At its conference in Berlin, Nvidia unveiled the Minerva software platform designed to predict AI server failures by mapping a dynamic knowledge graph linking hundreds of thousands of processors and their ongoing training tasks.
Nvidia launches Minerva platform at GTC Berlin
Unified knowledge map for hyperscale data centers
Early performance results and Vera Rubin platform integration
Early customers and distribution partners in AI factories
Frequently asked questions
Nvidia launches Minerva platform at GTC Berlin
Nvidia on Thursday launched a new software platform named “Minerva,” specifically designed to detect and monitor parts and components that begin to fail or degrade within massive GPU clusters before they slow down workflows or disrupt computational tasks. The company says Minerva merges telemetry data streaming from the infrastructure and AI workloads together, then runs purpose-built AI models on top of that data to provide data center operators with accurate early warnings and alerts.
The announcement of the new platform came during Nvidia’s GTC Berlin conference, which ran from October 20 to 22. Minerva reinforces Nvidia’s strategic push to sell integrated software and service suites that go beyond simply supplying electronic chips to companies building massive AI data centers around the world.
Unified knowledge map for hyperscale data centers
Nvidia explains that in data centers of this massive scale, a single declining component can slow down an entire AI model training task. An early warning sign might be a GPU running hotter than neighboring processors, or a high-speed NVMe storage drive losing efficiency under heavy operational load. The problem is compounded because traditional monitoring tools typically separate networking, compute, site reliability, and machine learning engineering teams.
To address this fragmentation, Minerva pulls operational signals and data from GPUs, networking equipment, storage units, job schedulers, and workloads, merging them all into a single dynamic knowledge graph. This graph links hardware components such as SuperNIC network cards, switches, and cables directly to the software tasks running on them, enabling operators to trace any slow task down to the specific hardware component causing the delay, even across clusters containing hundreds of thousands of GPUs.
Using supervised and unsupervised machine learning methods, the platform analyzes billions of data points per hour. It learns the normal operational behavior of the system, predicts emerging failures before they happen, and consolidates related alert storms into a single clear incident with its probable root cause identified. Additionally, an AI assistant built into the platform answers operators’ questions about the health of the compute cluster in simplified natural language and suggests appropriate fixes, with technical teams able to apply those fixes manually or connect Minerva to their own automation systems.
Early performance results and Vera Rubin platform integration
Nvidia reported that early deployments of the Minerva platform showed tangible improvements in operational efficiency, representing a 20% increase in mean time between failures, alongside a significant 63% reduction in mean time to repair.
The company also recorded a reduction in false or repetitive alert noise by up to 99.9%, and a 2% gain in actual productive GPU operating hours. The official product page describes Minerva as an integral part of the Vera Rubin advanced computing architecture.
Early customers and distribution partners in AI factories
IREN has already begun using the Minerva platform in its operations. Denis Skrynikov, chief technology officer at IREN, said: “Nvidia Minerva gives our operations team a unified, comprehensive view across the network, GPUs, and running jobs, allowing us to focus more on proactive maintenance rather than wasting time troubleshooting.” Alongside IREN, companies like Olani, Bitdeer AI, IBM, Geo, Lambda, Quick AI, and Tesla plan to deploy the platform in their centers. Under a deal struck last May, Nvidia has the right to invest up to $2.1 billion in IREN, which previously operated in Bitcoin mining.
On the commercial distribution front, global server manufacturers Dell Technologies, Lenovo, and Supermicro will offer Minerva as part of their AI factory operations portfolios. Lambda stated in a blog post on Wednesday that its users’ primary goal is to achieve “fewer workload interruptions, faster repairs, higher net productivity, and more code tokens generated.” Nvidia has not yet disclosed pricing details for the new platform.
Frequently asked questions
Question: What is the function of Nvidia’s new Minerva platform?
Answer: Minerva is an AI-powered software platform for monitoring degraded parts and components in massive GPU clusters and predicting failures before they slow down AI training and inference tasks.
Question: What are the most prominent digital results achieved by the platform in early trials?
Answer: The platform achieved a 20% increase in mean time between failures, a 63% reduction in mean time to repair, alert noise reduction of up to 99.9%, and a 2% increase in effective GPU operating hours.
Question: How does Minerva address the problem of fragmented monitoring teams in data centers?
Answer: The platform collects signals from processors, networks, storage, and job schedulers into a unified dynamic knowledge graph that analyzes billions of data points every hour and links each software task to its running hardware components.
Question: Who are the most prominent early customers and partners for Minerva?
Answer: IREN is currently using the platform, and IBM, Tesla, Lambda, Geo, Bitdeer AI, Olani, and Quick AI plan to deploy it, while Dell, Lenovo, and Supermicro will provide it within their solutions.
