当前位置:首页 > 报告详情

帕累托最优光子张量核的零样本自动电路拓扑搜索.pdf

上传人: 芦苇 编号:651860 2025-05-01 22页 4.17MB

1、ADEPT-Z:Zero-Shot Automated Circuit Topology Search for Pareto-Optimal Photonic Tensor Cores1Ziyang Jiang1,Pingchuan Ma1,Meng Zhang2,Rena Huang2,Jiaqi Gu11Arizona State University,2Rensselaer Polytechnic Institute1School of Electrical,Computer and Energy Engineeringzjian124asu.edujiaqiguasu.edu|scop

2、ex-asu.github.ioCopyright 2022 Arizona Board of RegentsPhotonic ML Accelerators2Evolve from electronics to integrated photonicsSource:Mitchell A.Nahmias,Bhavin J.Shastri,Alexander N.Tait,Thomas Ferreira de Lima and Paul R.Prucnal,“Neuromorphic photonics,”Optics&Photonics News,Jan 2018.Optical-electr

3、onic hybrid chips(10 TOPS/W)Fully-optical chipsNeuromorphic photonics(1E6 TOPS/W)Energy Efficiency(TOPS/W)Compute Density(TOPS/mm2)Speed-of-light,Massive parallelism,low-powerpotential105Copyright 2022 Arizona Board of RegentsPTC:Photonic Tensor Core3Basic Optical ComponentsSVD-based ONNs with MZIs

4、Nature17,Shen+WButterfly-style PTC ASP-DAC20,Gu+PTC with MMI Devices Sci.Rep.24,Zelaya+Modular Coherent PTC DesignSRepeated block structures-Phase shifter arrays(s)-Coupler arrays()-Crossing arrays()DCCRPScouplercrossingphase shifterMatrix:=222111Coherent PTCs:Complex-valued Matrix MultiplicationCop

5、yright 2022 Arizona Board of RegentsManual PTC Design Auto-Searched Design?MMI ArrayDA for PhotonicsManual Design Automated PTC DesignCan we automatically&efficiently find Pareto-optimal photonic circuits designs?Copyright 2022 Arizona Board of RegentsPrior Work on Auto PTC DesignRelax as continuous

6、 optimizationGradient descent to search(similar to DARTS in one-shot neural architecture search)Challenges in Differentiable ADEPT Limited design space:fixed slots for 2x2 coupler,cannot handle multi-port couplers Optimization difficulty:hard to handle multiple non-differentiable objectives/constrai

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。
本文介绍了一种名为ADEPT-Z的零样本自动电路拓扑搜索方法,旨在为光子张量核心(PTC)寻找帕累托最优解。与传统的需要梯度下降搜索的方法不同,ADEPT-Z使用进化算法,能够在更大的设计空间中进行探索,支持多目标优化和硬件约束。该方法在Global Foundries的PDK上搜索16×16 PTC(2/8-端口直流耦合器),在2.7小时内获得了40个帕累托最优解,比ADEPT的6-10小时单次搜索快了100倍。ADEPT-Z搜索到的设计在准确性加权面积能量效率产品(AAEE)方面优于手动设计和其他光子阵列。此外,该方法还适应了定制foundry PDK,并在不同的任务和模型上进行了验证,平均AAEE比手动PTC设计提高了1.6倍。
如何实现零样本光子张量核心搜索?" "如何通过进化算法探索巨大的光子电路设计空间?" ADEPT-Z的优势与挑战是什么?"
客服
商务合作
小程序
服务号
折叠