These case studies show how NVIDIA Minerva anomaly detection and correlation found the cause of infrastructure problems in AI factory clusters, in some cases before they affected workloads.
-
Pause Frame Anomaly Correlation: A switch page fault caused a restart and network congestion. NVIDIA Minerva correlated 81 anomalies into one incident and identified the root cause.
-
Performance Degradation: Jobs slowed down because adaptive routing hardware resources were exhausted. NVIDIA Minerva identified the cause, which shortened the investigation by about one to two weeks.
-
Predictive Maintenance: Log analysis detected a ConnectX NIC PCI error, so operators could remove the node from the allocation pool before a job failed.
-
Switch Upgrade Failure: During an in-service software upgrade (ISSU), NVIDIA Minerva detected a rare switch control-plane failure, so operators could stop the maintenance window before any jobs were affected.
Last updated: