一种高效的通用神经网络光学加速器.pdf

1、An Efficient General-Purpose Optical Accelerator for Neural NetworksASP-DAC 2025Sijie Fei(Technical University of Munich)Amro Eldebiky(Technical University of Munich)Grace Li Zhang(Technical University of Darmstadt)Bing Li(University of Siegen)Ulf Schlichtmann(Technical University of Munich)1Why Opt

2、ical Neural Networks(ONN)2Strict time constraintsMillions/Trillions ParametersONNhigher speedBasics of ONN3Mach-Zehnder Interferometers(MZI)MxN Matrix (M2+N2)/2 MZIs Microring Resonator(MRR)6 YichenShen,NicholasCHarris,Skirlo,etal.Deep learning with coherent nanophotonic circuits.Nature photonics,11

3、(7):441446,2017.17 Viraj Bangari,Bicky A Marquez,Heidi Miller,et al.Digital electronics and analog photonics for convolutional neural networks(DEAP-CNNs).IEEE Journal of Selected Topics in Quantum Electronics,26(1):113,2019.UnitarySVD6Weight Matrices with cascading MZIslimited expressivity due to wd

4、m17Architecture dependent on matrix dimensionLimitation of General-purpose Optical Accelerators(GOA)Represent a 4x4 matrix using an 8x8 GOAa 4x4 matrixMismatch of the matrix dimensions and GOA dimensionsLow utilization efficiency High mapping effort (Mapping:tuning PSs on GOA one time)(18+4)/28 MZIs

5、 are affected and unusedGOA:the same optical accelerator that can be reused for different neural networks4GOA ArchitectureHigher utiliztiom efficiency and low mapping effort with independent MZI modules5Mapping NNs and Determining GOA ParametersGenetic algorithm with metrics:Mapping cost-necessary m

6、appings for one NN Area cost Power E/O conversions MZIs,MRRs and peripheral devices6NN adjustment and Hardware-aware Trainingh:kernel size d:kernel depth n:kernel number Restoring by columns Training7l_a=h_a2 x d_an_b=d_aExperimental Results-Mapping/Energy/Lat