当前位置:首页 > 报告详情

20260218_E-103_Hosseini.PDF

上传人: 彩旗 编号:1158893 2026-03-02 15页 2.11MB

1、Chiplet SummitChiplet SummitFebruary 17February 17-19,2026,Santa Clara,CA,USA19,2026,Santa Clara,CA,USAChiplet Packaging to Scale AIChiplet Packaging to Scale AIAditya Vaidya,Si-Pkg Arch and Customer Enabling LKaveh Hosseini PhD,Principal EFebruary 18,2026Acknowledgements:R.Mahajan,several Intel tea

2、m members Chiplet Summit 2026,Aditya Vaidya2 What is AI scaling and what is needed to keep up with it Basics of an AI chip and how to scale it Challenges and opportunities Warpage Power Delivery Thermal challenges Yield and Redundancy ConclusionOutlineChiplet Summit 2026,Aditya Vaidya3AI scalingScal

3、ing requires more data used for training,more model parameters and faster compute(1)Amount of FLOP/s per$doubles every 2.5 years(2)AI scaling requires more compute and memory which in turn requires more Si packed on the packageHeterogeneous Si and memory integration is cost effective and scalable wa

4、y to achieve thisTrend of substrate and silicon area for AI packagesChiplet Summit 2026,Aditya Vaidya4Chiplets in AI package and its scalabilityHigh BW communication needed between the different elements of the unit cellOff package communication handled via IO dieScaling for the AI chip will require

5、 multiple such unit cells placed on a packageWafer scale packaging is first attempt at scaling but is limited to wafer sizeIntels EMIB-T based architecture allows an elegant and sustained scaling solutionSource:Intel on YoutubeAI scaling using EMIB-T to achieve 10 x reticle complexBlow up view of Te

6、sla Dojo AI training tile(3)Unit CellxPU=Compute chipletHBM=Memory chipletIO=Data in and out of packagexPUHBMHBMIONVIDIA Blackwell GPU(4)AMD MI300(5)Intel Gaudi 3(6)All Si to Si solution via EMIBEMIB-T TChiplet Summit 2026,Aditya Vaidya5EMIB-T PrimerCurrently in High volume:EMIBThe Future:EMIB-T(7)E

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. **AI scaling需求**:AI算力需每2.5年翻倍,依赖更多数据、参数及计算力,需通过异构集成(如Chiplet)在封装中集成更多硅片。 2. **EMIB-T技术**:Intel的EMIB-T(含TSV)支持>10倍光刻复杂度扩展,降低功耗噪声,提升D2D带宽,支持多晶圆厂异构集成(如UCIe、HBM)。 3. **核心挑战**: - **翘曲**:大尺寸封装需玻璃基板或低温焊料等创新。 - **供电**:传统横向供电不足,转向垂直供电(如嵌入式电容、MIM)。 - **散热**:高功率密度需优化TIM、冷板设计及2.5D/3D堆叠。 - **良率**:需多级冗余(内存/逻辑级、互连级如UCIe 3.0的2冗余通道/32数据通道)。 4. **结论**:异构封装是AI扩展关键,但需解决材料、供电、散热及良率问题,同时需负责任地使用AI能源。
**AI芯片如何突破?** **封装技术如何助力AI?** **AI散热有何新方案?**
客服
商务合作
小程序
服务号
折叠