AI网络定义AI超级工厂

AI Fabric Unlocks The Full Potential Of AI Factory NVIDIA NetworkQingchunSong The Data Center Is The Computer, The Network Defines The Data Center Spectrum-X Ethernet Brings High Performance to Ethernet For TrainingThe Network Defines The Data Center The New Inference Architecture-Distributed Inference Spectrum-X Ethernet Expands To Inference SpectrumX Ethernet Low Jitter Communications for AI Switch-To-SuperNIC, End-To-End Network Processing, Bringing High Performance To Ethernet 3X Higher Expert Dispatch PerformanceSpectrum-X Ethernet Load Balancing And Congestion Control Show 3X Higher Performance •ModernMoEapplications (training andinference) use asymmetric and stochasticExpert token dispatch •This pattern is largely exposed and cancreateextreme congestion that can causepacket drops and victim flows 42% Higher LLAMA3 70B Training In Multi-Tenant Data Center Spectrum-X Ethernet Isolates Multi-Tenant Application Noise, Higher Performance With Deterministic Results Spectrum-X Ethernet Multi-TenantOTS Ethernet Multi-Tenant What Makes Spectrum-X Ethernet Special Switch-to-SuperNIC, End-To-End Network Processing, Bringing High Performance To Ethernet Scale-Out And AI Density Depend On Optical ConnectivityThe Optical Network Power Consumption Represents 10% Of Compute Resources 100KServers 100KServers 2.3 MWTransceiver Power 40 MWTransceiver Power Introducing Spectrum-X Ethernet PhotonicsWorld’s First 200G/SerDes Co-Packaged Optics NVIDIA Photonics Solves Power & Reliability Challenges Of AI Scale-out NVIDIA Photonics Switch Systems Co-Packaged Optics Networking Switches To Scale AI factories To Millions Of GPUs 10XHigher Resiliency 3.5XPower Efficiency 1.3X Faster Time To Operation Scaling AI Beyond The Data Center Requires New Infrastructure Distributed AI Between Remote Locations, Overcoming Power And Physical Limitations 1.9X Higher NCCL Multi-Site Performance With Spectrum-XGS Ethernet Auto-Adjust Load Balancing And Topology-Aware Congestion Control Maximize NCCL Throughout Summary The Data Center Is The Computer, The Network Defines The AI Factory •The first AI Ethernet network-NVIDIA Spectrum-X Ethernet•Both training and Inference request the high performance network-Spectrum-X Ethernet•NVIDIA Spectrum-X powers the largest AI Factory in worldwide•NVIDIA CPO switches are the first production level data center CPO product, reduce power, higher resiliency•Spectrum-XGS Ethernet overcomes the power and physical limitations of AI