NVIDIA Minerva：用 AIOps 把遥测混乱转化为 GPU 产出
原标题：How NVIDIA Minerva Turns Telemetry Chaos Into Peak GPU Yield
AI 摘要
NVIDIA 发布技术博客介绍其 AI 原生 AIOps 平台 Minerva，用于解决 AI 工厂中 GPU 集群遥测数据分散、故障诊断困难的问题。Minerva 跨 GPU、Spectrum-X/NVLink 交换机、存储、调度器等采集指标与日志，通过自动发现构建动态知识图谱，将基础设施与运行中的训练任务关联，并用监督与无监督机器学习识别“灰色故障”。平台还提供 AI agent 助手和 MCP server，让运维人员以对话方式查询实时证据，目标是保住有效 GPU 小时、减少训练中断。
正文节选
AI Factories are some of the most technologically complex operational environments ever created. Hundreds of thousands of GPUs, NVIDIA Spectrum-X Ethernet scale-out, scale-across and scale-in fabrics, NVIDIA NVLink scale-up fabrics, distributed schedulers, and multi-week training runs operate in continuous lockstep.These layers function as one production system even when different teams own their telemetry. In synchronous distributed training, a job advances at the pace of its slowest rank. One