Certify a Cluster

Before you begin

• Install nvcrectl and set up the controller
• Confirm your kubeconfig points at the target cluster

AWS

GB200 (EFA interconnect)

1
apiVersion: cre.nvidia.com/v1alpha1
2
kind: Certification
3
metadata:
4
  name: gb200-cert
5
spec:
6
  target:
7
    nodeSelector:
8
      nvidia.com/gpu.product: NVIDIA-GB200
9
  enableMNNVL: true
10
  categories:
11
    - domain: communication
12
      variant: nccl-all-reduce
13
    - domain: training
14
      variant: nemotron5-8b

$
nvcrectl certification run --cert-file gb200-cert.yaml --wait

The controller auto-detects AWS + GB200 and applies EFA-specific resources (hugepages-2Mi, vpc.amazonaws.com/efa: 4, EFA hostPath volume) automatically.

GB300 (RoCE interconnect)

Same spec as GB200 with nvidia.com/gpu.product: NVIDIA-GB300. The controller detects GB300 and applies RoCE resource claims (roce-channel) instead of EFA — no hugepages, no EFA volumes.

H100

1
spec:
2
  target:
3
    nodeSelector:
4
      nvidia.com/gpu.product: NVIDIA-H100-80GB-HBM3
5
  enableMNNVL: false

H100 on AWS uses vpc.amazonaws.com/efa: 32. No hugepages or ComputeDomain.

GCP

Content coming soon.

Azure

Content coming soon.

Monitoring progress

$
# Watch overall status
$
kubectl get certifications.cre.nvidia.com -w
$
$
# Watch individual workflows
$
kubectl get workflows.cre.nvidia.com -w
$
$
# Tail controller logs
$
kubectl logs -n cluster-readiness-engine deploy/cluster-readiness-engine-controller -f

Reviewing results

$
nvcrectl certification report <name>

• Passed — all categories met their thresholds. Cluster is ready.
• Failed — one or more categories failed. See Interpret Results for how to read the failed node list and act on it.