NVIDIA previews Vera Rubin cluster recipes

NVIDIA’s announcement introduces Preview configuration recipes for Vera Rubin clusters, covering training and inference on RKE2 and Ubuntu. The company reports successful deployment and conformance checks for both variants, but inference serving performance remains unvalidated in this work. Published evidence records the configuration captured during testing; because these Preview recipes keep changing, it does not attest to today’s recipe contents.

The project guide explains the operator’s part: provide an existing GPU Kubernetes cluster and deployment tooling; AICR does not provision the cluster. Describe the target hardware, operating system and workload to obtain a recipe with fixed component versions, then render a bundle for the chosen deployment tool. After deploying that bundle, validate the running cluster against the recipe stored inside it. The validator compares configuration requirements with a snapshot of actual cluster state and can run additional checks inside the cluster.

The maturity rules reserve Supported status for recipes that resolve and bundle, deploy on real hardware, pass their declared checks and have public, verifiable evidence. Its signer must be recorded, allowlisted in recipes/evidence/allowlist.yaml and verifiable; first-party evidence is not required. Qualification concerns a particular service, accelerator, operating system, workload and platform combination. A configuration file’s presence alone proves none of this. Planned work is outside the support commitment.