当前位置:首页 > 报告详情

20250121_PreConE_Joly.PDF

上传人: 哆哆 编号:631036 2025-04-19 21页 3.41MB

1、Simplifying the Use of Co-Packaged Optics with ChipletsSylvie JOLYPartnerships Manager 3D integration&PackagingThe computing power required for AIis doubling every 100 days andis projected to increase by more thana million times over the next 5 yearsZhu&al,Intelligent computing,2023Photo:28/10/2022E

2、xemple de pied de page(A modifier dans longlet Insertion/En-tte/Pied3AI Performance challengesMemory and Interconnect WallAI and Memory WallAmir Gholami,UC Berkeley,Berkeley,CA,94720,USAPhoto:28/10/2022Exemple de pied de page(A modifier dans longlet Insertion/En-tte/Pied4Packaging&Chiplet answer to

3、performance?Intel Clearwater Forest for Server CPUAMD MI300 for AIAWS Graviton4Nvidias BlackwellBroadcomChiplets and Interposers becoming mainstreamScale-out at package levelHeterogeneous integrationRecover yieldpartitioning creates communication bottlenecksOptical reach went from 100km to 100m to m

4、 to board-levelWhat opportunities down to chip-level?Die-to-die optical links:the right technology for interposer-level communication!Relieve constraints of dense on-chip interconnectsReduce latency created by distance&routingImprove performance&power by overcoming latency overheadsAppl.Phys.Lett.20

5、21(Broadcom,UCSB,Intel)Enabling Communication technology roadmapEnabling Communication technology roadmapMature CMOS ecosystem applied to silicon photonicModern packaging and 3D integration Up to optical links for chiplets on interposer Bring silicon-photonic and microelectronic world together Integ

6、rationModern foundries are in 300mm and InP/GaAs in lower wafer sizeDesign ecosystem capabilities,development of assembly PDKs,and thermal design managementHigh Yield required for co-integration of 100s of devices together.Reliability:laser implementation is still an issues,most PIC use desagragated

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。
本文探讨了人工智能时代,为满足计算能力的需求,芯片封装技术的创新至关重要。据文中所述,AI性能挑战主要体现在内存和互联墙上,即数据处理速度受到内存访问和芯片间通信的限制。为解决这些问题,文章提出了封装和芯片小片(Chiplet)技术作为性能提升的潜在解决方案。具体关键点如下: 1. 计算能力需求激增:AI计算能力每100天翻倍,预计未来五年将增长超过一百万倍。 2. 封装技术革新:文章提到Intel、AMD、Nvidia等公司正在探索的封装技术,如硅光子学、3D集成和先进封装技术。 3. 光电集成:光电链接作为互联层次通信的技术,可减少密集芯片互联的约束,降低延迟,提高性能和能效。 4. 技术挑战:文章讨论了光电集成中的挑战,包括如何在微电子和光电子世界之间架起桥梁,以及如何处理激光实现、模块组装标准化等问题。 5. 行业合作:强调了建立稳固的供应链、供应商网络和标准化流程的重要性。 综上所述,为应对AI性能挑战,文章强调了光电集成、先进的封装技术以及行业合作的重要性,并提出了CEA-Leti在光电集成和微电子封装领域的研究进展和技术解决方案。
如何缓解内存和互联墙问题?" 如何优化高性能计算和AI领域的封装?" 硅光子技术如何推动芯片级光学集成?"
客服
商务合作
小程序
服务号
折叠