NVIDIA DSX AI Factory Enhances GPU Validation with Runtime AI Inventory
News
NVIDIA introduces AICR with k8s-aibom integration to validate GPU platforms and track AI workloads on Kubernetes, enhancing AI governance and traceability.
NVIDIA has announced significant advancements in GPU platform validation and AI workload tracking through the integration of its AI Cluster Runtime (AICR) with the k8s-aibom project. This development, part of the NVIDIA DSX AI Factory Platform, addresses critical gaps in AI infrastructure governance by providing enhanced traceability and reproducibility for AI workloads running on Kubernetes.
Enhanced Validation with AICR
AICR, an open-source modular component of the NVIDIA DSX AI Factory, generates version-locked recipes and reproducible deployment bundles for GPU-accelerated Kubernetes environments. Each recipe defines a specific configuration, including hardware, cloud, operating system, and workload intent. These recipes are validated against deployment paths, ensuring consistency and security. For example, AICR binds upstream sources, charts, and digest-pinned controller images into a tested set, verifying provenance and reviewing security defaults.
Deployment bundles created by AICR include checksums for integrity and optional signing through the aicr bundle --attest command. These bundles are deployable using various tools such as Helm, Flux, and Argo CD, providing flexibility for platform teams. AICR v1.2.0 has demonstrated its capability to deliver validated GPU platforms quickly and repeatably.
AI Workload Inventory with k8s-aibom
k8s-aibom, an open-source project developed by Google, complements AICR by tracking AI workloads running on Kubernetes. It emits CycloneDX 1.6 AI bills of materials (AIBOMs) for recognized workloads, providing detailed insights into models, inference runtimes, and dependencies. The system supports native Dynamo graphs, NIMService, LeaderWorkerSet, DisaggregatedSet, and CronJob resources, among others.
Two stock recipes—h100-gke-cos-inference and h100-gke-cos-inference-dynamo—select k8s-aibom by default, while other Google Kubernetes Engine (GKE) recipes can opt in using the --runtime-inventory flag. During steady-state operation, k8s-aibom has shown efficient resource usage, including approximately four API-server requests per minute, 2.2 millicores average CPU usage, and a 28 to 34 MiB memory working set.
Unified AI Governance
The integration of AICR and k8s-aibom addresses two ends of the AI infrastructure lifecycle. While AICR ensures that GPU platforms are built from validated configurations, k8s-aibom provides visibility into the AI workloads running on those platforms. Together, they enhance AI governance by enabling security reviews, compliance audits, and incident response.
This unified approach helps identify unregistered or shadow AI workloads, detect changes in dependencies, and track external AI APIs and vector databases. By bridging the gap between platform validation and workload traceability, NVIDIA aims to provide a robust foundation for designing, building, and operating AI factories.
