %d: %v", metricName, expectedMultiGPUCount, len(gpus), gpus)
+		})
+	}
+}
+
+// TestGPUDRAClaimedVsUnclaimedCorrelation is the DRA-path per-device correlation
+// guard. On a 4-GPU node, the burn pod claims exactly 1 GPU via DRA, so exactly 1
+// GPU correlates to that pod and the remaining 3 carry NO pod. This fails if DRA
+// correlation collapses GPUs onto one pod, over-correlates unclaimed GPUs, or maps
+// a GPU to the wrong pod.
+func TestGPUDRAClaimedVsUnclaimedCorrelation(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "DCGM_FI_DEV_GPU_UTIL")
+	require.NoError(t, err, "querying DCGM_FI_DEV_GPU_UTIL")
+	multi := filterByHostType(results, multiGpuInstanceType)
+	require.NotEmpty(t, multi, "No DCGM_FI_DEV_GPU_UTIL results from %s node", multiGpuInstanceType)
+
+	// For each GPU, collect the distinct pods it is attributed to (empty = unclaimed).
+	gpuPods := make(map[string]map[string]struct{})
+	for _, r := range multi {
+		r := r
+		gpu := r.Labels.Datapoint["gpu"]
+		if gpu == "" {
+			continue
+		}
+		if gpuPods[gpu] == nil {
+			gpuPods[gpu] = make(map[string]struct{})
+		}
+		if pod := r.Labels.Resource["k8s.pod.name"]; pod != "" {
+			gpuPods[gpu][pod] = struct{}{}
+		}
+	}
+	require.Len(t, gpuPods, expectedMultiGPUCount,
+		"expected %d GPUs on the node, got %d", expectedMultiGPUCount, len(gpuPods))
+
+	var claimed, unclaimed []string
+	for gpu, pods := range gpuPods {
+		switch len(pods) {
+		case 0:
+			unclaimed = append(unclaimed, gpu)
+		case 1:
+			var pod string
+			for p := range pods {
+				pod = p
+			}
+			require.True(t, strings.HasPrefix(pod, burnPodPrefix),
+				"GPU %s correlated to unexpected pod %q (expected %s*)", gpu, pod, burnPodPrefix)
+			claimed = append(claimed, gpu)
+		default:
+			t.Errorf("GPU %s correlated to multiple pods", gpu)
+		}
+	}
+
+	require.Len(t, claimed, expectedClaimedGPUs,
+		"expected %d claimed GPU correlated to %s*, got %d: %v "+
+			"(collapse over-correlates unclaimed GPUs onto a pod)",
+		expectedClaimedGPUs, burnPodPrefix, len(claimed), claimed)
+	require.Len(t, unclaimed, expectedUncorrelated,
+		"expected %d uncorrelated GPUs with no pod, got %d: %v",
+		expectedUncorrelated, len(unclaimed), unclaimed)
+}
+
+// TestGPUDRABurnPodLabels validates the correlated GPU's pod labels via a targeted
+// PromQL query (avoids stale series in the shared OTLP store).
+func TestGPUDRABurnPodLabels(t *testing.T) {
+	t.Parallel()
+	ctx := context.Background()
+	promql := fmt.Sprintf(
+		`DCGM_FI_DEV_GPU_UTIL{"@resource.k8s.cluster.name"="%s","@resource.k8s.pod.name"=~"%s.*"}`,
+		escapePromQL(cfg.ClusterName), burnPodPrefix)
+	burn, err := client.Query(ctx, promql)
+	require.NoError(t, err, "querying DCGM_FI_DEV_GPU_UTIL for %s", burnPodPrefix)
+	require.Equal(t, expectedClaimedGPUs, len(burn),
+		"Expected %d GPU correlated to %s, got %d", expectedClaimedGPUs, burnPodPrefix, len(burn))
+
+	for _, r := range burn {
+		r := r
+		require.Equal(t, "default", r.Labels.Resource["k8s.namespace.name"], "%s namespace", burnPodPrefix)
+	}
+}
diff --git a/test/otel/gpu_dra/helpers_test.go b/test/otel/gpu_dra/helpers_test.go
new file mode 100644
index 00000000..80774242
--- /dev/null
+++ b/test/otel/gpu_dra/helpers_test.go
@@ -0,0 +1,46 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+package gpu_dra
+
+import (
+	"sort"
+	"strings"
+
+	"github.com/aws/amazon-cloudwatch-agent-test/util/otelmetrics"
+)
+
+// escapePromQL escapes a string value for safe use inside PromQL label matches.
+func escapePromQL(s string) string {
+	return strings.NewReplacer(`\`, `\\`, `"`, `\"`).Replace(s)
+}
+
+// filterByHostType returns only results whose resource host.type matches.
+func filterByHostType(results []otelmetrics.MetricResult, instanceType string) []otelmetrics.MetricResult {
+	var out []otelmetrics.MetricResult
+	for _, r := range results {
+		if r.Labels.Resource["host.type"] == instanceType {
+			out = append(out, r)
+		}
+	}
+	return out
+}
+
+// uniqueDatapointValues returns the sorted unique non-empty values of a
+// datapoint-level attribute across all results.
+func uniqueDatapointValues(results []otelmetrics.MetricResult, attr string) []string {
+	seen := make(map[string]struct{})
+	for _, r := range results {
+		if v, ok := r.Labels.Datapoint[attr]; ok && v != "" {
+			seen[v] = struct{}{}
+		}
+	}
+	out := make([]string, 0, len(seen))
+	for v := range seen {
+		out = append(out, v)
+	}
+	sort.Strings(out)
+	return out
+}
diff --git a/test/otel/gpu_dra/metrics_test.go b/test/otel/gpu_dra/metrics_test.go
new file mode 100644
index 00000000..24e6de09
--- /dev/null
+++ b/test/otel/gpu_dra/metrics_test.go
@@ -0,0 +1,31 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+package gpu_dra
+
+import "github.com/aws/amazon-cloudwatch-agent-test/util/otelmetrics"
+
+// The DRA cluster has a single multi-GPU node (g4dn.12xlarge, 4 GPUs).
