当前位置:首页 > 报告详情

20260219_K-6_Yusuf.PDF

上传人: 彩旗 编号:1158852 2026-03-02 19页 2.33MB

1、1Arms Open Chiplet Ecosystem for the Converged AI DatacenterArm KeynoteDirector Hardware Ecosystem,Cloud AIImran Yusuf20262AI is Redefining ComputeScaling of AI ModelsIncreased Demand for AI ApplicationsAgentic/On-Device AIDesign Complexity&CostPower&Energy ConstraintsSupply Chain Management&Operati

2、onsCUSTOM SILICON CustomN E T W O R KS T O R A G E&I/OM E M O R YC O M P U T EA C C E L E R A T I O NCustomCustomCustomCustomConstraint-specific optimizationCo-designed for max performance per wattDeep feature integration$249$449$581$725 7nm5nm3nm2nmSource:IBS Global Semiconductor Industry Service R

3、ising Costs and Complexity of Advanced SoC Designs7NM TO 2NM ADVANCED DESIGN NODE COST:3x growth24-36 monthsAvg development cycle for SoCsSource:EE Times 1nm$1BsProjectedCUSTOMIZATIONSoCArchCSSIPCOST,COMPLEXITY&TIME(log)CUSTOMIZATIONSoCCSSIPArchCOST,COMPLEXITY&TIME(log)CUSTOMIZATIONSoCCSSIPArchChipl

4、etEcosystemStandardsDesign Methodology8D E S I G N S E R V I C E S&PA C K A G I N G F O U N D R Y S E R V I C E SF I R M W A R E,E D A&3 R D PA R T Y I PC O M PA N I O N S I L I C O N&O E M sHow can chiplets help solve the AI compute problem?Tighter IntegrationReduce RiskLeverage Investment10Chiplet

5、s Connect Silicon Modularity to AI System DeliveryFirmwareComputeStandardsIOMEMORYSYSTEMSECURITYOPTIMIZED ARM CPU CHIPLETSINTEGRATED AND VALIDATED SUBSYSTEMS ToFromArchitectureSoftware EnablementKernel librariesSubsystems*CSA,AMBA,UCIeInstruction SetBIOS,BMCAI applicationSystem Demands Feedback Loop

6、AI is the full stackPlatform11The Value of Arm Neoverse Compute Subsystems(CSS)CSS delivers significant customer value beyond existing licensing models IP DevelopmentCompute SubsystemTop-Level SoCBackEndSoftwareArchCPUCMNSystemPOP/RFMArchIPConfigPerfRTLVerify/SBSAFPGAImageSoCArch3PIP1Config3PIP Perf

word格式文档无特别注明外均可编辑修改,预览文件经过压缩,下载原文更清晰!
三个皮匠报告文库所有资源均是客户上传分享,仅供网友学习交流,未经上传用户书面授权,请勿作商用。
1. **AI算力挑战**:AI模型规模扩大、应用需求激增,但先进SoC设计(7nm至2nm)成本增长约3倍,开发周期长达24-36个月,1nm节点预计超10亿美元。 2. **Chiplet解决方案**:通过模块化设计(如Arm Neoverse CSS)降低风险、复用投资,将集成点从PCB移至SoC,预验证计算与子系统,节省21个工程师年。 3. **开放生态构建**:Arm推动UCIe、OCP等开放标准,定义Foundation Chiplet System Architecture(FCSA),实现跨厂商互操作,加速SoC设计“左移”。 4. **性能优化**:定制化Chiplet针对AI场景优化,平衡标准化与灵活性,提升每瓦性能,满足数据中心与边缘设备需求。
**Chiplet如何助力AI?** **AI芯片成本如何降?** **Chiplet生态如何建?**
客服
商务合作
小程序
服务号
折叠