当前位置:首页 > 报告详情

Scalable Optical IO for Disaggregated Infrastructure.pdf

上传人: 张** 编号:158263 2024-03-31 12页 856.11KB

1、OCP Global Summit October 18,2023|San Jose,CASYM Title SlideLK Bhupathi,Ayar LabsScalable Optical I/O for Disaggregated InfrastructureSYM-Content100s to 1000s xPUs working togetherRadix and throughput demands growing for inter-node connectivityxPU chips to escape multiple TBps to support clusteringS

2、witch chips to offer 10s of TBps of connectivityExample topology shown;other partitions possibleDatacenter Trends Need for Larger Clusters Pooled/Shared DRAM/HBMInter-chassis high throughput low-latency interconnect for pooled/shared memoryCapacity expansion via disaggregation of HBM demands TBps li

3、nks(not shown)Enabling partitioning-flexibility and composability of resourcesExample topology shown;other partitions possibleMemory Pools and Capacity ExpansionApplications and Use-cases of Optical I/OArchitectural Parameter DescriptionApplication requirement or technical capabilityApplication requ

4、irement:Capable and efficient memory-semantic fabrics that deliver scale-out,memory pools,and composability.Technical capabilities:Multi-TB/s connectivity,power-efficiency,BW density,low-latency,and easy-to-manage connectorized fiberUse case Node-node connectivity CXL-based resource disaggregation H

5、BM disaggregation and capacity expansionFuture with Optical I/OTechnology Basics1,000 x smaller than traditional devices WDM for TBps throughput capabilityMicroring ResonatorsTeraPHY chiplet for in-package optical I/ODense 300mm CMOS integrationElectro-Optic ChipletIn-Package IntegrationIntegration

6、with state-of-the-art SoCsDirect from the package optical I/OCapability Scaling(Tx+Rx)Chiplet bandwidth=2*(#of ports/chiplet)x(#of wavelengths/port)x(data rate/wavelength)Chiplet bi-directional BW#of ports/chiplet#of wavelengths/portData rate/wavelength4.096 Tbps(0.5 TBps)8832 Gbps8.192 Tbps(1 TBps)

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。
在OCP全球峰会上,Ayar Labs的Calk Bhupathi介绍了了一种可扩展的光学I/O技术,旨在支持 disaggregated infrastructure。该技术能提供100s到1000s xPUs的高效协作,满足不断增长的节点间连接需求。通过采用WDM技术,该光学I/O解决方案具备了 TBps级别的吞吐量能力。Microring Resonators和TeraPHY™芯片实现了高密度的CMOS集成和封装内光I/O。该技术还支持多链接、多波长和双向带宽,可扩展到32.768 Tbps (4 TBps)。此外,该光学I/O技术还具有低延迟、高能效和远距离传输的能力。Ayar Labs呼吁业界帮助发展光学I/O生态系统,标准化外形尺寸、连接器和光纤管理,并提出了相关的技术参数和应用场景。
"2023年OCP全球峰会将展示哪些技术突破?" "如何通过光学I/O实现数据中心的集群扩展?" "Ayar Labs的光学I/O技术如何推动数据中心未来发展趋势?"
客服
商务合作
小程序
服务号
折叠