+var clusterHostTypes = []string{multiGpuInstanceType}
+
+// DCGM metric definitions (same source names as the device-plugin gpu package).
+var dcgmMetrics = []otelmetrics.MetricDefinition{
+	{Name: "DCGM_FI_DEV_GPU_UTIL", MetricType: "gauge", Scope: otelmetrics.ScopePod, Unit: "%"},
+	{Name: "DCGM_FI_DEV_MEM_COPY_UTIL", MetricType: "gauge", Scope: otelmetrics.ScopePod, Unit: "%"},
+	{Name: "DCGM_FI_DEV_FB_USED", MetricType: "gauge", Scope: otelmetrics.ScopePod, Unit: "MiBy"},
+	{Name: "DCGM_FI_DEV_GPU_TEMP", MetricType: "gauge", Scope: otelmetrics.ScopePod, Unit: "Cel"},
+	{Name: "DCGM_FI_DEV_POWER_USAGE", MetricType: "gauge", Scope: otelmetrics.ScopePod, Unit: "W"},
+	{Name: "DCGM_FI_DEV_FB_FREE", MetricType: "gauge", Scope: otelmetrics.ScopePod, Unit: "MiBy"},
+}
+
+func metricNames(defs []otelmetrics.MetricDefinition) []string {
+	names := make([]string, len(defs))
+	for i, d := range defs {
+		names[i] = d.Name
+	}
+	return names
+}
+
+var dcgmMetricNamesList = metricNames(dcgmMetrics)
diff --git a/test/otel/gpu_dra/setup_test.go b/test/otel/gpu_dra/setup_test.go
new file mode 100644
index 00000000..9aeb334f
--- /dev/null
+++ b/test/otel/gpu_dra/setup_test.go
@@ -0,0 +1,95 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+package gpu_dra
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"os"
+	"testing"
+	"time"
+
+	awsconfig "github.com/aws/aws-sdk-go-v2/config"
+	"github.com/aws/aws-sdk-go-v2/service/sts"
+
+	"github.com/aws/amazon-cloudwatch-agent-test/environment"
+	"github.com/aws/amazon-cloudwatch-agent-test/util/otelmetrics"
+)
+
+var (
+	cfg        otelmetrics.TestConfig
+	client     *otelmetrics.OtelMetricsClient
+	queryCache *otelmetrics.QueryCache
+)
+
+func TestMain(m *testing.M) {
+	environment.RegisterEnvironmentMetaDataFlags()
+	flag.Parse()
+	env := environment.GetEnvironmentMetaData()
+
+	region := env.Region
+	if region == "" {
+		region = os.Getenv("AWS_REGION")
+	}
+	if region == "" {
+		fmt.Fprintf(os.Stderr, "Region not set\n")
+		os.Exit(1)
+	}
+
+	clusterName := env.EKSClusterName
+	if clusterName == "" {
+		clusterName = os.Getenv("CLUSTER_NAME")
+	}
+	if clusterName == "" {
+		fmt.Fprintf(os.Stderr, "Cluster name not set\n")
+		os.Exit(1)
+	}
+
+	ctx := context.Background()
+	awsCfg, err := awsconfig.LoadDefaultConfig(ctx, awsconfig.WithRegion(region))
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "AWS config error: %v\n", err)
+		os.Exit(1)
+	}
+	stsClient := sts.NewFromConfig(awsCfg)
+	identity, err := stsClient.GetCallerIdentity(ctx, &sts.GetCallerIdentityInput{})
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "STS GetCallerIdentity error: %v\n", err)
+		os.Exit(1)
+	}
+
+	cfg = otelmetrics.TestConfig{
+		Region:         region,
+		Endpoint:       fmt.Sprintf("https://monitoring.%s.amazonaws.com", region),
+		Timeout:        30 * time.Second,
+		MaxRetries:     3,
+		ClusterName:    clusterName,
+		AccountID:      *identity.Account,
+		SigningService: "monitoring",
+	}
+
+	client, err = otelmetrics.NewClient(ctx, cfg)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Client error: %v\n", err)
+		os.Exit(1)
+	}
+
+	hostMappings := []otelmetrics.SourceHostMapping{
+		{Source: otelmetrics.SourceDCGM, HostTypes: clusterHostTypes},
+	}
+
+	registry := otelmetrics.NewSourceRegistry(clusterHostTypes, hostMappings,
+		otelmetrics.SourceMapping{Source: otelmetrics.SourceDCGM, Metrics: dcgmMetrics},
+	)
+
+	queryCache = otelmetrics.NewQueryCache(client, cfg.ClusterName,
+		otelmetrics.WithHostTypes(clusterHostTypes),
+		otelmetrics.WithSourceRegistry(registry),
+	)
+
+	os.Exit(m.Run())
+}
diff --git a/test/otel/multi_efa_dra/helpers_test.go b/test/otel/multi_efa_dra/helpers_test.go
new file mode 100644
index 00000000..15748188
--- /dev/null
+++ b/test/otel/multi_efa_dra/helpers_test.go
@@ -0,0 +1,71 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+package multi_efa_dra
+
+import (
+	"sort"
+
+	"github.com/aws/amazon-cloudwatch-agent-test/util/otelmetrics"
+)
+
+// getAnyValue returns the attribute value from either resource or datapoint scope,
+// preferring resource scope if set in both.
+func getAnyValue(r otelmetrics.MetricResult, attr string) string {
+	if v, ok := r.Labels.Resource[attr]; ok && v != "" {
+		return v
+	}
+	return r.Labels.Datapoint[attr]
+}
+
+// filterByNodeLabel returns results where the given resource attribute equals the given value.
