How NVIDIA Minerva Turns Telemetry Chaos Into Peak GPU Yield

NVIDIA is addressing the critical operational challenges of large-scale AI infrastructure with Minerva, an AIOps platform designed to maximize GPU yield. As AI factories scale to hundreds of thousands of GPUs, operators face an increase in subtle hardware and network issues, known as “gray failures,” which degrade performance without triggering traditional alarms. These issues, such as a single underperforming network link or a slightly overheating GPU, can slow down entire distributed training jobs, wasting expensive compute hours. Minerva aims to solve this by unifying telemetry data from across the stack into a cohesive, real-time model of the entire AI factory.

Under the Hood: A Knowledge Graph for AI Factories

At its core, NVIDIA Minerva ingests metrics, logs, and topology data from GPUs, switches, and schedulers into a dynamic knowledge graph. This graph maps the relationships between physical hardware and active workloads, providing critical context that is often missing in siloed monitoring tools. By applying purpose-built machine learning models, the platform moves beyond static thresholds to detect anomalies through peer comparison—identifying, for instance, a GPU that is running hotter than its peers under an identical load. Its topology-aware correlation engine can collapse thousands of related alerts into a single, actionable incident. In one case, 81 separate alerts from a network pause-frame storm were consolidated into one incident, saving an estimated 54 hours of cluster uptime.

• Dynamic Knowledge Graph: Maps infrastructure to active jobs in real-time.
• ML-Powered Anomaly Detection: Uses supervised and unsupervised models to identify gray failures that rule-based systems miss.
• Topology-Aware Correlation: Reduces alert storms by grouping dependent symptoms into single incidents.
• Agentic Assistant: An AI agent named Sherlock allows operators to query system status and evidence in natural language.

The Impact: Proactive Maintenance and Increased Yield

The platform’s real-world impact stems from its ability to shift operations from a reactive to a proactive model. By detecting rare issues like a SuperNIC PCIe fault or a single degrading NVLink interface among 400,000, Minerva allows operators to drain and repair affected nodes before they impact customer workloads. This proactive stance has been shown to prevent hundreds to thousands of wasted GPU hours. For the broader market, this represents a crucial step in maturing the operational discipline required to run AI factories efficiently. By improving reliability and simplifying diagnostics through its AI agent, Sherlock, NVIDIA is providing the tooling necessary to maximize the return on massive investments in accelerated computing hardware.

Minerva is NVIDIA's answer to a fundamental challenge of scaling AI: system reliability is as important as raw performance. By bundling advanced, topology-aware AIOps with its hardware ecosystem, the company is solidifying its position by ensuring customers can actually achieve the massive utilization required to justify their capital expenditure.