GPU device passthrough is an experimental feature available in commercial Deckhouse Platform (DP) editions.

DP attaches physical GPU devices to virtual machines using DRA (Dynamic Resource Allocation). A device is requested by a reference to a GPUClass in the .spec.gpus block of the VirtualMachine resource.

An administrator prepares the GPUClass resources, so ask them which classes are available in the cluster.

To request a GPU device, add the .spec.gpus block to the machine specification:

apiVersion: virtualization.deckhouse.io/v1alpha2
kind: VirtualMachine
metadata:
  name: linux-vm
spec:
  # ... other VM settings ...
  gpus:
    - gpuClassName: nvidia-h100

In the gpuClassName parameter, specify the name of an existing GPUClass resource. To attach several devices, add more elements to the list, their order doesn’t matter. A single machine takes no more than 16 devices.

A change to the .spec.gpus block applies only after the virtual machine restarts.

Additional resources