ACF-S：用于人工智能和加速计算网络中高性能数据移动的8-Tbit每秒的超级网卡.pdf

1、 2024 ENFABRICA CORPORATION.ALL RIGHTS RESERVED.UNLEASH THE REVOLUTION IN NEXT-GEN COMPUTING ACF-S:An 8-Terabit/sec SuperNIC for High-Performance Data Movement in AI&Accelerated Compute NetworksHot Chips 2024August 27,2024Shrijeet Mukherjee,EnfabricaThomas Norrie*,OpenAI 2024 ENFABRICA CORPORATION.A

2、LL RIGHTS RESERVED.*work previously done at Enfabrica 2024 ENFABRICA CORPORATION.ALL RIGHTS RESERVED.3:missionredefine networking for distributed accelerated computing to deliver peak performance,resiliency and node scale:teamstarted 2020120+engineerspreviously built high-performance NICs,switches/r

3、outers,TPUs,graphics,host networking stacks:productaccelerated compute fabric superNIC(ACF-S)1st chip codename millennium 8Tbps bandwidth:scale-up supercomputing /mainframe,ccNUMA Fully coherent memory system operating on a“large”problem by sharding computation CPUs synchronize state and move memory

4、 closer using IPC transactions with latencies in nanoseconds Communication protocols deeply embedded in the processor to enable“transparent”communication 2024 ENFABRICA CORPORATION.ALL RIGHTS RESERVED.4All blue links are IPC communicationCPUCPUCPUCPUCPUCPUCPUCPU:borg /the rise of scale-out computing

5、 Client-server design,built for extreme,resilient application scaling All communication uses retargetable,resilient software managed RPCs Workers and data pipelines are imminently reconfigurable Heterogenous elements with high aggregate bandwidth needs and high tolerance to latency(microseconds to m

6、illiseconds)2024 ENFABRICA CORPORATION.ALL RIGHTS RESERVED.5All green links are sharded RPC communicationLoad Balancer/RequestorSharded/ReplicatedWorkersSharded/ReplicatedWorkersCPUCPUCPUCPUCPUCPU:hyperscale AI/ML systems /super,meet borg A modern,truly scalable solution demands:Tight performance of