Nvidia launches Minerva to monitor massive AI factory hardware

The complexity of AI factories

Large-scale AI factories are becoming massive industrial environments. These facilities contain hundreds of thousands of graphic processing units (GPUs) linked together by high-speed networks to train models and serve inference. Because these systems are so interconnected, a single degraded component can cause significant problems for the entire operation.

For example, a graphics card operating at an elevated temperature or a disk drive experiencing a drop in data speed can hinder a training task. Traditionally, different teams have had to manage different parts of the infrastructure, often leaving them with incomplete views of the cluster's health. This fragmentation makes it difficult to see how a small hardware glitch ripples through the whole system.

Nvidia launched Minerva to solve this by providing end-to-end observability. The platform sits within the full-stack Nvidia Vera Rubin platform. It gathers telemetry from a wide range of hardware, including GPUs, Spectrum-X Ethernet, NVLink switches, SuperNICs, and data processing units.

Mapping the hardware web

At the core of Minerva is a dynamic knowledge graph. This technology maps the logical relationships between all the different elements of the infrastructure. It connects the physical hardware, such as switches and cables, to the specific GPUs and the workloads running on them. This unified topology is what allows operators to see the big picture.

By understanding these connections, the platform helps teams identify the exact piece of equipment responsible for a lagging training process. In a cluster with a vast number of GPUs, finding one faulty cable or switch manually would be nearly impossible. Minerva automates this discovery process by linking the physical layout with the digital workload.

The system also integrates with common schedulers like Slurm and Nvidia Run:ai. By pulling data from these sources alongside hardware telemetry, Minerva creates a complete map of the factory. This ensures that operators aren't just looking at isolated parts, but at the entire flow of data and compute power.

Turning noise into action

Minerva processes billions of telemetry data points every hour. To make sense of this volume, it uses both supervised and unsupervised machine learning. These models learn what normal operating patterns look like for a specific factory and flag any unusual behavior that deviates from the norm.

One of the biggest challenges in large-scale computing is alert fatigue. Instead of forwarding every individual symptom to a human, Minerva consolidates cascading alerts into focused incidents. It identifies the probable root cause of a problem so that engineers don't have to sift through thousands of minor notifications to find the real issue.

The platform is designed to provide three specific operational outcomes. It aims to extend the time between workload interruptions through predictive maintenance. It also seeks to decrease the duration of fixes by quickly pinpointing errors and filtering out unnecessary signals so that only relevant alerts reach the human operators.

AI-assisted maintenance

Nvidia has built a native AI assistant directly into the Minerva platform. This provides a natural-language interface grounded in live telemetry data. Operators can ask questions about the health of the cluster and receive specific recommendations on how to fix identified problems.

The platform also allows for connections to external agentic AI tools. This means the system can be integrated into broader automation workflows. While the AI provides the recommendations, teams still have the choice to apply those fixes manually or route them through infrastructure APIs for automated execution.

This approach moves the operational model from reactive troubleshooting to proactive maintenance. By using AI to monitor the hardware, companies can focus on keeping the GPUs productive. This visibility is intended to strengthen the operating model for the next generation of infrastructure being built by major tech firms.

Industry adoption and scale

Several major companies have already confirmed plans to deploy Minerva. This includes high-profile names such as Tesla, IBM, Jio, Lambda, Quake AI, and Bitdeer AI. These organizations are looking to maximize the productivity of their GPUs as they scale their respective AI factories.

The platform is also being integrated into the portfolios of major hardware providers. Dell Technologies, Lenovo, and Supermicro will incorporate Minerva into their AI factory operations. This move extends the platform's availability across broader enterprise infrastructure channels, making it accessible to more organizations.

Nvidia measures the success of these factories using specific metrics like GPU hours, goodput, cluster uptime, and token production. By providing a clearer view of these metrics, Minerva helps companies ensure their expensive hardware is running as efficiently as possible. The goal is to ensure that the underlying infrastructure can keep up with the demands of modern AI development.