NVIDIA has issued the first stable release of AI Cluster Runtime (AICR), open-source software for configuring and checking GPU-accelerated Kubernetes clusters. Version 1.0 freezes the command-line interface, REST API, Go SDK, bundle layout and artifact schemas; breaking changes to those surfaces must wait for a major release.
GPU clusters combine software components that can change on different schedules, including drivers, kernels, network operators and workload tools. AICR recipes describe a specific mix of cloud service, accelerator, operating system and workload, then generate deployment files for tools such as Helm, Argo CD, Flux or Helmfile. Those tools install the configuration; AICR checks the live cluster against the recipe.
The v1.0 upgrade check can compare a proposed recipe with components already installed and warn about live objects an upgrade could delete. The release also adds tests of GPU and network capability, rather than checking only whether Kubernetes resources exist. NVIDIA says its GKE checks verify that each A3 node exposes all eight GPU network interfaces; a Slurm test checks that a job assigned one GPU can use only that GPU, while a job assigned none has no GPU access.
AICR generates configuration and validation evidence but does not provision or manage clusters. Compatibility is limited to the combinations its recipes cover.
Sources
- NVIDIA/aicr: project documentation
- NVIDIA/aicr: v1.0.0 release notes
- Illustrative server-room photo by Kevin Ache on Unsplash, used under the Unsplash License.