diff --git a/docs/docs/howto/cluster-setup/bare-metal.md b/docs/docs/howto/cluster-setup/bare-metal.md
index 33a75fc3c9..f6663db364 100644
--- a/docs/docs/howto/cluster-setup/bare-metal.md
+++ b/docs/docs/howto/cluster-setup/bare-metal.md
@@ -205,6 +205,8 @@ helm install --wait --generate-name \
   --set sandboxDevicePlugin.repository=ghcr.io/nvidia \
   --set sandboxDevicePlugin.image=nvidia-sandbox-device-plugin \
   --set sandboxDevicePlugin.version=8e76fe81 \
+  --set 'sandboxDevicePlugin.env[0].name=P_GPU_ALIAS' \
+  --set 'sandboxDevicePlugin.env[0].value=pgpu' \
   --set nfd.enabled=true \
   --set nfd.nodefeaturerules=true
 ```
@@ -296,7 +298,7 @@ The above command should yield an output similar to the following, depending on
 
 ```json
 {
-  "nvidia.com/GH100_H100_PCIE": "1"
+  "nvidia.com/pgpu": "1"
 }
 ```
 
diff --git a/docs/docs/howto/workload-deployment/GPU-configuration.md b/docs/docs/howto/workload-deployment/GPU-configuration.md
index 9f871d33f2..c2a08fcba9 100644
--- a/docs/docs/howto/workload-deployment/GPU-configuration.md
+++ b/docs/docs/howto/workload-deployment/GPU-configuration.md
@@ -63,7 +63,7 @@ spec:
     - # ...
       resources:
         limits:
-          "nvidia.com/GH100_H100_PCIE": 1
+          "nvidia.com/pgpu": 1
       env:
         # ...
         - name: NVIDIA_VISIBLE_DEVICES
diff --git a/e2e/gpu/gpu_test.go b/e2e/gpu/gpu_test.go
index 80a6cb610b..9c28f3ba5e 100644
--- a/e2e/gpu/gpu_test.go
+++ b/e2e/gpu/gpu_test.go
@@ -42,19 +42,17 @@ func TestGPU(t *testing.T) {
 	runtimeHandler, err := manifest.RuntimeHandler(platform)
 	require.NoError(t, err)
 
-	var deviceURI, gpuName string
+	var gpuName string
 	switch platform {
 	case platforms.MetalQEMUTDXGPU:
-		deviceURI = "nvidia.com/GB100_B200"
 		gpuName = "NVIDIA B200"
 	case platforms.MetalQEMUSNPGPU:
-		deviceURI = "nvidia.com/GH100_H100_PCIE"
 		gpuName = "NVIDIA H100 PCIe"
 	default:
 		t.Errorf("platform %s does not support GPU tests", platform)
 	}
 
-	resources := kuberesource.GPU(deviceURI)
+	resources := kuberesource.GPU()
 
 	// Since the TDX-GPU testing cluster has multiple GPUs, we run into the drift
 	// explained in [1]. To avoid this, we need to remove the deployment of the direct GPU tester,
diff --git a/internal/kuberesource/resourcegen/main.go b/internal/kuberesource/resourcegen/main.go
index 971b14f642..475296d991 100644
--- a/internal/kuberesource/resourcegen/main.go
+++ b/internal/kuberesource/resourcegen/main.go
@@ -68,7 +68,7 @@ func main() {
 		case "vault":
 			subResources = kuberesource.PatchRuntimeHandlers(kuberesource.Vault(*namespace), "contrast-cc")
 		case "gpu":
-			subResources = kuberesource.PatchRuntimeHandlers(kuberesource.GPU("placeholder-gpu"), "contrast-cc")
+			subResources = kuberesource.PatchRuntimeHandlers(kuberesource.GPU(), "contrast-cc")
 		default:
 			log.Fatalf("Error: unknown set: %s\n", set)
 		}
diff --git a/internal/kuberesource/sets.go b/internal/kuberesource/sets.go
index e65d44a173..edb7f963de 100644
--- a/internal/kuberesource/sets.go
+++ b/internal/kuberesource/sets.go
@@ -658,7 +658,7 @@ done
 }
 
 // GPU returns the resources for deploying a GPU test pod.
-func GPU(deviceURI string) []any {
+func GPU() []any {
 	tester := Deployment("gpu-tester", "").
 		WithSpec(DeploymentSpec().
 			WithReplicas(1).
@@ -679,7 +679,7 @@ func GPU(deviceURI string) []any {
 							WithResources(ResourceRequirements().
 								WithMemoryLimitAndRequest(500). // This accounts for nvidia-smi and the guest pull overhead.
 								WithLimits(corev1.ResourceList{
-									corev1.ResourceName(deviceURI): resource.MustParse("1"),
+									corev1.ResourceName("nvidia.com/pgpu"): resource.MustParse("1"),
 								}),
 							),
 						Container().
diff --git a/packages/upgrade-gpu-operator.sh b/packages/upgrade-gpu-operator.sh
index df1408a806..155fd013cb 100644
--- a/packages/upgrade-gpu-operator.sh
+++ b/packages/upgrade-gpu-operator.sh
@@ -50,9 +50,6 @@ kubectl delete crd nvidiadrivers.nvidia.com --ignore-not-found
 helm repo add nvidia https://helm.ngc.nvidia.com/nvidia && helm repo update
 
 # Upstream instructions from https://github.com/kata-containers/kata-containers/pull/12257
-# The P_GPU_ALIAS environment variable is stripped to get per-GPU annotations for heterogenous
-# clusters. With it, all GPUs will look like this to CDI clients: `nvidia.com/pgpu`. Without
-# it, the annotations are specific to the GPU type: `nvidia.com/GB100_B200`.
 helm install --wait --generate-name \
   -n gpu-operator --create-namespace \
   nvidia/gpu-operator \
@@ -72,6 +69,8 @@ helm install --wait --generate-name \
   --set sandboxDevicePlugin.repository=ghcr.io/nvidia \
   --set sandboxDevicePlugin.image=nvidia-sandbox-device-plugin \
   --set sandboxDevicePlugin.version=8e76fe81 \
+  --set 'sandboxDevicePlugin.env[0].name=P_GPU_ALIAS' \
+  --set 'sandboxDevicePlugin.env[0].value=pgpu' \
   --set nfd.enabled=true \
   --set nfd.nodefeaturerules=true