NVIDIA Minerva is an AIOps platform for operating AI factories. It collects telemetry from GPUs, NICs, DPUs, NVLink switches, Spectrum-X Ethernet switches, NCCL, schedulers such as Slurm, and AI workloads, and models how devices, links, and jobs connect. Machine learning on that data detects anomalies, groups related events into incidents, and points to likely root causes, so you can investigate one problem instead of many alerts.
You install NVIDIA Minerva on Docker Swarm or on a Kubernetes cluster you operate. Use its UI, preconfigured dashboards, and AI assistant, Sherlock, or send its data and insights to your own tools, such as Prometheus, Slack, ServiceNow, or AI assistants that support the Model Context Protocol (MCP).
For more information about the benefits and features, see Benefits and Features for NVIDIA Minerva.
Skip Ahead
-
Get Started: Choose Docker Swarm or Kubernetes, and a single-node, multi-node, or hybrid layout.
-
Complete Initial Setup: Connect data sources, assign device roles, import dashboards, and turn on correlation after you install.
-
AI Factory Operational Readiness Review: Validate a new AI factory with step-by-step checks before handoff to production.
-
What's New: See what's new in this release and the known issues.
Use Cases
-
Validate an AI Factory Before Production: Confirm that telemetry arrives from every data source and that GPUs, NVLink, links, BGP, and Slurm jobs pass their checks, then record sign-off.
-
Investigate Incidents and Find Root Causes: Review related events and anomalies grouped into one incident, see where they sit in the topology, and ask Sherlock about them in plain language.
-
Catch Problems Before They Affect Jobs: Get alerted when telemetry deviates from its learned baseline, and send notifications to Slack, email, ServiceNow, or other endpoints.
-
Trace a Job Problem to the Infrastructure: Follow a job slowdown or timeout to the switch, link, or NIC responsible.
Core Concepts
-
Architecture: The Collector Stack gathers telemetry close to your devices. The Platform Stack models, analyzes, stores, and presents it.
-
Network Model: How NVIDIA Minerva represents devices, interfaces, and their relationships as a graph, with dimensions and measures for each object.
-
Anomalies and Incidents: An anomaly is telemetry that deviates from a learned baseline. An incident groups related events and anomalies so you investigate one problem.
Learn More
Last updated: