NVIDIA’s AICR 1.2.0 Adds Operator Preview, Updates GKE Recipes

NVIDIA’s Nov. 2 AICR 1.2.0 update adds an opt-in operator that checks deployment bundle recipes and requires approval before rollout, giving Kubernetes operators another validation step. The release also updates AICR’s Google Kubernetes Engine (GKE) recipes for Dynamo with k8s-aibom 1.6.2.

The operator runs a pinned AICR image inside a Kubernetes cluster, checks bundle recipes and fingerprints inventory specifications. Its rollout gate requires approval before a bundle proceeds. The operator defaults to AICR 1.1.2, the version used to test it.

The --split flag outputs two bundle groups, labeled “core” and “ops.” Core covers requirements for a GPU pod to run, including fabric and installation prerequisites. Ops covers other components, including workload platforms.

Version 1.2.0 offers an Ubuntu-based training configuration for GB200 on NVIDIA Kubernetes Engine (NVKE), including a Kubeflow variant. That setup pairs SR-IOV InfiniBand with deployment of Multus and CNI plugins. The release lists published evidence for a Cloud Native Computing Foundation AI Conformance submission for NVKE v1.36.

Dynamo is NVIDIA’s distributed inference framework for coordinating model serving across GPU workers. AICR creates environment-specific recipes with pinned versions and validation checks, then renders deployment bundles. The changes build on earlier AICR deployment guide coverage.

A recipe marked “pass” means it resolves and its charts are pinned; that status alone does not establish hardware validation. Performance validation is a separate phase.