当前位置:首页 > 报告详情

20260219_A-201_Jose.PDF

上传人: 彩旗 编号:1158829 2026-03-02 18页 1.54MB

1、Chiplet Summit 2026Title:Managing Workloads in Chiplet Based Data Centers using Digital TwinsPresenter:Deepak Shankar|LinkedInPresenter:Tom Jose|LinkedInPRESENTATION OUTLINE01The Chiplet RevolutionWhy data centers are shifting from monolithic to chiplet-based architectures02Challenges at ScaleSystem

2、-level tradeoffs in compute,memory,interconnect,power,and thermal03Data Centre Digital Twin ArchitectureScalable modeling from chiplets to pods with parameterized workloads04Key Results&MetricsQuantified benefits across scheduling,interconnect,memory,and power05Design Rules&ConclusionsActionable gui

3、delines for chiplet-based data center deploymentTHE CHIPLET REVOLUTION IN DATA CENTERSFrom Monolithic to ModularModern data centers now integrate modular chiplet-driven systems connecting diverse silicon dies through advanced interconnects,replacing traditional monolithic server architectures.CPUsGe

4、neral computeGPU ModulesNVIDIA DGX B300Memory ChipletsDisaggregatedDPUsData processingINTERCONNECT FABRICUCIeUniversal Chiplet Interconnect Express die-to-die within packageNVLinkGPU-to-GPU high-bandwidth interconnect for aggregated computePCIe/CXLHost-to-device connectivity and memory expansion pro

5、tocolsSpectrum-XHigh-performance Ethernet switching for distributed AI trainingWHY CHIPLETS?KEY ADVANTAGESDomain-OptimizedSiliconEach chiplet is purpose-built for its function compute,memory,I/O enabling best-in-class performance per tile.Greater YieldPer WaferSmaller dies have exponentially higher

6、manufacturing yield than monolithic designs,reducing cost per good die.FieldUpgradeabilityIndividual chiplets can be swapped or upgraded without replacing the entire system,extending platform lifecycle.Thermal&PowerPartitioningFine-grained control over power delivery and cooling per chiplet,enabling

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. **Chiplet革命**:数据中心从单片转向模块化Chiplet架构,通过UCIe、NVLink等互连实现异构集成(CPU/GPU/内存/DPU),提升性能与良率。 2. **核心挑战**:系统级权衡(计算/内存/互连/功耗/热管理),需通过数字孪生架构建模解决。 3. **数字孪生架构**:从Chiplet到Pod的分层建模,支持参数化AI工作负载,动态调度任务映射,提升硬件利用率>30%。 4. **量化收益**: - 互连能效提升25倍(vs PCIe); - 分布式训练吞吐量提升1.6倍(Spectrum-X); - 内存按需扩展,避免过度配置; - 功耗/热设计节省10-20% CapEx。
**Chiplet如何提升数据中心效率?** **数字孪生如何优化芯片负载?** **异构芯片如何协同工作?**
客服
商务合作
小程序
服务号
折叠