6608 - Learnings from AI SI Workshops and Next Steps.pdf

《6608 - Learnings from AI SI Workshops and Next Steps.pdf》由会员分享，可在线阅读，更多相关《6608 - Learnings from AI SI Workshops and Next Steps.pdf（8页珍藏版）》请在三个皮匠报告上搜索。

1、Cliff Grossner,Ph.D.-OCPOCP AI Strategic Initiative AI/ML Physical Infra Workshop 1OCP AI Strategic Initiative AI/ML Physical Infra Workshop 1Cliff Grossner Ph.D.,OCPArtificial Intelligence(AI)Thanks to all the ContributorsSession 1:Rack&Power Enabling xxx kW-1 MW racksSession 2:Cooling Enabling xxx

2、 kW-1 MW racksSession 3:DCF Enabling xxx kW-1 MW racksRack&Power Larger,wider racks need Pods vs racks(CPU,GPU)?Large number of cables(Rack,Side Car Rack)Must consider optical impact on rack power Need for robust telemetry like for power management of CPUs GFCI(Ground Fault Interrupt)feature standar

3、dization?ORV based,HPR Next HPR V3 250 kW roofline power,50 VDC to ITE rack HPR V4 Side Car AC-to-DC(PSU+BBU)HPR V4 can support 1.5 MW/rack(21”wide),4 Side Cars supporting 1 ITE Rack if Supercapacitor,no Supercap 2 Sidecars to support 1.6 MW/rack for ITE Mount Diablo+/-400 VDC Power rack to supply h

4、igher voltage power to the ITE rack Larger racks,large trays and more trays,more accelerators More bandwidth very challenging,to stay in copper More liquid flow rate,lower temperaturesCooling 250 kW,500 kW,1 MW rack cooling,greater liquid to air load ratios Need to consider higher water flow rates a

5、t the higher power racks Pressure drops approaching 50 PSI dP at 1 MW rack?2000 LPM for 1 MW rack(0.5 GPM/kW)?Will need Racks per CDU guidelines Number of Idle to full power(TDP)cycles guidelines for reliability Design for more quick disconnects and different sizes of couplings?Must design for fungi

6、bility,Quality,Reliability Impact on colocation DCs important,what are the requirements Standardizing optical interconnect cold plates to fit inside the cages Prepare for lidless bare-die packages for higher power chips Must consider System management protocol