Drivers for GPU node groups in a Managed Kubernetes cluster

You can create Managed Kubernetes clusters on a cloud server with GPU:

with pre-installed drivers — kernel modules in pre-installed drivers depend on the GPU architecture;
without pre-installed drivers — you select and install drivers yourself. GPU node groups without drivers do not support cluster autoscaling.

Kernel modules in pre-installed drivers

Kernel modules in pre-installed drivers can be:

proprietary — used in GPUs with Pascal, Maxwell, and Volta architecture;
open-source — used in GPUs with Turing architecture and all subsequent architecture generations.

Among the available GPUs, only the NVIDIA® GTX 1080 GPU with Pascal-based architecture has proprietary kernel modules, while the rest have open-source ones.

For more information about kernel modules, see the Kernel Modules section in the NVIDIA® documentation.

Install drivers

To install the driver yourself, use NVIDIA® GPU Operator.

Connect to the cluster.
Install the Helm package manager version 3.7.0 or higher.

Add the nvidia repository to Helm:

helm repo add nvidia https://helm.ngc.nvidia.com/nvidia

Update the nvidia repository in Helm:
```
helm repo update
```
Install NVIDIA GPU Operator and specify the required GPU driver version:
```
helm install \
  --namespace gpu-operator \
  --create-namespace \
  --set driver.version=<driver_version> \
  gpu-operator nvidia/gpu-operator
```
Specify <driver_version> — NVIDIA® driver version. You can find it in the NVIDIA GPU Driver row of the GPU Operator Component Matrix table in the NVIDIA® documentation.

To verify that NVIDIA GPU Operator and the GPU driver are installed correctly, run a GPU application. For example, the CUDA VectorAdd vector addition application:

cat << EOF | kubectl create -f -
apiVersion: v1
kind: Pod
metadata:
name: cuda-vectoradd
spec:
restartPolicy: OnFailure
containers:
- name: cuda-vectoradd
    image: "nvidia/samples:vectoradd-cuda11.2.1"
    resources:
    limits:
        nvidia.com/gpu: 1
EOF

Make sure that the CUDA VectorAdd application has completed successfully — the pod status should be Completed:

kubectl get pods

In the response, the cuda-vectoradd pod will have the status Completed:

NAME                                  READY   STATUS        RESTARTS   AGE
cuda-vectoradd                        0/1     Completed     0          51s