AICR v1.0 commits NVIDIA to stable interfaces for version-locked GPU cluster recipes

AICR splits the work into four narrow parts. Snapshot records observed cluster state, including Kubernetes, operating system, kernel, GPU and topology. Recipe describes the desired version-locked configuration, along with the constraints and validation phases attached to it. Bundle turns the recipe into artifacts for whatever deployment tool the operator uses. Validation compares the recipe with the observed state ^[7].

AICR does not apply its own bundles. According to NVIDIA, a recipe does not reconcile a cluster. Helm, Argo CD, Flux or Helmfile apply the bundle, and AICR afterward validates the running cluster and records signed evidence of the result ^[8]. I think this is the right split for teams already running GitOps, because the CD tool remains the only component reconciling cluster state ^[8].

Validation exists because a clean install proves little. NVIDIA writes that installation does not confirm components are healthy, that gang scheduling or accelerator discovery work, or that measured results meet a recipe's performance thresholds ^[9]. Validation runs deployment, conformance and performance checks only where the recipe declares them ^[10]. A recipe with no declared phases gets the state comparison and nothing more ^[17].

I treat a recipe's signed evidence the way I treat a vendor benchmark table: a true statement about someone else's cluster. NVIDIA's own post says a configuration can pass validation on one combination of service, GPU generation, fabric type, machine shape and Kubernetes release and still fail on another ^[11]. The evidence carries over to your cluster only when all five match. The validation dashboard filters recipes by service, GPU, operating system, workload intent and an optional platform ^[12]. Fabric type and machine shape are not among those filters, so confirming a match on them means opening each candidate recipe's evidence ^[18].

The compatibility contract is enforced in CI. For v1.0, NVIDIA added committed compatibility baselines and merge-blocking checks around the public integration surfaces ^[3]. A change that breaks a committed baseline is supposed to stop before it merges. Integrators building on the public interfaces get that protection under the v1.x compatibility policy ^[4]. NVIDIA's summary of the policy names the surfaces it covers but not a deprecation window ^[2].

Coverage outside NVIDIA's lab depends on other people's clusters. Contributors can propose recipes for environments the maintainers cannot test, validate them on their own hardware and submit signed evidence for maintainer review ^[13]. AICR has more than 100 distinct contributors, almost half from outside NVIDIA, a share the post announces with an exclamation mark ^[14].