+func filterByNodeLabel(results []otelmetrics.MetricResult, labelKey, labelValue string) []otelmetrics.MetricResult {
+	var out []otelmetrics.MetricResult
+	for _, r := range results {
+		if r.Labels.Resource[labelKey] == labelValue {
+			out = append(out, r)
+		}
+	}
+	return out
+}
+
+// deviceKeys returns the sorted device names of a device->pods map (for messages).
+func deviceKeys(m map[string]map[string]struct{}) []string {
+	out := make([]string, 0, len(m))
+	for k := range m {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}
+
+// setKeys returns the sorted keys of a string set (for messages).
+func setKeys(m map[string]struct{}) []string {
+	out := make([]string, 0, len(m))
+	for k := range m {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}
+
+// uniqueAnyValues returns the sorted unique non-empty values of an attribute
+// found in either resource or datapoint scope across all results.
+func uniqueAnyValues(results []otelmetrics.MetricResult, attr string) []string {
+	seen := make(map[string]struct{})
+	for _, r := range results {
+		if v, ok := r.Labels.Resource[attr]; ok && v != "" {
+			seen[v] = struct{}{}
+		} else if v, ok := r.Labels.Datapoint[attr]; ok && v != "" {
+			seen[v] = struct{}{}
+		}
+	}
+	out := make([]string, 0, len(seen))
+	for v := range seen {
+		out = append(out, v)
+	}
+	sort.Strings(out)
+	return out
+}
diff --git a/test/otel/multi_efa_dra/metrics_test.go b/test/otel/multi_efa_dra/metrics_test.go
new file mode 100644
index 00000000..7aa4afeb
--- /dev/null
+++ b/test/otel/multi_efa_dra/metrics_test.go
@@ -0,0 +1,25 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+package multi_efa_dra
+
+import "github.com/aws/amazon-cloudwatch-agent-test/util/otelmetrics"
+
+var efaMetrics = []otelmetrics.MetricDefinition{
+	{Name: "efa_rx_bytes", MetricType: "counter", Scope: otelmetrics.ScopePod, Unit: "By"},
+	{Name: "efa_tx_bytes", MetricType: "counter", Scope: otelmetrics.ScopePod, Unit: "By"},
+	{Name: "efa_rx_dropped", MetricType: "counter", Scope: otelmetrics.ScopePod},
+	{Name: "efa_rdma_read_bytes", MetricType: "counter", Scope: otelmetrics.ScopePod, Unit: "By"},
+}
+
+var efaMetricNamesList = metricNames(efaMetrics)
+
+func metricNames(defs []otelmetrics.MetricDefinition) []string {
+	names := make([]string, len(defs))
+	for i, d := range defs {
+		names[i] = d.Name
+	}
+	return names
+}
diff --git a/test/otel/multi_efa_dra/multi_efa_dra_test.go b/test/otel/multi_efa_dra/multi_efa_dra_test.go
new file mode 100644
index 00000000..f24027f3
--- /dev/null
+++ b/test/otel/multi_efa_dra/multi_efa_dra_test.go
@@ -0,0 +1,233 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+// Package multi_efa_dra validates per-device EFA metric correlation on the
+// Dynamic Resource Allocation (DRA) path — the same behavior the multi_efa
+// package validates for the EFA device-plugin path, but with EFA devices
+// allocated via DRA (the dranet driver, driver name "dra.net") through a
+// ResourceClaimTemplate instead of the vpc.amazonaws.com/efa device-plugin
+// resource.
+//
+// This exercises the awsdevicepodcorrelation processor's DRA path end to end:
+// the processor watches Pods/ResourceClaims/ResourceSlices via the K8s API and
+// bridges the DRA device identity (a PCI name) to the EFA metric label
+// (rdmapXsY) via the ResourceSlice attribute dra.net/rdmaDevice. The emitted
+// metrics are identical to the device-plugin path, so the assertions mirror the
+// multi_efa package.
+//
+// Cluster topology:
+//   - 1x c6in.32xlarge with 2 EFA interfaces (1 per network card)
+//   - EFA exposed via DRA (dranet), not the EFA device plugin
+//   - Node label: ci-test.example.com/multi-efa-dra=true
+package multi_efa_dra
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+const (
+	expectedMultiEFACount = 2
+	multiEfaDraNodeLabel  = "k8s.node.label.ci-test.example.com/multi-efa-dra"
+)
+
+func TestMultiEFADeviceCount(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "efa_rx_bytes")
+	require.NoError(t, err, "querying efa_rx_bytes")
+	require.NotEmpty(t, results, "efa_rx_bytes not available")
+
+	multi := filterByNodeLabel(results, multiEfaDraNodeLabel, "true")
+	require.True(t, len(multi) > 0,
+		"No efa_rx_bytes results from multi-EFA DRA node (label %s)", multiEfaDraNodeLabel)
+
+	devices := uniqueAnyValues(multi, "aws.efa.device")
+	require.Equal(t, expectedMultiEFACount, len(devices),
+		"Expected %d distinct aws.efa.device, got %d: %v",
+		expectedMultiEFACount, len(devices), devices)
+}
+
+func TestMultiEFAUniqueENIs(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "efa_rx_bytes")
+	require.NoError(t, err, "querying efa_rx_bytes")
+	require.NotEmpty(t, results, "efa_rx_bytes not available")
+
+	multi := filterByNodeLabel(results, multiEfaDraNodeLabel, "true")
+	require.True(t, len(multi) > 0, "No results from multi-EFA DRA node")
+
+	enis := uniqueAnyValues(multi, "aws.efa.eni.id")
+	devices := uniqueAnyValues(multi, "aws.efa.device")
+	require.Equal(t, len(devices), len(enis),
+		"EFA device count (%d) != ENI count (%d) — duplicate ENIs?",
+		len(devices), len(enis))
+	for _, eni := range enis {
+		require.True(t, strings.HasPrefix(eni, "eni-"),
+			"aws.efa.eni.id should start with 'eni-', got '%s'", eni)
+	}
+}
+
+func TestMultiEFADeviceNamesAreRDMA(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "efa_rx_bytes")
+	require.NoError(t, err, "querying efa_rx_bytes")
+	require.NotEmpty(t, results, "efa_rx_bytes not available")
+
+	multi := filterByNodeLabel(results, multiEfaDraNodeLabel, "true")
+	require.True(t, len(multi) > 0, "No results from multi-EFA DRA node")
+
+	for _, r := range multi {
+		dev := getAnyValue(r, "aws.efa.device")
+		require.NotEmpty(t, dev, "aws.efa.device is empty")
+	}
+}
+
+func TestMultiEFAAllMetricsPerDevice(t *testing.T) {
+	t.Parallel()
+	for _, metricName := range efaMetricNamesList {
+		metricName := metricName
+		t.Run(metricName, func(t *testing.T) {
+			t.Parallel()
+			results, err := queryCache.Get(context.Background(), metricName)
+			require.NoError(t, err, "querying %s", metricName)
+			multi := filterByNodeLabel(results, multiEfaDraNodeLabel, "true")
+			require.True(t, len(multi) > 0, "No %s results from multi-EFA DRA node", metricName)
+
+			devices := uniqueAnyValues(multi, "aws.efa.device")
+			require.Equal(t, expectedMultiEFACount, len(devices),
+				"%s: expected %d EFA devices, got %d: %v",
+				metricName, expectedMultiEFACount, len(devices), devices)
+		})
+	}
+}
+
+func TestMultiEFAPortPresent(t *testing.T) {
+	t.Parallel()
+	for _, metricName := range efaMetricNamesList {
+		metricName := metricName
+		t.Run(metricName, func(t *testing.T) {
+			t.Parallel()
+			results, err := queryCache.Get(context.Background(), metricName)
+			require.NoError(t, err, "querying %s", metricName)
+			multi := filterByNodeLabel(results, multiEfaDraNodeLabel, "true")
+			require.True(t, len(multi) > 0, "No %s results from multi-EFA DRA node", metricName)
+
+			for _, r := range multi {
+				port := getAnyValue(r, "aws.efa.port")
+				require.NotEmpty(t, port,
+					"%s missing aws.efa.port (device: %s)",
+					metricName, getAnyValue(r, "aws.efa.device"))
+			}
+		})
+	}
+}
+
+func TestMultiEFACorrelatedCount(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "efa_rx_bytes")
+	require.NoError(t, err, "querying efa_rx_bytes")
+	multi := filterByNodeLabel(results, multiEfaDraNodeLabel, "true")
+	require.True(t, len(multi) > 0, "No results from multi-EFA DRA node")
+
+	var correlated int
+	for _, r := range multi {
+		if strings.HasPrefix(r.Labels.Resource["k8s.pod.name"], "efaburn") {
+			correlated++
+		}
+	}
+	require.True(t, correlated >= 1,
+		"Expected at least 1 EFA correlated to efaburn, got %d", correlated)
+}
+
+// TestMultiEFACorrelatedPodLabels validates that the correlated EFA result
+// has the expected pod labels.
+func TestMultiEFACorrelatedPodLabels(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "efa_rx_bytes")
+	require.NoError(t, err, "querying efa_rx_bytes")
+	multi := filterByNodeLabel(results, multiEfaDraNodeLabel, "true")
+
+	for _, r := range multi {
+		if strings.HasPrefix(r.Labels.Resource["k8s.pod.name"], "efaburn") {
+			require.Equal(t, "efaburn", r.Labels.Resource["k8s.container.name"], "efaburn container name")
+			require.Equal(t, "default", r.Labels.Resource["k8s.namespace.name"], "efaburn namespace")
+			return
+		}
+	}
+	t.Fatal("No efa_rx_bytes result correlated to efaburn pod")
+}
+
+// expectedClaimedEFACount is how many of the node's EFA devices are claimed by a
+// pod. efaburn (replicas: 1) requests 1 EFA via a ResourceClaimTemplate, so
+// exactly one device is claimed and the remaining device(s) must stay unclaimed.
+const expectedClaimedEFACount = 1
+
+// TestMultiEFAClaimedVsUnclaimedCorrelation validates per-device pod correlation
+// on a multi-EFA node whose devices are allocated via DRA: the device claimed by
+// efaburn is correlated to that pod, and every remaining (unclaimed) device
+// carries NO pod attributes.
+//
+// This is the DRA-path counterpart of the device-plugin regression guard. It
+// exercises the processor's DRA correlation (ResourceClaim/ResourceSlice keying
+// via dra.net/rdmaDevice) together with the groupbyattrs/efa split before the
+// resource-level promote. Without correct per-device correlation, ALL of the
+// node's EFA devices — including unclaimed ones — collapse onto a single pod;
+// this test fails in that case because (a) more than expectedClaimedEFACount
+// devices carry a pod, and (b) no device is left unclaimed. It also catches a
+// single device attributed to multiple pods.
+func TestMultiEFAClaimedVsUnclaimedCorrelation(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "efa_rx_bytes")
+	require.NoError(t, err, "querying efa_rx_bytes")
+	multi := filterByNodeLabel(results, multiEfaDraNodeLabel, "true")
+	require.NotEmpty(t, multi,
+		"No efa_rx_bytes results from multi-EFA DRA node (label %s)", multiEfaDraNodeLabel)
+
+	// For each device, collect the distinct pods it is attributed to (empty = unclaimed).
