Nvidia announced Minerva on 22 October, an AI operations platform that brings hardware and workload data together to help operators find problems in AI factories. The company says a single degraded component can slow an entire training or inference job.
Minerva maps GPUs, network equipment, storage and the jobs running across them in a knowledge graph — a connected record of which components each job depends on. Nvidia says its supervised and unsupervised machine-learning systems analyse billions of data points per hour to spot unusual behaviour, anticipate failures and group related alerts around a likely cause. Operators can also investigate cluster health through an AI assistant and choose whether to implement a recommended fix themselves.
Getting an answer from an AI application could take longer when a degraded component slows the job serving it. Connecting that job to the affected hardware would give operators a way to investigate where the slowdown began.
IREN is using Minerva, according to Nvidia. Denis Skrinnikoff, IREN’s chief technology officer, said it gives his operations team one view of the network, GPUs and jobs running on them.
