NVIDIA ships AICR 1.0 for GPU clusters

NVIDIA has released version 1.0 of AI Cluster Runtime (AICR), an open source tool for configuring GPU-accelerated Kubernetes clusters. Those clusters depend on many components that ship on separate release cycles, including kernels, GPU drivers, container runtimes, networking, storage, operators and workload frameworks. A combination that works for one service, GPU generation or Kubernetes release can fail silently on another.

AICR handles this with version-locked "recipes" that pin component combinations known to work together. The tool has four separate functions. Snapshot records the observed cluster state. Recipe defines the desired configuration. Bundle renders that configuration for Helm, Argo CD, Flux or Helmfile. Validation checks the running cluster against the recipe and can run deployment, conformance and performance tests. The validation results are recorded as signed evidence that identifies what was tested, which checks passed and who signed off.

The main change in v1.0 is a compatibility contract. It covers the aicr CLI, the aicrd REST API and its OpenAPI spec, the Go SDK package pkg/client/v1, and the generated bundle layout and artifact schemas. Each interface has a committed baseline that is checked before changes merge. Removing or breaking a stable interface now requires a new major version.

NVIDIA says the recipe library has grown over six months from a handful of recipes to cover major Kubernetes services and its current accelerators. A public validation dashboard lets users filter recipes by service, GPU, OS and workload. Outside contributors can submit signed evidence for environments that NVIDIA does not test.

Pulumi Labs offers AICR through an infrastructure-as-code provider, and Mirantis has integrated it into k0rdent for multi-cluster management. NVIDIA says the project has more than 100 contributors, nearly half from outside the company.