DC25_SLIDES_Track02_PanelPoweringTheAISemiconductorStack_Zou.pdf

1、Information Classification:GeneralPowering the AI Semiconductor Stack Power Strategies for High-Performance AI ComputingModerated byMobashar Yazdani1Information Classification:GeneralImagePanelistShenli ZouPower Engineer,Meta Platforms,IShenli works as power engineer at Meta,supporting in-house AI A

2、SIC(MTIA)module power delivery2Information Classification:General Power Trend Driven by:oBigger scale-up domain to support AI workloadoThe need to keep the scale-up domain within copper cable reach to reduce unavailability in comparison to connecting multiple racks with opticsoCorresponding system o

3、verhead power increase oSingle GPU/accelerator power increase for higher performance AI Power Trend3Figure referenceInformation Classification:General Emerging Areas:oHVDC IT rack ecosystemHigh voltage busbar/connector/cable,PSU/HSC/dc-dc converters/passivesSafety,reliability,complianceoHigher power

4、 delivery technology to ChipVertical power delivery(VPD)module,iVR,embedded passivesoAdvanced cooling technologies Industry Collaborations:oComponent availabilityoStandardization in design and qualificationoReduced development cycle for customized power solutionsCall for Collaborations4Time-to-productionStandardization Customization