扩展数据中心基础设施.pdf

1、Ihab Tarazi Chief Technology Officer/SVPDell TechnologiesScaling Data Center InfrastructureDell OCP Contributions2015 2019 2021 2022 2023 2024 2025 SSD 2.6NIC 3.0NW Switch InterfaceDC-MHSM-XIOM-PICM-PESTIM-FLWM-CRPSLTPIDC-SCM 2.0Server Project Co-LeadProject LeadsSSD 1.0HDD DynamicsSSD 2.5HDD Dynami

2、csNIC 3.0 1.5NIC 3.0 1.6DC SAS-SATACooling LoopFault Mngmt Infra7 PSU SpecsMHS S-DNOOpen System for AINetworkingStorageComputeRackMHS Co-LeadCold Plate Co-LeadDC-SCM Co-LeadHPM Common Circuit Type 1 GPU and CPU Power TrendsCPUAI Cluster Growth20232024202520264K GPU Cluster4K GPU ClusterTotal Power 1

3、0MWTotal Data Hall Size 3500Sq.ft#of Optical Transceivers 7K#of Optical Fibers 10K18K GPU Cluster18K GPU ClusterTotal Power 40MWTotal Data Hall Size 10,000Sq.ft#of Optical Transceivers 30K#of Optical Fibers 65K82K GPU Cluster82K GPU ClusterTotal Power 162MWTotal Data Hall Size 42,000Sq.ft#of Optical

4、 Transceivers 125K#of Optical Fibers 285K110K GPU Cluster110K GPU ClusterTotal Power 220MWTotal Data Hall Size 55,000Sq.ft#of Optical Transceivers 165K#of Optical Fibers 380KOCP Enabled High Performance Cooling SystemsManifolds integrated into the rack:Blind-mate universal quick disconnects(UQDB06)H

5、igh flow rates to enable 480kW per rackPromotes gen-over-gen rack infrastructure re-use Optimized server cooling:Enabling 10 kW per OUImpedance-matched cold plate loops enable diverse liquid cooled platforms in the same ORv3 rack Including CPU and GPU servers,network switches and other equipment wit

6、h widely varying cooling requirementsServers blind mate to manifold via UQDBsEarly adoption of ORv3 native CDUs:Busbar powered CDUs Blind-mate power interface with ORv3 rack In-rack CDUDell Integrated Rack 7000Enhanced DLC blind mate manifolds10kW per OU Superior power whip routi