+	devicePods := make(map[string]map[string]struct{})
+	for _, r := range multi {
+		dev := getAnyValue(r, "aws.efa.device")
+		if dev == "" {
+			continue
+		}
+		if devicePods[dev] == nil {
+			devicePods[dev] = make(map[string]struct{})
+		}
+		if pod := r.Labels.Resource["k8s.pod.name"]; pod != "" {
+			devicePods[dev][pod] = struct{}{}
+		}
+	}
+	require.Len(t, devicePods, expectedMultiEFACount,
+		"expected %d EFA devices on the node, got %d: %v",
+		expectedMultiEFACount, len(devicePods), deviceKeys(devicePods))
+
+	var claimed, unclaimed []string
+	for dev, pods := range devicePods {
+		switch len(pods) {
+		case 0:
+			unclaimed = append(unclaimed, dev)
+		case 1:
+			claimed = append(claimed, dev)
+		default:
+			// A single device correlated to multiple pods is itself a collapse symptom.
+			t.Errorf("EFA device %s correlated to multiple pods %v", dev, setKeys(pods))
+		}
+	}
+
+	// Claimed side: exactly the number of EFAs efaburn requested map to a pod.
+	require.Len(t, claimed, expectedClaimedEFACount,
+		"expected %d correlated (claimed) EFA device(s), got %d: %v "+
+			"(collapse over-correlates unclaimed devices onto a pod)",
+		expectedClaimedEFACount, len(claimed), claimed)
+
+	// Unclaimed side: the remaining devices must carry no pod — this is the
+	// coverage that distinguishes correct correlation from the collapse.
+	require.Len(t, unclaimed, expectedMultiEFACount-expectedClaimedEFACount,
+		"expected %d unclaimed EFA device(s) with no pod, got %d: %v",
+		expectedMultiEFACount-expectedClaimedEFACount, len(unclaimed), unclaimed)
+}
diff --git a/test/otel/multi_efa_dra/setup_test.go b/test/otel/multi_efa_dra/setup_test.go
new file mode 100644
index 00000000..b522836d
--- /dev/null
+++ b/test/otel/multi_efa_dra/setup_test.go
@@ -0,0 +1,98 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+package multi_efa_dra
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"os"
+	"testing"
+	"time"
+
+	awsconfig "github.com/aws/aws-sdk-go-v2/config"
+	"github.com/aws/aws-sdk-go-v2/service/sts"
+
+	"github.com/aws/amazon-cloudwatch-agent-test/environment"
+	"github.com/aws/amazon-cloudwatch-agent-test/util/otelmetrics"
+)
+
+var (
+	cfg        otelmetrics.TestConfig
+	client     *otelmetrics.OtelMetricsClient
+	queryCache *otelmetrics.QueryCache
+)
+
+// Instance types in this cluster.
+var clusterHostTypes = []string{"c6in.32xlarge"}
+
+func TestMain(m *testing.M) {
+	environment.RegisterEnvironmentMetaDataFlags()
+	flag.Parse()
+	env := environment.GetEnvironmentMetaData()
+
+	region := env.Region
+	if region == "" {
+		region = os.Getenv("AWS_REGION")
+	}
+	if region == "" {
+		fmt.Fprintf(os.Stderr, "Region not set\n")
+		os.Exit(1)
+	}
+
+	clusterName := env.EKSClusterName
+	if clusterName == "" {
+		clusterName = os.Getenv("CLUSTER_NAME")
+	}
+	if clusterName == "" {
+		fmt.Fprintf(os.Stderr, "Cluster name not set\n")
+		os.Exit(1)
+	}
+
+	ctx := context.Background()
+	awsCfg, err := awsconfig.LoadDefaultConfig(ctx, awsconfig.WithRegion(region))
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "AWS config error: %v\n", err)
+		os.Exit(1)
+	}
+	stsClient := sts.NewFromConfig(awsCfg)
+	identity, err := stsClient.GetCallerIdentity(ctx, &sts.GetCallerIdentityInput{})
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "STS GetCallerIdentity error: %v\n", err)
+		os.Exit(1)
+	}
+
+	cfg = otelmetrics.TestConfig{
+		Region:         region,
+		Endpoint:       fmt.Sprintf("https://monitoring.%s.amazonaws.com", region),
+		Timeout:        30 * time.Second,
+		MaxRetries:     3,
+		ClusterName:    clusterName,
+		AccountID:      *identity.Account,
+		SigningService: "monitoring",
+	}
+
+	client, err = otelmetrics.NewClient(ctx, cfg)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Client error: %v\n", err)
+		os.Exit(1)
+	}
+
+	hostMappings := []otelmetrics.SourceHostMapping{
+		{Source: otelmetrics.SourceEFA, HostTypes: clusterHostTypes},
+	}
+
+	registry := otelmetrics.NewSourceRegistry(clusterHostTypes, hostMappings,
+		otelmetrics.SourceMapping{Source: otelmetrics.SourceEFA, Metrics: efaMetrics},
+	)
+
+	queryCache = otelmetrics.NewQueryCache(client, cfg.ClusterName,
+		otelmetrics.WithHostTypes(clusterHostTypes),
+		otelmetrics.WithSourceRegistry(registry),
+	)
+
+	os.Exit(m.Run())
+}
diff --git a/test/otel/neuron_dra/helpers_test.go b/test/otel/neuron_dra/helpers_test.go
new file mode 100644
index 00000000..506781d0
--- /dev/null
+++ b/test/otel/neuron_dra/helpers_test.go
@@ -0,0 +1,87 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+package neuron_dra
+
+import (
+	"sort"
+	"strconv"
+	"strings"
+
+	"github.com/aws/amazon-cloudwatch-agent-test/util/otelmetrics"
+)
+
+// escapePromQL escapes a string value for safe use inside PromQL label matches.
+func escapePromQL(s string) string {
+	return strings.NewReplacer(`\`, `\\`, `"`, `\"`).Replace(s)
+}
+
+// filterByHostType returns results matching the given host.type.
