Meta 主讲的“人工智能网络：规模化、扩展和面向未来”研讨会.pdf

1、Role of Network in AI InfrastructureWEBSERVERCACHEPYMLNEWSFEEDADSCOEFFICIENTLASERSEARCHEVERSTOREPTAILSCRIBEDBEvery arrow goes over the networkA global network of fiber to bring AI to the world20242024129KMulti-building scale cluster129KMulti-building scale cluster1281282K2K4K4K8K8K2023202324KBuildin

2、g-scale cluster24KBuilding-scale clusterMulti-millionMulti-region scale clustersMulti-millionMulti-region scale clustersPrometheus:1GW+cluster in 2026Prometheus:1GW+cluster in 2026Increasing cluster size129KMulti-building Scale Cluster129KMulti-building Scale Cluster1281282K2K4K4K8K8K24KBuilding-Sca

3、le Cluster24KBuilding-Scale ClusterEven bigger haystacks of network componentsIncreasing cluster sizeLots of networking in the racksThe Role of Network:1.Enable the entire AI infrastructure stack1.Provide optionalityScaling out to gigawatts129KMulti-building Scale Cluster129KMulti-building Scale Clu

4、ster1281282K2K4K4K8K8K24KBuilding-Scale Cluster24KBuilding-Scale ClusterScale-out networks for large clustersPrometheus:1GW+cluster in 2026Prometheus:1GW+cluster in 2026Hyperion:5GW over the next few yearsHyperion:5GW over the next few years growing to gigawatt-scale Took“intra-chassis”networking an

5、d disaggregated across a data-hall for 4K GPUs Heavily tuned for AI Flexibly supports multiple generations and types of accelerators and NICs Took“intra-chassis”networking and disaggregated across a data-hall for 4K GPUs Heavily tuned for AI Flexibly supports multiple generations and types of accele

6、rators and NICsLast year:Disaggregated Scheduled Fabric Expanding to building-size:4K=20K GPUs Still flexibly supports multiple types of accelerators and NICs Can be combined for even larger clusters Expanding to building-size:4K=20K GPUs Still flexibly supports multiple types of accelerators and NI