NVIDIA launches Minerva, an AIOps platform for GPU 'AI factories'
AspireAI · NVIDIA
NVIDIA unveiled Minerva on October 22, 2026, an AIOps platform aimed at the growing problem that the bottleneck in a GPU data centre is no longer just the chips. Minerva pulls telemetry from GPUs, networking, storage, schedulers and the jobs running on them into a single view, then applies purpose-built AI models to spot problems earlier. Its stated target metric is 'AI factory yield' - useful work measured through GPU hours, goodput, cluster uptime and token production. At that scale one GPU running hot or one NVMe drive losing performance under load can quietly slow an entire training job, and the warning signs are easy to miss inside an alert storm. Minerva builds a dynamic knowledge graph linking hardware to the jobs on it, claims to analyse billions of data points per hour, and tries to collapse an event storm into one incident with a likely root cause. It ships with a natural-language assistant grounded in live telemetry, and remediation can go to a human or to automation. The honest caveat: NVIDIA cites no independent benchmark and gives neither pricing nor an availability date. IREN's CTO is quoted endorsing it, and IREN, IBM, Tesla, Jio, Bitdeer AI and others are named as planning deployments, with Dell, Lenovo and Supermicro to offer it in their AI factory operations portfolios.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.