当前位置:首页 > 报告详情

DC25_SLIDES_Track14_FlowModelCanGiveYou_HyunjunAn_V2.pdf

上传人: S** 编号:1240914 2026-05-16 46页 2.55MB

1、Information Classification:GeneralWelcome to ConferenceJanuary 2830,2025Santa Clara Convention Center1ExpoJanuary 2930,2025Information Classification:GeneralFlow Model Can Give You Optimal and Diverse Design Guidance for TSV Array Placement in HBMSpeaker:Hyunjun An(KAIST)Paper Authors:Hyunjun An(KAI

2、ST),Keunwoo Kim(KAIST),Haeyeon Rachel Kim(KAIST),Junghyun Lee(KAIST),Taein Shin(KAIST),Keeyoung Son(KAIST)and Joungho Kim(KAIST)2Information Classification:GeneralImageSPEAKERSHyunjun AnM.S.Student,Korea Advanced Institute of Science and Technology(KAIST)Email address:anhyunjunkaist.ac.kr-Machine Le

3、arning-based SI/PI Analysis-Design Automation of 2.5-D/3-D ICs3Information Classification:GeneralContents4I.Introduction to TSV Array Placement Problem II.Proposal of Hardware Problem Specialized-Generative Flow Network(HPS-GFN)III.Problem Setup for TSV Array Placement Problem IV.Verification of the

4、 Proposed MethodV.ConclusionInformation Classification:GeneralContents5I.Introduction to TSV Array Placement Problem II.Proposal of Hardware Problem Specialized-Generative Flow Network(HPS-GFN)III.Problem Setup for TSV Array Placement Problem IV.Verification of the Proposed MethodV.ConclusionInforma

5、tion Classification:GeneralRapid Development of Generative AI6Currently,a diverse services based on generative AI are rapidly emerging.Information Classification:GeneralTrend of Generative AI:Aggressively Increasing Model Size7Model size is aggressively increasing,encouraged by introduction to multi

6、-modal AI.Visual SystemAuditory SystemRecognitiveSystemInformation Classification:GeneralHBM:The Solution for Intense Computational Demand8High bandwidth memory(HBM)is a memory with high density interconnect,providing ultra-wide bandwidth and large memory capacity.Information Classification:GeneralT

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. **研究问题**:针对HBM中TSV阵列放置的复杂组合优化问题,提出硬件问题专用生成流网络(HPS-GFN),以优化信号完整性(SI)。 2. **核心方法**:HPS-GFN基于生成流网络(GFlowNet),采用顺序学习、多模型和流归一化方案,避免局部最优,提供多样化解。 3. **性能验证**:在9种测试案例中,HPS-GFN的平均眼图开度(EO)达61.9%(Top-10为65.7%),优于深度强化学习(DRL, 60.3%)、遗传算法(GA, 59.4%)和随机搜索(RS, 55.9%)。 4. **设计指导**:通过流图(Flow Map)提供TSV放置的直观指导,例如在差分信号TSV周围优先放置地TSV,避免过设计。 5. **未来方向**:扩展至更大规模TSV阵列、提升模型复用性(如差分信号位置适应性)及整合电源TSV放置。
**HPS-GFN如何优化TSV布局?** **AI如何解决HBM信号完整性问题?** **TSV布局的多样性如何实现?**
客服
商务合作
小程序
服务号
折叠