面向大规模智能计算的集合通信与在网计算研究综述OA

In recent years,artificial intelligence,represented by large-scale models,has developed rapidly.However,the severe mismatch between the growth rates of accelerator computing power and network bandwidth has confronted large-scale AI computing with a formidable"communication wall"challenge.Furthermore,the inherent heavy-flow characteristics of AI workloads exhibit a deep-seated discrepancy with traditional network protocols,rendering conventional single-node power stacking inca-pable of maintaining linear system scalability.In response,collective communication mechanisms have been introduced to align predictable data interactions with physical topologies,aiming for theoretically optimal end-to-end bandwidth utilization.Nevertheless,when confronted with massive and frequent syn-chronization demands,traditional host-based collective communication is often constrained by the passive forwarding overhead of network devices.Consequently,collective communication urgently requires un-derlying hardware acceleration support from in-network computing technologies.This paper systemati-cally reviews the evolution over the past decade of collective communication and in-network computing for large-scale AI computing.It concludes that future AI computing will evolve toward memory-centric architectures,all-optical interconnect networks,and ultra Ethernet consortium standards,where the deep integration of computing and networking will serve as the cornerstone of the artificial general intelligence era.