NVIDIA has released a new open-source tool called the Cluster Readiness Engine, or NVCRE. It is built to check whether GPU clusters are ready to handle real AI work.
The tool runs on Kubernetes, a popular system for managing computer workloads. Instead of relying only on standard checks, it runs actual AI jobs to test how a cluster performs.
These jobs include training benchmarks and communication tests. The results help find slow nodes, worn-out hardware, and setup errors that basic diagnostics can miss.
Why AI Clusters Need Testing
AI training jobs can fail or slow down because of small problems. A single slow GPU or a mistake in how network traffic is routed can drag down an entire job.
NVIDIA gave the example of a training run using 512 GPUs. In a job that size, one weak point can slow everything down.
NVCRE is designed to find the exact node or link causing the trouble. According to NVIDIA, this can save teams days of troubleshooting.
The tool runs its tests before workloads go live. This means problems can be found and fixed early.
How NVCRE Works
NVCRE works as a Kubernetes controller. It uses custom resources called Certification, Workflow, and Job to organize testing.
These resources set which nodes to test, run the distributed jobs, and report failures. Reports include machine-readable reasons, such as flagging a node for bandwidth issues or hardware wear.
The tool can test at different sizes. It supports checks within a single node, within a rack, and across a full cluster.
A feature called adaptive fault isolation helps narrow down problems when several nodes fail at once. NVIDIA says this task can otherwise take a lot of engineering time.
NVIDIA says the tool is meant for cloud providers, infrastructure teams, and businesses running large AI systems. Uses include acceptance testing before production, burn-in certification, and checking distributed training.
NVCRE supports NVIDIA's GB300 architecture, which is built for high-performance AI. It also works with other NVIDIA software.
That includes NVIDIA AI Cluster Runtime, which standardizes cluster setups, and NVSentinel, which monitors cluster health while it runs. Together, the three tools cover setup, testing, and monitoring.
The software works on AWS, Google Cloud, and Microsoft Azure, as well as on company-owned hardware. It requires Kubernetes version 1.29 or later and NVIDIA's GPU Operator.
Teams can install it using a command-line installer or a Helm chart. Documentation is available on GitHub.
NVIDIA currently has a market value of about $5.48 trillion, according to the report.
The latest stable version of NVCRE, version 0.3.0, was released on September 14, 2026. It adds Helm chart digest publication and Sigstore and SLSA provenance, which help verify that deployments are secure.
