跳到正文
arXiv cs.CY· Yingbo Zhang, Eric Masanet, Shengwei Wang·· 23 小时前AI 评分34

AI 算力激增推动数据中心转向液冷:重新审视冷却能耗与水耗

AI Surge Accelerates the Transition to Liquid Cooling: Rethinking Energy and Water Use in Data Center Cooling

AI 导读

研究评估 AI 负载推动的液冷转型对全球能源与水需求的影响,液冷相比风冷可降低冷却能耗 25-56%。最佳实践 PUE 可达 1.07-1.08,新兴设计低至 1.03-1.04,WUE 为 0.38-2.63 L/kWh。模型还给出蒸发式系统 33°C、干式系统 49°C 的全球自然冷却进风温度阈值,并指出最低 PUE 不一定对应最低总能耗,因为冷却液温度升高会增加服务器功耗。

正文 · 原文

View PDF

Abstract:The power density of AI workloads is rapidly driving shifts to liquid cooling. This study rethinks and assesses the global impact on energy and water demand of this cooling transformation. Results suggest that liquid cooling can reduce cooling energy use by 25-56% relative to air cooling, offering substantial energy savings for GW-scale AI data centers. Best-practice PUE values could be as low as 1.07-1.08 for today's systems and 1.03-1.04 for emerging designs, and WUE values range from 0.38 to 2.63 L/kWh, depending on temperature conditions. We identify global free-cooling rack inlet temperature thresholds of 33 $^\circ$C for evaporative systems and 49 $^\circ$C for dry systems, demonstrating the practicality of 100% free cooling in many locations. The model finds that minimum PUEs do not necessarily coincide with minimum total energy use, because warmer coolant raises server power. Results quantify the large sustainability potential of liquid cooling and provide actionable targets for data center design and policies.
Subjects: Computers and Society (cs.CY)
Cite as: arXiv:2611.23903 [cs.CY]
  (or arXiv:2611.23903v1 [cs.CY] for this version)
  https://doi.org/10.48550/arXiv.2611.23903

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Yingbo Zhang [view email]
[v1] Wed, 18 Nov 2026 05:55:53 UTC (2,145 KB)

来源:arXiv cs.CY · arxiv.org