Verified, then logged.
Sun 22 Nov 2026 · 00:34 KST / 15:34 UTC한국어 | ENSearch /

Breaking02 / IT

NVIDIA publishes DSX OS procedures for managing GPU clusters

In a data center aisle, an out-of-focus technician seen from behind slides a repaired GPU server tray back into an empty rack slot beside thick bundles of network cables. AI-generated image

NVIDIA published operating procedures for DSX OS on its developer blog on 5 Nov 2026, setting out how GPU cluster operators configure, certify and run their systems. The certification example uses API version v1alpha1.

The work is split across three open source tools, according to the blog. AI Cluster Runtime (AICR) pins configuration versions in a recipe. Cluster Readiness Engine (NVCRE) checks readiness by running training jobs. NVSentinel continuously monitors GPU, node and network health.

NVIDIA said revalidation starts automatically once it is enabled in the configuration. The NVSentinel controller requests validation when a node joins the cluster or returns after repair. A node that passes has its scheduling block lifted. A node that still fails after retries stays blocked.

NVIDIA said a performance drop in one GPU can slow distributed training. NVCRE has no thresholds by default, so a node that misses its target can still pass. In the words of the post, "A node performing below the target can pass."

Compatibility. NVIDIA said AICR recipes are converted into artefacts for tools including Helm, Argo CD and Flux. A default NVSentinel installation only reports faults. After a driver upgrade or fabric change, operators have to request validation themselves.

Where it runs. NVIDIA said the three projects are released under the Apache 2.0 licence. The blog's examples assume a Kubernetes GPU cluster. The CLI tools, controllers and an NVCRE test provider must also be configured.

NVIDIA's post gives no release number for DSX OS as a whole and no schedule for a general release.