+func filterByHostType(results []otelmetrics.MetricResult, hostType string) []otelmetrics.MetricResult {
+	var out []otelmetrics.MetricResult
+	for _, r := range results {
+		r := r
+		if r.Labels.Resource["host.type"] == hostType {
+			out = append(out, r)
+		}
+	}
+	return out
+}
+
+// uniqueDatapointValuesList returns the sorted unique non-empty values of a
+// datapoint-level attribute across all results.
+func uniqueDatapointValuesList(results []otelmetrics.MetricResult, key string) []string {
+	seen := make(map[string]struct{})
+	for _, r := range results {
+		r := r
+		if v, ok := r.Labels.Datapoint[key]; ok && v != "" {
+			seen[v] = struct{}{}
+		}
+	}
+	out := make([]string, 0, len(seen))
+	for v := range seen {
+		out = append(out, v)
+	}
+	sort.Strings(out)
+	return out
+}
+
+// uniqueDatapointPairs collects distinct (a, b) pairs from two datapoint keys.
+func uniqueDatapointPairs(results []otelmetrics.MetricResult, keyA, keyB string) [][2]string {
+	set := make(map[[2]string]struct{})
+	for _, r := range results {
+		r := r
+		a := r.Labels.Datapoint[keyA]
+		b := r.Labels.Datapoint[keyB]
+		if a == "" || b == "" {
+			continue
+		}
+		set[[2]string{a, b}] = struct{}{}
+	}
+	out := make([][2]string, 0, len(set))
+	for p := range set {
+		out = append(out, p)
+	}
+	return out
+}
+
+// isIntLike returns true if s is a non-negative integer string.
+func isIntLike(s string) bool {
+	if s == "" {
+		return false
+	}
+	_, err := strconv.Atoi(s)
+	return err == nil
+}
+
+// setKeys returns the sorted keys of a string set (for messages).
+func setKeys(m map[string]struct{}) []string {
+	out := make([]string, 0, len(m))
+	for k := range m {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}
diff --git a/test/otel/neuron_dra/metrics_test.go b/test/otel/neuron_dra/metrics_test.go
new file mode 100644
index 00000000..d20a38e0
--- /dev/null
+++ b/test/otel/neuron_dra/metrics_test.go
@@ -0,0 +1,18 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+package neuron_dra
+
+import "github.com/aws/amazon-cloudwatch-agent-test/util/otelmetrics"
+
+// Neuron metric definitions (same source names as the device-plugin neuron package;
+// OTel preserves raw neuron-monitor names regardless of allocation mechanism).
+// neuroncore_utilization_ratio is the per-core metric the correlation tests use.
+var neuronMetrics = []otelmetrics.MetricDefinition{
+	{Name: "neuroncore_utilization_ratio", MetricType: "gauge", Scope: otelmetrics.ScopePod, Unit: "1"},
+	{Name: "neuroncore_memory_usage_model_shared_scratchpad", MetricType: "gauge", Scope: otelmetrics.ScopePod, Unit: "By"},
+	{Name: "neuron_runtime_memory_used_bytes", MetricType: "gauge", Scope: otelmetrics.ScopePod, ExpectedLabels: []string{"memory_location"}, Unit: "By"},
+	{Name: "execution_latency_seconds", MetricType: "gauge", Scope: otelmetrics.ScopePod, ExpectedLabels: []string{"percentile"}, Unit: "s"},
+}
diff --git a/test/otel/neuron_dra/neuron_dra_test.go b/test/otel/neuron_dra/neuron_dra_test.go
new file mode 100644
index 00000000..45d0f808
--- /dev/null
+++ b/test/otel/neuron_dra/neuron_dra_test.go
@@ -0,0 +1,219 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+// Package neuron_dra validates per-device/per-core Neuron pod correlation on the
+// Dynamic Resource Allocation (DRA) path — the counterpart of the device-plugin
+// `neuron` package's multi_device tests. Neuron devices are allocated via the
+// AWS Neuron DRA driver (DeviceClass neuron.aws.com, driver neuron.aws.com) through
+// a ResourceClaimTemplate instead of the aws.amazon.com/neuron device-plugin
+// resource. The emitted metrics are identical, so the assertions mirror the
+// device-plugin tests.
+//
+// Cluster topology:
+//   - 1x trn1.2xlarge = 1 Neuron (Trainium) device × 2 cores = 2 cores
+//   - Neuron exposed via the Neuron DRA driver, not the device plugin
+//   - A burn workload (neuron-burn-dra) claims 1 whole device via DRA
+//
+// Why Trainium (trn1) and not Inferentia (inf2): the AWS Neuron DRA driver
+// (neuron-helm-chart, driver image 1.2.0) supports Trainium only — it explicitly
+// rejects inf1/inf2 at device discovery ("unsupported instance type"). trn1.2xlarge
+// is the smallest, most reliably-available Trainium instance (1 device / 2 cores),
+// which is enough to exercise the DRA correlation code path end to end: a claimed
+// device's two cores must both attribute to the claiming pod, and to no other pod.
+// The multi-device "1 claimed + N-1 unclaimed" matrix is covered by the
+// device-plugin multi_device_test; here the device count is 1 so the whole node's
+// single device is the claimed device.
+package neuron_dra
+
+import (
+	"context"
+	"fmt"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+const (
+	neuronInstanceType        = "trn1.2xlarge"
+	expectedNeuronDeviceCount = 1
+	expectedNeuronCoreCount   = 2
+	expectedCoresPerDevice    = 2
+
+	// burnPodPrefix is the DRA burn workload that claims 1 whole Neuron device
+	// (= 2 cores) via a ResourceClaimTemplate.
+	burnPodPrefix = "neuron-burn-dra"
+	// A claim for 1 whole device yields 2 cores on 1 device.
+	expectedClaimedDevices = 1
+	expectedClaimedCores   = 2
+)
+
+// TestNeuronDRADeviceCount validates the node exposes the expected Neuron devices.
