借助思科人工智能网络将部署挑战转化为机遇.pdf

1、Transforming deployment challenges into opportunitiesOctober 2025Murali GandluruVP,Product Management 2025 Cisco and/or its affiliates.All rights reserved.Cisco ConfidentialTypical AI Network ArchitectureFrontend NetworkStorageComputeGPUDPUBackend NetworkHigh-ThroughputLow JitterLow-LatencyLosslessD

2、ata CenterColocation10G25G50G100G400G800GServersBut there are challenges when building the AI networking infrastructureNetworkingKeeping latency low when GPUs need to talk to each other at massive scaleSecurity&IdentityEliminate eastwest security blind spots for high-bandwidth AI flows,and provide d

3、eep application and identity awarenessAutomation&VisibilityAutomating and seeing whats really happening in networks that change as fast as AI workloads 2025 Cisco and/or its affiliates.All rights reserved.Challenge#1:Traditional Networking approachesUnreliable Data DeliveryNetwork DelayNetwork Conge

4、stionNon-uniform link utilizationPoor Burst absorption 2025 Cisco and/or its affiliates.All rights reserved.Cisco ConfidentialTransforming AI Networks for Performance and ScaleGPU NodesInter-GPU Backend NetworkManagement networkStorage networkFrontend networkAdvanced Load BalancingCongestion Managem

5、ent through TelemetryAutonomous RecoveryLow Job Completion Time and High Performance 2025 Cisco and/or its affiliates.All rights reserved.Node-1Node-32Hybrid CPU/GPU Workloads:Dynamic Paths,Deterministic ResultsFlexible policy-driven load balancing:per flowlet,per-packet,a mix of both,or dynamically

6、 switch betweenSpineLeafServerFlow2*Flow3XFlow1*Flow4*DLB enabled flowXQoS Policy flowConverged Physical NetworkStorage server 2025 Cisco and/or its affiliates.All rights reserved.Challenge#2:Automation&VisibilityFragmented&complex automationLacks Unified Fabric ViewNo unified life cycle managementI