用于人工智能_机器学习数据中心电源基础设施的固态变压器.pdf

1、Solid-State-Transformers for AI/ML Datacenter Power InfrastructureDelta ElectronicsSolid-State-Transformers for AI/ML Datacenter Power InfrastructurePeter BarbosaVP,R&DDelta ElectronicsRACK&POWERSources:EDN,Google,Daelim,miSci19552024SAGE(Semi-Automatic Ground Environment)Single Computer:AN/FSQ7:60,

2、000 Vacuum TubesPower Consumption:3MW,cooling systemAI-Based Hyperscale DatacenterSingle GPU Chip:Blackwell 208B TransistorsPower Consumption/DC:100MW GW,Vacuum TubesBlackwell GPURubin Ultra Rack:600kW 100s Billion X more computing 100s X more power consumptionLoad Characteristic Changed:Power distr

3、ibution infrastructure remains nearly the same!Datacenter Evolution70 Years10 35 kVac480 VacDiesel GeneratorATS277/480VacSubstationTransformer50 VdcAI Server RacksxPU serverDC-DC0.8 VdcBBUnvidiaPower ShelfAC-DCSimplified Datacenter architecture with AC Distribution and 50V bus:0100200300400500600201

4、42016201820202022202420262028Electricity Consumption(TWh)Future Scenario Low(TWh)Future Scenario High(TWh)Total data center electricity use from 2014 to 2028Electricity Consumption(TWh)4.4%of US total12%6.7%Source:2024 Report on U.S.Data Center Energy UsePower Infrastructure&TrendsMore Electricity U

5、tilizationReduce grid dependencyMicrogrid on-site generation and energy storageRenewable resources(Wind,Solar,SMR)Better Electricity DistributionLimitation of the 50V rack busLimitation of existing 480VacHVDC distribution requiredThe scarcity of Cu and abundance of Si enabled trend in inverter price

6、s.SST utilizes more power electronics(Si)to replace the traditional Cu and Fe.SST vs Line Frequency Transformer(LFT)Power Electronics enables HF isolationOrders of magnitude smaller dimensionReduce the use of Cu and enclosure materialSources:Enchems,Statista,ONR,GEGrid-to-Chip Power Architectures400