+func TestNeuronDRADeviceCount(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "neuroncore_utilization_ratio")
+	require.NoError(t, err, "querying neuroncore_utilization_ratio")
+	require.NotEmpty(t, results, "neuroncore_utilization_ratio not available")
+
+	node := filterByHostType(results, neuronInstanceType)
+	require.True(t, len(node) > 0, "No results from %s node", neuronInstanceType)
+
+	devices := uniqueDatapointValuesList(node, "aws.neuron.device")
+	require.Equal(t, expectedNeuronDeviceCount, len(devices),
+		"Expected %d distinct aws.neuron.device on %s, got %d: %v",
+		expectedNeuronDeviceCount, neuronInstanceType, len(devices), devices)
+}
+
+// TestNeuronDRACoreCount validates the total (device, core) pair count.
+func TestNeuronDRACoreCount(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "neuroncore_utilization_ratio")
+	require.NoError(t, err, "querying neuroncore_utilization_ratio")
+	require.NotEmpty(t, results, "neuroncore_utilization_ratio not available")
+
+	node := filterByHostType(results, neuronInstanceType)
+	require.True(t, len(node) > 0, "No results from %s node", neuronInstanceType)
+
+	pairs := uniqueDatapointPairs(node, "aws.neuron.device", "aws.neuron.core")
+	require.Equal(t, expectedNeuronCoreCount, len(pairs),
+		"Expected %d (device, core) pairs on %s, got %d",
+		expectedNeuronCoreCount, neuronInstanceType, len(pairs))
+}
+
+// TestNeuronDRACoresPerDevice validates each device exposes the expected cores.
+func TestNeuronDRACoresPerDevice(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "neuroncore_utilization_ratio")
+	require.NoError(t, err, "querying neuroncore_utilization_ratio")
+	require.NotEmpty(t, results, "neuroncore_utilization_ratio not available")
+
+	node := filterByHostType(results, neuronInstanceType)
+	require.True(t, len(node) > 0, "No results from %s node", neuronInstanceType)
+
+	coresByDevice := make(map[string]map[string]struct{})
+	for _, r := range node {
+		r := r
+		dev := r.Labels.Datapoint["aws.neuron.device"]
+		core := r.Labels.Datapoint["aws.neuron.core"]
+		if dev == "" || core == "" {
+			continue
+		}
+		if coresByDevice[dev] == nil {
+			coresByDevice[dev] = make(map[string]struct{})
+		}
+		coresByDevice[dev][core] = struct{}{}
+	}
+	require.True(t, len(coresByDevice) > 0, "No devices with core metrics on %s", neuronInstanceType)
+	for dev, cores := range coresByDevice {
+		require.Equal(t, expectedCoresPerDevice, len(cores),
+			"Neuron device %s: expected %d cores, got %d", dev, expectedCoresPerDevice, len(cores))
+	}
+}
+
+// TestNeuronDRADeviceIndicesAreIntegers validates device and core are integers.
+func TestNeuronDRADeviceIndicesAreIntegers(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "neuroncore_utilization_ratio")
+	require.NoError(t, err, "querying neuroncore_utilization_ratio")
+	node := filterByHostType(results, neuronInstanceType)
+	require.True(t, len(node) > 0, "No results from %s node", neuronInstanceType)
+
+	for _, r := range node {
+		r := r
+		dev := r.Labels.Datapoint["aws.neuron.device"]
+		core := r.Labels.Datapoint["aws.neuron.core"]
+		require.True(t, isIntLike(dev), "aws.neuron.device should be integer, got '%s'", dev)
+		require.True(t, isIntLike(core), "aws.neuron.core should be integer, got '%s'", core)
+	}
+}
+
+// TestNeuronDRAClaimedDeviceCorrelation is the DRA-path per-device correlation guard.
+// The burn pod claims exactly 1 whole Neuron device via DRA, so both of that device's
+// cores must attribute to that single pod — and to exactly one pod. This fails if DRA
+// correlation collapses cores onto the wrong pod, attributes an unclaimed core to a
+// pod, or splits a device across pods. On this single-device node the node's one
+// device is the claimed device.
+func TestNeuronDRAClaimedDeviceCorrelation(t *testing.T) {
+	t.Parallel()
+	results, err := queryCache.Get(context.Background(), "neuroncore_utilization_ratio")
+	require.NoError(t, err, "querying neuroncore_utilization_ratio")
+	node := filterByHostType(results, neuronInstanceType)
+	require.NotEmpty(t, node, "No neuroncore_utilization_ratio results from %s node", neuronInstanceType)
+
+	// For each device, collect the distinct pods its cores are attributed to
+	// (empty = unclaimed), and count how many of its cores carry the burn pod.
+	devicePods := make(map[string]map[string]struct{})
+	claimedCoresByDevice := make(map[string]map[string]struct{})
+	for _, r := range node {
+		r := r
+		dev := r.Labels.Datapoint["aws.neuron.device"]
+		core := r.Labels.Datapoint["aws.neuron.core"]
+		if dev == "" {
+			continue
+		}
+		if devicePods[dev] == nil {
+			devicePods[dev] = make(map[string]struct{})
+		}
+		pod := r.Labels.Resource["k8s.pod.name"]
+		if pod == "" {
+			continue
+		}
+		devicePods[dev][pod] = struct{}{}
+		if len(pod) >= len(burnPodPrefix) && pod[:len(burnPodPrefix)] == burnPodPrefix {
+			if claimedCoresByDevice[dev] == nil {
+				claimedCoresByDevice[dev] = make(map[string]struct{})
+			}
+			claimedCoresByDevice[dev][core] = struct{}{}
+		}
+	}
+	require.Len(t, devicePods, expectedNeuronDeviceCount,
+		"expected %d Neuron device(s) on the node, got %d", expectedNeuronDeviceCount, len(devicePods))
+
+	var claimedDevices []string
+	for dev, pods := range devicePods {
+		switch len(pods) {
+		case 0:
+			// No pod on any core of this device.
