Understand this story
On November 5, 2026, NVIDIA launched the open-source software layer DSX OS for configuring, certifying, and operating Kubernetes GPU clusters. According to an article published by NVIDIA Developer, DSX OS comprises three Apache 2.0 projects: AICR, NVCRE, and NVSentinel. AICR is responsible for version-locked cluster configuration, rendering the configuration into artifacts for deployment tools such as Argo CD, Flux, and Helm, and supporting drift detection; NVCRE performs readiness certification based on NCCL and training workloads, using binary search to locate suspicious nodes; NVSentinel is responsible for continuous health monitoring and fault isolation and remediation, reporting only node status by default, and once enabled, can cordon, evict, and remediate faulty nodes. The three can be chained into a closed lifecycle loop, automatically triggering re-validation after a node joins or is remediated, and only lifting scheduling restrictions after it passes. Currently, this release remains a single-source report, with no subsequent developments yet.
AI-generated from public reports
Developments
1 confirmed developments
NVIDIA launches DSX OS for configuring, certifying, and operating Kubernetes GPU clusters
Source:NVIDIA Developer · Generative AI
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 6
NVIDIA Developer · Generative AIFirst-party
NVIDIA Open-Sources Three DSX OS Components for Configuring, Validating, and Operating Kubernetes GPU Clusters
NVIDIA has released DSX OS, an open-source software layer comprising three Apache 2.0 projects—AICR, NVCRE, and NVSentinel—responsible respectively for version-locked cluster configuration, readiness validation based on NCCL and training workloads, and continuous health monitoring with fault isolation and remediation. AICR renders configurations into artifacts for deployment tools such as Argo CD, Flux, and Helm, and supports drift detection; NVCRE uses binary search to pinpoint suspect nodes; NVSentinel by default only reports node status, and once enabled can cordon, evict, and remediate faulty nodes. The three can be chained into a closed lifecycle loop: after a node joins or is remediated, revalidation is triggered automatically, and scheduling restrictions are lifted only after it passes.