Scaling AI Infrastructure with Open Systems and Arm-Based Silicon.pdf

1、Scaling AI Infrastructure with Open Systems and Arm-Based SiliconEddie RamirezVice President of Marketing,Cloud AIArmAgentic explosion*Based on GitHub adoption trends20122026202420222020201820162014100k200kGitHub starsOpenClawKubernetesLinux20262026-20292031+500+110 Agents(per team/workflow)10000+So

2、urce:GartnerAgentic explosionAgents query 15x requests of humans AgentsAI data center CloudCPUCPUCPUCPUCPUCPUCPUCPUCPU orchestratesAccelerators generate tokensACCELERATORACCELERATORC P UAI agentAgentsCPU orchestratesAccelerators generate tokensACCELERATORACCELERATORC P UAI data center AnswerCloudCPU

3、CPUCPUCPUCPUCPUCPUCPUMassive amount of agent workloads swamp CPURequest queue28 waiting100%GPU 2WAITINGGPU 1WAITINGUtilization0%AI agentAnswer3-5X more CPUs needed to balance agent flow AgentsAI data center CloudCPUCPUCPUCPUCPUCPUCPUCPUCPUCPUCPUCPUCPUCPUCPUCPUCPU orchestratesAccelerators generate to

4、kensCPUACCELERATORACCELERATORACCELERATORACCELERATORCPUAI agentPerformanceScaleEfficiencyWorlds most efficient agentic CPUMemory tuned for compute6GB/s memory BW/core6TB per chip capacityUp to DDR5-8800World class Arm efficiencyIncredible 3nm efficiencyMaximum compute density300 watt TDPI/O for compo

5、sable AI systems96x lanes PCIe Gen6CXL 3.0 memory expansion and moreAMBA CHI extension linksLatency-optimized memory accessDual chiplet designMemory and I/O on same dieSub 100ns memory latencyUp to 136 Arm Neoverse V3 coresDedicated 2MB L2 cache per coreUp to 3.7Ghz frequencyResponsive performanceOp

6、en by design.Scaled by the ecosystem.1U customer reference serverOpen Rack V3Data Center Secure Control Module 2.0Data Center Modular Hardware System 1.1 RC2Foundation Chiplet System Architecture36kW Open Rack V330 x 2 node 1U servers8,160 performant CPU cores180TB+low latency me