Version v1.3 of the documentation is no longer actively maintained. The site that you are currently viewing is an archived snapshot. For up-to-date documentation, see the latest version.

NVIDIA GPU (OSS drivers)

In this guide we’ll follow the procedure to support NVIDIA GPU using OSS drivers on Talos.

Enabling NVIDIA GPU support on Talos is bound by NVIDIA EULA. Talos GPU support has been promoted to beta. The Talos published NVIDIA OSS drivers are bound to a specific Talos release. The extensions versions also needs to be updated when upgrading Talos.

The published versions of the NVIDIA system extensions can be found here:

Upgrading Talos and enabling the NVIDIA modules and the system extension

Make sure to use talosctl version v1.3.7 or later

First create a patch yaml gpu-worker-patch.yaml to update the machine config similar to below:

- op: add
  path: /machine/install/extensions
    - image:
    - image:
- op: add
  path: /machine/kernel
      - name: nvidia
      - name: nvidia_uvm
      - name: nvidia_drm
      - name: nvidia_modeset
- op: add
  path: /machine/sysctls
    net.core.bpf_jit_harden: 1

Update the driver version and Talos release in the above patch yaml from the published versions if there is a newer one available. Make sure the driver version matches for both the nvidia-open-gpu-kernel-modules and nvidia-container-toolkit extensions. The nvidia-open-gpu-kernel-modules extension is versioned as <nvidia-driver-version>-<talos-release-version> and the nvidia-container-toolkit extension is versioned as <nvidia-driver-version>-<nvidia-container-toolkit-version>.

Now apply the patch to all Talos nodes in the cluster having NVIDIA GPU’s installed:

talosctl patch mc --patch @gpu-worker-patch.yaml

Now we can proceed to upgrading Talos to the same version to enable the system extension:

talosctl upgrade

Once the node reboots, the NVIDIA modules should be loaded and the system extension should be installed.

This can be confirmed by running:

talosctl read /proc/modules

which should produce an output similar to below:

nvidia_uvm 1146880 - - Live 0xffffffffc2733000 (PO)
nvidia_drm 69632 - - Live 0xffffffffc2721000 (PO)
nvidia_modeset 1142784 - - Live 0xffffffffc25ea000 (PO)
nvidia 39047168 - - Live 0xffffffffc00ac000 (PO)
talosctl get extensions

which should produce an output similar to below:

NODE           NAMESPACE   TYPE              ID                                                                           VERSION   NAME                             VERSION   runtime     ExtensionStatus            1         nvidia-container-toolkit         515.65.01-v1.10.0   runtime     ExtensionStatus       1         nvidia-open-gpu-kernel-modules   515.65.01-v1.2.0
talosctl read /proc/driver/nvidia/version

which should produce an output similar to below:

NVRM version: NVIDIA UNIX x86_64 Kernel Module  515.65.01  Wed Mar 16 11:24:05 UTC 2022
GCC version:  gcc version 12.2.0 (GCC)

Deploying NVIDIA device plugin

First we need to create the RuntimeClass

Apply the following manifest to create a runtime class that uses the extension:

kind: RuntimeClass
  name: nvidia
handler: nvidia

Install the NVIDIA device plugin:

helm repo add nvdp
helm repo update
helm install nvidia-device-plugin nvdp/nvidia-device-plugin --version=0.11.0 --set=runtimeClassName=nvidia

Apply the following manifest to run CUDA pod via nvidia runtime:

cat <<EOF | kubectl apply -f -
apiVersion: v1
kind: Pod
  name: gpu-operator-test
  restartPolicy: OnFailure
  runtimeClassName: nvidia
  - name: cuda-vector-add
    image: "nvidia/samples:vectoradd-cuda11.6.0"
      limits: 1

The status can be viewed by running:

kubectl get pods

which should produce an output similar to below:

NAME                READY   STATUS      RESTARTS   AGE
gpu-operator-test   0/1     Completed   0          13s
kubectl logs gpu-operator-test

which should produce an output similar to below:

[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory