Nvidia launches Minerva to catch failing parts in giant AI data centres
Nvidia has launched Minerva, software that uses AI to monitor large AI data centres. It combines information from graphics chips (GPUs), networks, storage devices and computing jobs to spot parts that may fail before they cause problems. It groups related warnings into one incident and includes an AI assistant. IREN already uses it, while Lambda, IBM and Tesla plan to use it.
Nvidia has launched Minerva, software that watches the huge groups of GPUs it calls AI factories and warns about problems before they slow down work, the company said on Thursday. Nvidia announced the software at its GTC Berlin conference, just days after it signed an AI partnership with Germany.
In clusters this large, one faulty part can delay an entire AI training job, Nvidia says. For example, one GPU might run hotter than the others, or a storage drive might become slow when it is busy. These warning signs can be hard to spot because network, computing and machine learning teams often monitor different parts of the system separately.
Minerva brings information from GPUs, network equipment, storage devices, job scheduling systems and AI workloads together in one map of the data centre. Nvidia says this helps operators find the part causing a slow job, even in systems with hundreds of thousands of GPUs.
The software uses machine learning to learn what normal operation looks like and predict when parts might fail. Rather than sending operators a separate warning for every problem, it combines related warnings into one incident and suggests what may have caused it.
An AI assistant built into Minerva lets staff ask questions about the health of their GPU clusters using everyday language. It also suggests ways to fix problems. Teams can carry out these fixes themselves or connect Minerva to their own automated tools.
“NVIDIA Minerva gives our operations team one view across the network, the GPUs and the jobs running on them,” said Denis Skrinnikoff, chief technology officer of IREN. Under an agreement made in May, Nvidia can invest up to $2.1B in IREN, which previously focused on Bitcoin mining.
Nvidia said that Bitdeer AI, IBM, Jio, Lambda and Tesla also plan to use Minerva. Dell, Lenovo and Supermicro will make the software available to their customers.
The launch is part of Nvidia's effort to sell more than just chips to companies building AI data centres. On Wednesday, the company said European partners were adopting its DSX blueprint to reduce water and electricity use. It also released a design for running AI agents safely. Nvidia did not announce a price for Minerva.