将人工智能基础设施扩展到数据中心区域.pdf

1、Dan RabinovitsjVP,MetaScaling the AI Infrastructure toData Center RegionsServing 3.4B DAUsis a privilege,but presents significant challenges.WEBRAY-BAN METAMETA QUESTCHATVOICEIMAGINE YOURSELFIMAGE UNDERSTANDINGIMAGINE EDITAI is used across all our Apps Built with Llama Meta AI,launched 1.5 years ago

2、,is available across our apps and on the web,is more creative,smarterand used by 800M people every month.Source:Meta Q424 earnings,Jan 2025.24k GPUs100k GPUsLLAMA3LLAMA4The FutureGigawatt DatacentersGenAI Cluster Projectionsand NW BW push systemand cluster scaleGenAI Cluster Projections Expected To

3、OutgrowLarge-Pod FeasibilityMoE GrowthFuture GenAI Cluster Projections Expectedto Outgrow X-Large-Pod Feasibility AI SYSTEM INNOVATIONOur AI Journey -LLMsContent Evaluation,Image Recognition,Malicious Content ScreenShort videos drove requirements for AI Ranking and Recommendations128 GPUScale-OutClu

4、stersCOMPLEXITYGenAIAI is amazingand amazingly hard1282K4K8KData HallScale Cluster24KDC Scale Cluster2023Building out massive clustersto support our ambitions2024129KRegion Scale Cluster1282K4K8KData HallScale Cluster202324KDC Scale ClusterBuilding out massive clustersto support our ambitions1282K4K

5、8KData HallScale Cluster24KDC Scale Cluster20232024129KRegion Scale ClusterMulti-MillionMulti-Region Scale Cluster2025 and BeyondMassive and myriadof clusters to support our ambitions Large Scale Infrastructure Prometheus:1GW+cluster by 2026Hyperion:5GW over the next few yearsBeyond what weve seen b

6、eforeMTIA MI300XH200GB200GB300NetworkConnectivityDomainNote:Courtesy of Mark Papermaster,AMD CTO,slide from Keynote at ISC2025 Scale-ininterconnectScale-UpinterconnectScale-outinterconnect8x32x72x144xScale-Up SystemsScale-Up NetworkingNetwork BandwidthCompute Flopsexpertrouterinputsoutputsgatingweig