NVIDIA has announced the release of AICR v1.0 (AI Cluster Runtime version 1.0), a new framework designed to create open, stable, and verifiable configurations for GPU-accelerated Kubernetes clusters. This update establishes a crucial compatibility contract across its public interfaces, enabling developers and system integrators to build upon the platform with confidence.
The Need for Cluster Standardization
Building GPU-accelerated Kubernetes clusters is inherently complex because they rely on compatible versions of numerous components. These elements—including host kernels, GPU drivers, container runtimes, networking tools, storage systems, operators, and various workload frameworks—each operate on independent release schedules. A configuration that successfully operates for a specific service, GPU generation, or Kubernetes version might unexpectedly fail when deployed in a different environment. Furthermore, diagnosing version conflicts after deployment can be a slow and error-prone process.
NVIDIA AI Cluster Runtime (AICR) addresses this challenge by providing version-locked, validated recipes. Each recipe specifies the exact combination of components that are known to work together. These recipes then generate deployment artifacts compatible with major deployment tools such as Helm, Argo CD, Flux, or Helmfile, and they are accompanied by signed validation evidence proving the cluster’s operational status on the hardware used for testing.
Core Capabilities of AICR
The v1.0 release is significant because it formalizes stability across multiple technical layers, including the CLI, REST API, Go SDK, bundle layout, and artifact schemas. This stability allows system operators and integrators to rely on public interfaces.
AICR operates through four distinct, yet combinable, capabilities:
- Snapshot: Records the actual, observed state of a cluster, capturing details about the Kubernetes version, operating system, kernel, GPU type, and overall topology.
- Recipe: Defines the desired, version-locked configuration, along with the specific constraints and validation phases that must be met.
- Bundle: Translates the recipe into deployable artifacts suitable for the operator’s preferred deployment mechanism.
- Validation: Compares the defined recipe against the observed cluster state. When instructed, it can run deployment, conformance, and performance checks, providing signed evidence of the results.
The capabilities are designed to be independent. For instance, a Snapshot captures what exists, while a Recipe defines what *should* exist. Standard open-source Continuous Deployment (CD) tools handle the deployment (applying the Bundle), and AICR subsequently validates the running system against the Recipe, recording signed proof of the test.
AICR v1.0 Enhancements and Ecosystem Adoption
The validation dashboard allows operators to search for recipes based on criteria such as the service, GPU type, operating system, intended workload, and optional platform. Users can then review the status of the recipe and examine any published evidence from the hardware configuration tested.
The utility of the recipe model is already expanding across the technology landscape. Pulumi Labs has exposed AICR through an infrastructure-as-code provider, and Mirantis’s k0rdent integration has packaged it for multi-cluster management, demonstrating that the same GPU cluster configuration can be managed using diverse tools. Currently, AICR boasts over 100 distinct contributors, with nearly half of them coming from outside of NVIDIA.
For example, an operator can select specific criteria—such as EKS, GB300 GPUs, Ubuntu OS, for a training workload using Kubeflow—to resolve a pinned recipe. This recipe can then be rendered for Argo CD, deployed via the existing GitOps workflow, and finally validated against the initial recipe, regardless of whether the developer later chooses to render the same configuration for Helm or Flux.
Commitment to Public Interfaces
A key focus of v1.0 is formalizing compatibility rules across all public interfaces. This includes:
- The public commands, flags, exit semantics, and structured output of the `aicr` CLI.
- The OpenAPI contract for the `aicrd` REST API.
- The exported API within the `github.com/NVIDIA/aicr/pkg/client/v1` package.
- The generated bundle layout and AICR artifact schemas.
For robust integration, each public interface now features a committed baseline that must pass checks before any changes can be merged. The release policy strictly defines semantic breaking changes, meaning that any attempt to remove or change a stable public interface incompatibly will mandate a new major release, ensuring long-term stability for third-party developers.
Getting Started and Contributing
Users are encouraged to test available recipes for their specific environments and to utilize the dashboard’s `aicr evidence verify` command if validation evidence is available. Contributions are highly valued, particularly recipes for hardware or cluster combinations not yet covered by the project. Individuals can contribute by proposing new recipes, validating existing recipes in their own clusters and submitting signed proof, or submitting general feedback and bug reports through the designated GitHub issue tracker.
