当前位置:首页 > 报告详情

人工智能机架和集群的冷却解决方案.pdf

上传人: 明**** 编号:1011468 2025-12-21 15页 1.95MB

1、Cooling Solutions for AI Racks and ClustersFoxconn/IngrasysCooling Solutions for AI Racks and ClustersChris PaiEngineering VP/IngrasysAI CLUSTERSData Center Power and Cooling Solution TrendCRPS+Air Cooled2000 2000-201020101990 1990-200020002010 2010-201520151-3kW5-10kW8-20kW2026 2026 300+kWPower Rac

2、k+Fully Liquid CooledPower Shelf+Liquid/Air Hybrid Cooled2016 2016-202020202021 2021-202320232024 2024-2025202550+kW120+kW30kWData Center AI SuperPODsPUE:1.2132 KW/rackLiquid-cooledPUE:1.528 KW/rackAir-cooled Superior thermal efficiency Lower power usage effectiveness of datacenter Higher cooling ca

3、pacityper rack Reduced infrastructure foodprintWhy Liquid Cooling?CDU(L2L)CDU(L2A)Types of Liquid Cooling DLC(Direct Liquid-Cooled)Comparison of Cooling MethodsLiquid to AirNo need facility waterLow cooling capabilitySidecar(Liquid to Air)IT RackSidecar(Liquid to Air)IT RackSidecar(Liquid to Air)IT

4、RackLiquid to LiquidFacility water requiredHigh cooling capabilityIn-row CDU(Liquid to Liquid)IT RackTypes of Liquid Cooling Immersion CoolingSingle PhaseTwo PhasePUE:1.03PUE:1.01 Piping and infrastructure layout Temperature control and regulation Scalability and flexibility considerationsTypical Li

5、quid Cooling Datacenter Layout45-55deg C25-35deg C35-45deg CHeat Rejection(Cooling Tower/Dry Cooler)20-30deg CBuilding Blocks of AI ClusterGPU ClusterNetworkingControl and Management NodesStorageFull Liquid-cooled GPU ClusterCooling BankAI Rack with Power SidecarDiabloDiabloDiabloKey Considerations

6、of The Three Building BlocksRack Architecture for Next Generation AI ServerHigh Density AI RackPower SidecarCooling Bank High Power Consumption Full Liquid-cooled(Single Phase DLC and Two Phase DLC)High Density Good Serviceability High Power(Higher Power Capac

word格式文档无特别注明外均可编辑修改,预览文件经过压缩,下载原文更清晰!
三个皮匠报告文库所有资源均是客户上传分享,仅供网友学习交流,未经上传用户书面授权,请勿作商用。
根据报告的内容,全文主要概括了数据中心AI集群的冷却解决方案趋势。以下是关键点: 1. 冷却解决方案趋势: - 1990-2000年:1-3kW - 2000-2010年:5-10kW - 2010-2015年:8-20kW - 2026-2023年:2000-300+kW 2. 数据中心AI超级POD: - 液冷PUE:1.21 - 空冷PUE:1.52 3. 液冷优势: - 高热效率 - 低PUE - 高冷却能力 - 减少基础设施 4. 液冷类型: - 直接液冷(DLC) - 液气混合冷却 - 液液冷却 - 沉浸式冷却 5. AI集群构建块: - GPU集群网络 - 控制和管理节点 - 存储 - 全液冷GPU集群 6. 高密度AI服务器: - 高功率消耗 - 全液冷 - 高密度 - 高功率管理 - 液冷母线 - 高效电力分配 7. 未来趋势: - 提高功率密度 - 先进冷却技术 - 可持续性和环保实践
数据中心节能秘籍?" 液冷与空气冷却大对决!" 液冷技术引领潮流?"
客服
商务合作
小程序
服务号
折叠