当前位置:首页 > 报告详情

20260219_B-202_Nowroozi.PDF

上传人: 彩旗 编号:1158895 2026-03-02 30页 74.84MB

1、1H-103 Chiplet Summit 2026 LIGHTMATTERPhotonic Interposer Supports Terabit Chiplet-Based SystemsB-202:Die-to-Die Interfaces-6:New ApproachesBijan Nowroozi Head of Ecosystem Development2H-103 Chiplet Summit 2026 LIGHTMATTERThe Shift In The Semiconductor Scaling EconomyAI ROI has shifted from transist

2、or to system based scaling.Source:Intel,Our World in Data,Alger.https:/ourworldindata.org/Moores Law refers to length of time that it takes for number of transistors per integrated circuit to double.Doubling time for Al refers to the length of time it takes to double the amount of compute or trainin

3、g utilized by Al programs.The calculation period used for Al training was 2012-2026.4X Improvement3 3Master Slide TemplateH-103 Chiplet Summit 2026 LIGHTMATTERTHE NETWORK IS BECOMING THE COMPUTER4H-103 Chiplet Summit 2026 LIGHTMATTERThe Trend:Extending the Die with IODIEDIEDIEDIEDIEDIEDIEDIEDIEMCM O

4、rganic SubstrateLogic+IO w D2DReticle LimitedEverything2.5D InterposerAbstracted+Disaggregated3D InterposerAbstracted+Disaggregated+Stacked(RDL etc)TransistorDIE ScalingIO Scaling3D IO Scaling5H-103 Chiplet Summit 2026 LIGHTMATTERINTER-CONNECTTRAYRACKCLUSTERCHIPLETDIEMore Than Moore:The Network Did

5、Become The ComputerScaleOutUpAcrossAt every level,from transistors to clusters,performance is governed by network efficiency.6H-103 Chiplet Summit 2026 LIGHTMATTERMORE GPUS NEED A LOT MORE I/OBig AI supercomputers require links.SERDES LANES IN AN NVLINK SCALE-UP DOMAIN8 GPUS 202272 GPUS 2024576 GPUS

6、 20257H-103 Chiplet Summit 2026 LIGHTMATTERThe interconnect bottleneckRents Rule8H-103 Chiplet Summit 2026 LIGHTMATTERNetworking Scaling:Faster and More Dense224G448G448G224G224G224G800G Link=16 Wires.1 TX channel:2 wires 0101011 RX channel:2 wires 010101 With Copper,this 102T switch ASIC uses 2040

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. **半导体经济转型**:AI投资回报从晶体管级转向系统级扩展,性能提升4倍。 2. **光子互连突破**:Lightmatter的3D光子中介层(如M1000)实现114Tbps带宽,支持1024 SerDes,256根光纤,冗余光路交换。 3. **解决互连瓶颈**:铜缆在1米以上距离受限,硅光子技术通过DWDM单光纤支持>16通道,覆盖500米至数公里。 4. **可扩展架构**:2x4瓦片设计支持跨光罩拼接,128 SerDes/瓦片,56Gbps NRZ/112Gbps PAM4,支持未来224G/448G SerDes。 5. **生态与开放合作**:推动开放标准(OIF、UCIe),简化部署,降低AI系统互连复杂度,目标实现400Tbps交换与100万节点扩展。
**光子芯片未来?** **互联瓶颈如何破?** **AI算力新引擎?**
客服
商务合作
小程序
服务号
折叠