面向新时代的基础设施：为前沿规模的人工智能赋能.pdf

1、Saurabh DigheMicrosoftInfrastructure for a New Era:Powering AI at Frontier ScaleFaster than todays leading supercomputerThe worlds most powerful AI datacenterSquare feetBest performance at the lowest costPrompt price/1M tokens$0$10$20$30GPT-4May 2023GPT-4 TurboNov 2023GPT-4oJun 2024GPT-4oNov 2024GPT

2、-4.1Apr 202593%price reductionOptimizing across every layer of the stackOf ux psl johQpx f sDppm johTjm jdpo!boe!SbdlE D!Jog sbtusvduvsfOpen Letter to the Industry:on AI Data Center Infrastructure StandardsHXU Cooling Capacity IncreaseLegacy HXUNext-Gen HXUCurrent HXU202220242026vs.2022Liquid-coolin

3、g for AI systems in air-cooled datacentersUpcoming contribution to OCPRearchitecting datacenter power for the AI eraUPS(AC/AC)PDUs/Transformers(600VAC)Mt.Diablo:Next-generation datacenter power rack designOperates at 400VDC and 800VDC differentialfor efficient,high-capacity power deliveryScales from

4、 100s of kW to 1MW+Industry collaborationFuture innovations in power delivery:Solid State TransformersImprove ecosystem cost through innovation Leveraging the technology in other industriesHigh density power conversion for AI datacentersGrid integration and stabilization through energy storageFull-s

5、tack solutions for power stabilizationOf ux psl johEnough fiber inside the datacenter toBandwidth utilization is relatively modestHigher tolerance to latencySDN or BGP-level reroutingNeeds reordering buffersVariable packet sizeHigh bisectional bandwidth utilizationMore latency sensitiveTransport and

6、 application-level reroutingNatively supported(semantically)More homogenousAI Networking requirementsDm pve!E bub!Df ouf s!Of ux psl johScale UpXPUXPUXPUXPUPodScale UpScale OutXPUXPUXPUXPUPodScale UpXPUXPUXPUXPUPodScale UpXPUXPUXPUXPUPodScale UpScale OutAI-WAN