EnterpriseNVIDIAAMD

Rancher RKE2

SUSE

A security-focused, fully conformant Kubernetes distribution (FIPS-capable) common on-prem. GPUs are enabled with the NVIDIA GPU Operator; RKE2 uses containerd.

Bootstrap with Rancher RKE2
Enablement

NVIDIA GPU Operator (containerd) / AMD GPU Operator

Driver install

GPU Operator installs drivers, or pre-install host drivers and set driver.enabled=false.

Container runtime

Bundled containerd; the operator configures the NVIDIA runtime.

Scheduling

Default scheduler; label/taint GPU nodes; pairs well with Kueue/Volcano for batch.

Partitioning

MIG and time-slicing via the operator.

Networking

Canal/Cilium; Multus + SR-IOV for RDMA fabrics.

Upgrades

RKE2 channel upgrades; re-validate the operator after Kubernetes minor bumps.

GPU enablement
  1. 1

    Point the operator at RKE2 containerd

    RKE2's containerd socket differs from default; set the toolkit/runtime path.

    bash
    helm install --wait gpu-operator nvidia/gpu-operator \
  -n gpu-operator --create-namespace \
  --set toolkit.env[0].name=CONTAINERD_CONFIG \
  --set toolkit.env[0].value=/var/lib/rancher/rke2/agent/etc/containerd/config.toml
  2. 2

    Label GPU nodes

    So workloads and the operator target the right hosts.

  3. 3

    Validate

    Run a CUDA pod to confirm scheduling and driver access.

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
  • ·The containerd socket path must be set for the operator's toolkit.
  • ·AMD ROCm support depends on kernel/driver compatibility with the host OS.
Best for

Security-conscious on-prem and edge clusters, including air-gapped installs.