当前位置:首页 > 报告详情

DCON26_PAPER_Track01_Adv.SoCICPackagingArchitecture_309_163.pdf

上传人: 明**** 编号:1195383 2026-04-16 27页 6.52MB

1、1 Advanced SoC/IC Packaging Architecture Dr.Moh Kolbehdari,Socionext USA M Gopalakrishnan Ramakrishnan Socionext USA Gopalakrishnan.R 2 Abstract Monolithic scaling alone no longer delivers the bandwidth-per-watt or memory proximity demanded by AI/HPC and high-I/O platforms.Modern SoC IC package arch

2、itectures from 2.5D interposers to 3D stacks and hybrid bridges now set the envelope for interconnect density,power delivery,thermal paths,reliability,and cost.Recent systems such as NVIDIA GB200/GB300 NVL72 liquid-cooled racks,AMD MI450 Helios platforms,and Intel EMIB/Foveros hybrids exemplify the

3、shift toward rack-scale,package-centric integration.In parallel,wafer-to-wafer(W2W)hybrid bonding is emerging as a disruptive capability,enabling ultrafine-pitch,low-parasitic vertical links and opening new options for memory-on-logic and logic-on-logic stacks.This paper compares five architecture f

4、amilies 2.5D interposers,3D stacks(incl.W2W),3.5D bridges,fan-out RDL,and organic MCM using consistent,testable criteria across:(i)electrical(channel loss/return paths,SI/PI/PDN targets,VRM partitioning),(ii)thermal-mechanical(CTE,warpage,TIM/underfill,hotspot escape),(iii)manufacturability&yield(as

5、sembly windows,rework ability,KGD policies),(iv)test/DFT(diepackageboard access/coverage),and.(v)cost/TTM&sustainability 1-10.Authors Bio Dr.Kolbehdari holds B.Sc.,M.Sc.,and Ph.D.degrees in Electrical Engineering,with doctoral research completed at Temple University(U.S.).He bridges academic rigor a

6、nd industrial execution,developing governed,physics-driven packaging methodologies such as the Tri-Loop(Predict Choose Release)and G G gating framework to deliver predictable yield,reliability,and time-to-market.His current research centers on making packaging architecture auditable,efficient,and sc

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. **核心问题**:单芯片 scaling 无法满足 AI/HPC 高带宽、低功耗需求,先进封装架构(2.5D/3D/3.5D/扇出/MCM)成为关键。 2. **架构对比**: - **2.5D硅中介层**:最高带宽密度(HBM3e/4),但成本高(Z_PDN≤0.4–0.5mΩ,ΔV≤40–50mV@100A/10ns)。 - **3D堆叠(W2W)**:超短垂直互连(<10μm pitch),但热管理挑战大(ΔT_tile需≤15°C)。 - **3.5D桥接(EMIB)**:局部高密度,折中成本与性能(桥接对齐≤±3μm)。 - **扇出RDL/有机MCM**:成本低、TTM快,但SI/PI受限(长链路损耗需控制)。 3. **设计框架**:Tri-Loop协同设计(物理建模→治理→制造),通过G0-G5门控(Predict→Choose→Release)确保可制造性与可靠性。 4. **关键指标**:PDN阻抗(Z_PDN(f)≤Z_target)、热管理(Tj,max、θ_jc/θ_ja)、良率(Cp/Cpk)、成本(BOM/TTM)。
封装架构如何选? 三大设计如何协同? 成本与性能如何平衡?
客服
商务合作
小程序
服务号
折叠