Cloud-managedManagedNVIDIA

Google GKE

Google Cloud

Managed Kubernetes on GCP. GKE can auto-install NVIDIA drivers on GPU node pools via a DaemonSet, or you can run the GPU Operator.

Bootstrap with Google GKE
Enablement

GKE driver DaemonSet or NVIDIA GPU Operator

Driver install

GKE installs drivers via a Google-managed DaemonSet when the node pool has accelerators; Autopilot handles it automatically.

Container runtime

containerd with NVIDIA runtime.

Scheduling

Default scheduler; accelerator node pools; cluster autoscaler; Autopilot GPU pods.

Partitioning

MIG and time-slicing supported; configure on the node pool or via the operator.

Networking

GKE Dataplane V2 (Cilium); GPUDirect-TCPX/RDMA on A3 for high-throughput training.

Upgrades

Node auto-upgrade; pin driver version if a workload needs a specific CUDA.

GPU enablement
  1. 1

    Create a GPU node pool

    Attach accelerators and let GKE install the driver.

    bash
    gcloud container node-pools create gpu-pool \
  --cluster $CLUSTER --accelerator type=nvidia-h100-80gb,count=8 \
  --machine-type a3-highgpu-8g --num-nodes 2
  2. 2

    Install the driver DaemonSet (if not automatic)

    Applies Google's driver installer to accelerator nodes.

    bash
    kubectl apply -f https://raw.githubusercontent.com/GoogleCloudPlatform/container-engine-accelerators/master/nvidia-driver-installer/cos/daemonset-preloaded.yaml
  3. 3

    Request a GPU in the pod spec

    Set limits: nvidia.com/gpu on the container.

Validate the enablement
verify GPUs
kubectl get nodes -o custom-columns=NAME:.metadata.name,GPU:.status.allocatable.'nvidia\.com/gpu'
kubectl -n gpu-operator get pods
kubectl run cuda-check --rm -it --restart=Never \
  --image=nvidia/cuda:12.4.1-base-ubuntu22.04 \
  --limits=nvidia.com/gpu=1 -- nvidia-smi
Known limitations
  • ·GPUDirect-TCPX/RDMA is limited to specific machine families (A3).
  • ·Driver DaemonSet vs operator: pick one to avoid conflicts.
Best for

GCP platforms wanting minimal driver management, and A3 for large training.