AMAZINGINDEX.COM 每天一点 Amazing
51.5
VOL. 2026.10
2026.10.30
← 返回 2026.10.30 日报
日报快照 · Daily Snapshot
NO. 023

NVL72 调度优化榨干 27% GPU 算力

#ARTICLE HackerNews 2026.10.30
推荐指数 40.0 NO. 023 · 2026.10.30
发布2026/10/29Score110Comments18

Grace Blackwell 的 NVL72 架构通过机架级 NVLink 互联打破了多节点通信瓶颈,但跨机架任务仍会回落到慢 43% 的 InfiniBand。新调度策略通过机架内集中部署多节点作业,将集群有效 GPU 利用率提升 27%。

这个 27% 的提升不是算法优化,而是拓扑感知的调度策略收益,说明当前主流调度器(如 Kubernetes 默认调度、Slurm)对 GPU 互联拓扑是盲视的。云厂商的弹性集群尤其吃亏——按需实例往往打散到不同机架,用户付了 NVL72 的钱却用着 InfiniBand 的性能。

如果你在用 AWS P5en 或 GCP A3 Ultra,需要立刻检查作业是否落在同一 NVLink domain,这比自己写 kernel 优化性价比更高。长期看,调度器会成为 AI 基础设施的新战场,类似 Run:AI、Vast.ai 这类工具的价值会被重估。

负面 17 条评论

核心争论:文章被指营销味重、缺乏技术深度,拓扑感知调度并非新技术

jhgg

This blog post spends a lot of time telling why gang scheduling is needed here (a fact obvious to anyone using the nvl72s), but doesn't actually go into detail about how their scheduler works. At work, we have built something very similar, a custom topology aware scheduler for workloads, and the mos

ismailmaj

Feels AI-generated, so much text real estate is spent explaining the basics in sufficient detail then just two paragraphs on their scheduler with a lot of hand-waving.

porridgeraisin

> making it more of a fluff/marketing piece than an actually interesting technical blog in this area. /genuine do you recommend any other blog?

替代方案: HTCondorKueueSkyPilot
查看原文 →