Rohan Paul@rohanpaul_ai

2026-07-05 07:01·1天前

AI 摘要

Nvidia、Oracle 等合著新论文提出“功率灵活的 AI 数据中心”范式，允许数据中心在电网压力期间快速降低功耗，同时保护重要任务。系统将电网信号与 AI 任务调度、GPU 功率限制及实时功耗测量联动。在真实 130kW GPU 集群测试中，成功满足 200+ 个功率目标，40 秒内将功率降低约 30%，并维持低功耗数小时。系统还能跟随碳信号调度，并在区域受限时将推理流量从弗吉尼亚转移到伊利诺伊。论文指出训练、批量推理等任务可延迟或迁移，改变电网对 AI 数据中心的传统认知。

New Nvidia+Oracle and others paper says data centers can cut power fast and still protect important jobs during grid stress.

i.e. data center power can be scheduled much like computing work.

Changes the way AI data centers can be viewed by power grids.

Instead of treating them as giant machines that always need full power， it shows they can lower， delay， or move some computing work when the grid is under stress.

The problem is that grid planners often treat large AI data centers as fixed loads that always need full power.

This work shows that many AI tasks are more flexible than that because training， batch inference， and lower-priority jobs can slow down， wait， or move.

The system connects grid signals to AI job scheduling， GPU power limits， and live power measurements， so the cluster can follow requested power targets.

In a real 130 kW GPU cluster， it met 200+ power targets and reduced power by about 30% within 40 seconds.

It also held lower power for hours， followed carbon signals， and shifted inference traffic from Virginia to Illinois when one region was constrained.

----

arxiv. org/abs/2606.25098

Title： "Power-Flexible AI Data Centers： A New Paradigm for Grid-Responsive Compute"

Rohan Paul@rohanpaul_ai · X

30导出 Markdown