|
NVIDIA has released Minerva, software for the part of running an AI factory that comes after the racks are switched on. It follows factory-built DSX POD units, shown a day earlier, that NVIDIA says can cut deployment at a prepared site to as little as 14 days. Minerva is an AIOps platform that collects telemetry from GPUs, networking, storage, schedulers and workloads, then runs purpose-built AI models over that data to help operations teams find problems earlier.
NVIDIA launched it Thursday as part of its GTC Berlin announcements. In the blog post introducing Minerva, NVIDIA's Brian Sparks pointed to how small the warning signs can be, such as a GPU running hotter than its peers or an NVMe drive losing performance under load. At AI factory scale, one degraded component like that can slow an entire job, he wrote.
Traditional monitoring tools often split the data center into separate views, with networking, compute, site reliability and machine learning teams each watching their own piece, the post said. Minerva builds a dynamic knowledge graph instead, mapping GPUs, SuperNICs, switches and cables to the jobs running on them, so an operator can trace a slow training run back to the component behind it, even in a factory with hundreds of thousands of GPUs.
One incident instead of an alert storm
Minerva analyzes billions of data points per hour with supervised and unsupervised machine learning. It learns what normal behavior looks like, catches anomalies it hasn't seen before, predicts failures as they start to form and correlates related events using the factory's topology. When one fault sets off a burst of symptoms, the platform can group them into a single incident with a likely root cause instead of sending every symptom to an operator. |
