NVIDIA tests cluster readiness with real distributed workloads

NVIDIA Cluster Readiness Engine runs distributed workloads across GPU nodes before production jobs arrive. Its Kubernetes resources organize tests, record measurements and attribute failures to nodes and categories. When a whole group performs poorly, diagnostic mode splits it and reruns smaller groups to narrow the suspects. Passing component health checks alone does not establish that a cluster performs correctly together.

NVIDIA · Sep 23, 2026

Why it matters

Useful for platform teams diagnosing slow collective communication or failures that appear only under distributed load.

More context

Try the open-source controller on a test Kubernetes GPU cluster. Choose representative communication or training workloads and supply your own performance thresholds; none ship by default. Reserve capacity because these tests generate real load.

Keep in mind

The example API is v1alpha1. Results cover the selected workload and topology, not every future job. Suspect-node isolation is not proof of a unique hardware cause.

Original read