AICR v1.0: Open, stable, and verifiable GPU cluster configuration

Overview of AICR v1.0

NVIDIA AI Cluster Runtime (AICR) v1.0 introduces a structured, version-locked approach to GPU-accelerated Kubernetes cluster configuration. The system addresses the complexity of managing compatibility across a wide range of components, including host kernels, GPU drivers, container runtimes, networking, storage, and workload frameworks. Each of these components operates on its own release cycle, and ensuring compatibility across them is a significant challenge for operators and integrators.

AICR v1.0 provides a stable compatibility contract across its public interfaces, including the CLI, REST API, Go SDK, bundle layout, and artifact schemas. This ensures that operators and integrators can build on AICR’s public interfaces with confidence. The release also introduces four core capabilities—Snapshot, Recipe, Bundle, and Validation—that work together to define, deploy, and verify GPU cluster configurations.

Key Features and Capabilities

The AICR v1.0 release includes several key features that enhance the reliability and reproducibility of GPU cluster configurations. These include:

• Snapshot: Records the observed state of a cluster, including details about the Kubernetes version, operating system, kernel, GPU, and topology.
• Recipe: Describes the desired configuration, including version-locked component combinations and validation phases.
• Bundle: Renders the recipe into deployment artifacts compatible with tools like Helm, Argo CD, Flux, or Helmfile.
• Validation: Compares the recipe with the observed state and runs checks to ensure the cluster meets the intended configuration.

These capabilities are designed to be independent, allowing operators to use them in various sequences. For example, a snapshot can be used to generate a recipe, which can then be rendered into a bundle for deployment. The validation process ensures that the running cluster matches the recipe and records signed evidence of the result.

Validation and Evidence

AICR v1.0 introduces a live validation dashboard that allows operators to search for recipes based on service, GPU, operating system, workload intent, and platform. The dashboard provides visibility into the status of each recipe and any published evidence from the hardware it was tested on. This evidence includes details about what was tested, which checks passed, and who signed the results.

The system also includes supply-chain verification and public evidence aggregation, ensuring that configurations are validated and reproducible. Signed validation evidence is a critical component of AICR, as it provides a verifiable record of the configuration’s performance and compatibility.

Ecosystem Integration and Compatibility

AICR v1.0 has been integrated into various ecosystem tools, including Pulumi Labs and Mirantis’s k0rdent. These integrations demonstrate the value of defining GPU cluster configurations once and consuming them through different infrastructure-as-code and multi-cluster management tools. The system also supports a wide range of Kubernetes services, NVIDIA GPU accelerators, Linux operating systems, and deployment frameworks.

The v1.0 release includes compatibility rules for the AICR CLI, REST API, Go SDK, bundle layout, and artifact schemas. These rules are enforced through committed baselines and merge-blocking checks, ensuring that public interfaces remain stable and backward-compatible. The release policy also defines semantic breaking changes, requiring a new major release if a stable public interface is removed or incompatibly changed.

Getting Started and Contributions

Operators can try AICR by selecting a recipe based on their environment and deploying it through their preferred deployment tooling. The system also includes a command-line interface for inspecting recipe status and validating evidence. Contributions are encouraged, especially for hardware and cluster combinations outside the current project coverage. Users can propose recipes, validate them in their own clusters, and submit signed evidence for review.

The AICR project repository, contributing guide, and issue tracker provide resources for users interested in contributing to the project. These resources include documentation, integration examples, and guidance on proposing new recipes or submitting validation evidence.

Source: NVIDIA Developer. Analysis based on publicly available information; no hands-on testing is claimed.