+		case 1:
+			claimedDevices = append(claimedDevices, dev)
+		default:
+			// A single device whose cores span multiple pods is a collapse symptom.
+			t.Errorf("Neuron device %s correlated to multiple pods %v", dev, setKeys(pods))
+		}
+	}
+
+	// Exactly 1 device correlated, and exactly 2 of its cores → the burn pod.
+	require.Len(t, claimedDevices, expectedClaimedDevices,
+		"expected %d claimed Neuron device(s), got %d: %v",
+		expectedClaimedDevices, len(claimedDevices), claimedDevices)
+	for _, dev := range claimedDevices {
+		require.Len(t, claimedCoresByDevice[dev], expectedClaimedCores,
+			fmt.Sprintf("expected %d cores of device %s correlated to %s*, got %d",
+				expectedClaimedCores, dev, burnPodPrefix, len(claimedCoresByDevice[dev])))
+	}
+}
+
+// TestNeuronDRABurnPodLabels validates the correlated burn pod's labels via a
+// targeted PromQL query (avoids stale series in the shared OTLP store).
+func TestNeuronDRABurnPodLabels(t *testing.T) {
+	t.Parallel()
+	ctx := context.Background()
+	promql := fmt.Sprintf(
+		`neuroncore_utilization_ratio{"@resource.k8s.cluster.name"="%s","@resource.k8s.pod.name"=~"%s.*"}`,
+		escapePromQL(cfg.ClusterName), burnPodPrefix)
+	burn, err := client.Query(ctx, promql)
+	require.NoError(t, err, "querying neuroncore_utilization_ratio for %s", burnPodPrefix)
+	require.Equal(t, expectedClaimedCores, len(burn),
+		"Expected %d cores correlated to %s (1 whole device × 2 cores), got %d",
+		expectedClaimedCores, burnPodPrefix, len(burn))
+
+	devices := make(map[string]struct{})
+	for _, r := range burn {
+		r := r
+		devices[r.Labels.Datapoint["aws.neuron.device"]] = struct{}{}
+		require.Equal(t, "default", r.Labels.Resource["k8s.namespace.name"], "%s namespace", burnPodPrefix)
+	}
+	require.Len(t, devices, expectedClaimedDevices,
+		"Expected %s cores on %d device, got %d", burnPodPrefix, expectedClaimedDevices, len(devices))
+}
diff --git a/test/otel/neuron_dra/setup_test.go b/test/otel/neuron_dra/setup_test.go
new file mode 100644
index 00000000..ac18d598
--- /dev/null
+++ b/test/otel/neuron_dra/setup_test.go
@@ -0,0 +1,98 @@
+//go:build integration
+
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: MIT
+
+package neuron_dra
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"os"
+	"testing"
+	"time"
+
+	awsconfig "github.com/aws/aws-sdk-go-v2/config"
+	"github.com/aws/aws-sdk-go-v2/service/sts"
+
+	"github.com/aws/amazon-cloudwatch-agent-test/environment"
+	"github.com/aws/amazon-cloudwatch-agent-test/util/otelmetrics"
+)
+
+var (
+	cfg        otelmetrics.TestConfig
+	client     *otelmetrics.OtelMetricsClient
+	queryCache *otelmetrics.QueryCache
+)
+
+// The DRA cluster has a single Neuron (Trainium) node (trn1.2xlarge).
+var clusterHostTypes = []string{neuronInstanceType}
+
+func TestMain(m *testing.M) {
+	environment.RegisterEnvironmentMetaDataFlags()
+	flag.Parse()
+	env := environment.GetEnvironmentMetaData()
+
+	region := env.Region
+	if region == "" {
+		region = os.Getenv("AWS_REGION")
+	}
+	if region == "" {
+		fmt.Fprintf(os.Stderr, "Region not set\n")
+		os.Exit(1)
+	}
+
+	clusterName := env.EKSClusterName
+	if clusterName == "" {
+		clusterName = os.Getenv("CLUSTER_NAME")
+	}
+	if clusterName == "" {
+		fmt.Fprintf(os.Stderr, "Cluster name not set\n")
+		os.Exit(1)
+	}
+
+	ctx := context.Background()
+	awsCfg, err := awsconfig.LoadDefaultConfig(ctx, awsconfig.WithRegion(region))
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "AWS config error: %v\n", err)
+		os.Exit(1)
+	}
+	stsClient := sts.NewFromConfig(awsCfg)
+	identity, err := stsClient.GetCallerIdentity(ctx, &sts.GetCallerIdentityInput{})
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "STS GetCallerIdentity error: %v\n", err)
+		os.Exit(1)
+	}
+
+	cfg = otelmetrics.TestConfig{
+		Region:         region,
+		Endpoint:       fmt.Sprintf("https://monitoring.%s.amazonaws.com", region),
+		Timeout:        30 * time.Second,
+		MaxRetries:     3,
+		ClusterName:    clusterName,
+		AccountID:      *identity.Account,
+		SigningService: "monitoring",
+	}
+
+	client, err = otelmetrics.NewClient(ctx, cfg)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "Client error: %v\n", err)
+		os.Exit(1)
+	}
+
+	hostMappings := []otelmetrics.SourceHostMapping{
+		{Source: otelmetrics.SourceNeuron, HostTypes: clusterHostTypes},
+	}
+
+	registry := otelmetrics.NewSourceRegistry(clusterHostTypes, hostMappings,
+		otelmetrics.SourceMapping{Source: otelmetrics.SourceNeuron, Metrics: neuronMetrics},
+	)
+
+	queryCache = otelmetrics.NewQueryCache(client, cfg.ClusterName,
+		otelmetrics.WithHostTypes(clusterHostTypes),
+		otelmetrics.WithSourceRegistry(registry),
+	)
+
+	os.Exit(m.Run())
+}