The system groups thousands of overlapping alerts into single incidents and drains faulty nodes before jobs land.

NVIDIA Minerva collects telemetry from GPUs, NVLink and Spectrum-X fabrics, storage, and running workloads, then links all of it in a dynamic knowledge graph that tracks which infrastructure serves which job at any moment. Supervised and unsupervised models scan that graph for gray failures, such as a single degrading NVLink interface among 400,000 or a GPU running hotter than its peers under identical load.

AI factories run hundreds of thousands of GPUs, scale-up and scale-out fabrics, distributed schedulers, and multi-week training runs as one production system, even when separate teams own each telemetry stream. Network operators watch interface counters, compute operators watch node health, and ML teams watch step times, so the links between symptoms stay hidden.

Topology-aware correlation collapses event storms into single incidents. NVIDIA cites one case where 81 separate pause-frame alerts became one actionable incident, which cut repair time and preserved roughly 54 hours of cluster uptime. Proactive detection of a rare SuperNIC PCIe fault and an NVMe power-state failure drained affected nodes before customer workloads landed, saving hundreds to thousands of GPU hours.

NVIDIA pairs the platform with an AI agent called Sherlock and an MCP server, so operators can query live evidence in natural language and feed operational context into external agentic workflows. The query layer is the piece to watch, because it moves fault evidence out of separate dashboards and into agent workflows that act on it.

What matters

• NVIDIA Minerva maps GPUs, fabric, storage, and jobs into one live knowledge graph.
• Topology-aware correlation turned 81 pause-frame alerts into one incident, saving about 54 cluster hours.
• Sherlock, an AI agent, and an MCP server let operators query live evidence in natural language.

Why it matters

Sherlock, an AI agent, and an MCP server let operators query live evidence in natural language.

This GenAI News article was prepared in original wording using reporting and materials published by NVIDIA Developer Blog. Source reference: https://developer.nvidia.com/blog/how-nvidia-minerva-turns-telemetry-chaos-into-peak-gpu-yield/.

Drafted by the GenAI News review pipeline.