面向人工智能工厂的人工智能原生网络.pdf

1、 2025 Juniper Networks 2025 Juniper NetworksAmit SanyalOct,2025AIAI-Native Networking for Native Networking for AI FactoriesAI Factories 2025 Juniper Networks 2025 Juniper Networks 2024 Juniper NetworksInternet/SP CoreDistributed AI clusters in todays Hybrid AI Cloud Distributed AI clusters in today

2、s Hybrid AI Cloud Driven by power,data gravity&privacy,latency,regulations&cost.Ent AI DCCo-Lo AI DCPrivate Cloud(Co-Lo or on-prem)Private WAN BackbonePublic/AI Cloud X Public/AI Cloud YTypical AI DCPrivate WAN BackboneStorageGPUsEdge AI DCsAI DC 2AI DC 1AI DC 3Region 1DCIRegion NAI FabricsGPUsNeeds

3、 High Performance AI Fabrics inside DC,DCI/DC Edge&Security 2025 Juniper NetworksNetworking plays a critical role in each phase of AI lifecycleNetworking plays a critical role in each phase of AI lifecycleFoundational Model TrainingInferenceModel Fine TuningModel Builders/NeoCloudsEnterpriseInferenc

4、eJob Completion TimeTime to First Token,Tokens/Sec The network plays a critical role in maximizing these metricsReasoning ModelsAgentic AIMixture of Experts 2025 Juniper NetworksAI workloads have unique requirementsAI workloads have unique requirements8x800G Backend GPU Network2x400G Frontend/Storag

5、e NetworkGeneral Purpose ServerNVIDIA DGX B300 Server2x25G50 Gbps per server7 Tbps per serverMassiveBandwidthNetworkRequirementsNumerous small,burst packetsTolerant of best-effort EthernetMinimal impact on CPU performanceFew,large RDMA flows(elephant flows)Requires ultra-low loss and latencyPacket l

6、oss and congestion impacts GPU performance negatively 2025 Juniper NetworksJuniper ConfidentialHPEHPE-Junipers Open&Validated AI DC SolutionJunipers Open&Validated AI DC SolutionOpen,Purpose-built Solution for Maximum Performance HPE-Juniper AI LabMLPerf SubmissionsBest Practices