Getting Started with GPU Workloads in Managed Kubernetes

GPU node groups in UpCloud Managed Kubernetes come with GPU drivers pre-installed and the GPU is exposed to the container runtime.

However, the NVIDIA device plugin is not pre-installed. You must install it so that Kubernetes can use nvidia.com/gpu resource limits.

Prerequisites

• An UpCloud Managed Kubernetes cluster with a GPU node group added

• kubectl and helm configured in your local terminal

Install NVIDIA Device Plugin

Run the following commands to install the NVIDIA device plugin:

helm repo add nvidia https://helm.ngc.nvidia.com/nvidia && helm repo update
helm install nvidia-device-plugin nvidia/nvidia-device-plugin -n kube-system

See nvidia-device-plugin documentation for additional configuration details.

If your cluster also has node groups without GPUs, you can set the device plugin DaemonSet to use a node selector:

helm install nvidia-device-plugin nvidia/nvidia-device-plugin -n kube-system \
  --set nodeSelector.gpu='NVIDIA-L40S'

Device plugin is then only scheduled to nodes with a matching label gpu: NVIDIA-L40S.

Verify the plugin is running

Check that the DaemonSet is scheduled on your GPU nodes and ready:

kubectl -n kube-system get ds nvidia-device-plugin
```xml

You should see `DESIRED` and `AVAILABLE` match the number of GPU nodes in your cluster.

Next, confirm the `nvidia.com/gpu` resource shows up on those nodes:

```bash
# Get nodes
kubectl get nodes

# Replace <GPU_NODE_NAME> with one of your GPU node names
kubectl describe node <GPU-NODE-NAME>

You should see Capacity and Allocatable list nvidia.com/gpu: <count>.

Run a quick GPU smoke test

Create a one-off Job that requests a GPU and prints hardware and driver details:

apiVersion: batch/v1
kind: Job
metadata:
  name: gpu-smoke-test
spec:
  template:
    spec:
      restartPolicy: Never
      containers:
        - name: cuda
          image: nvidia/cuda:12.4.1-base-ubuntu22.04
          command: ["bash", "-lc", "nvidia-smi"]
          resources:
            limits:
              nvidia.com/gpu: 1
```bash

Apply and view output:

```bash
kubectl apply -f gpu-smoke-test.yaml
kubectl wait --for=condition=complete job/gpu-smoke-test
kubectl logs job/gpu-smoke-test

If everything is set up, you’ll see the GPU model and driver info.

Once the plugin is installed and verified, your GPU nodes are ready to schedule real GPU workloads.