A Microdisk-based Photonic Accelerator For Deep Learning

Recently,convolutional Neural Networks(CNNs)are widely adopted in object recognition,speech processing and machine translation,due to their extremely high inference accuracy.However,as CNN model becomes more and more complex,it is very challenging to perform convolutional operations with high computational overhead on traditional CPU and GPU devices.Although some new CMOS-based deep learning accelerators are explored extensively,such as Google's TPU,these solutions still face high energy consumption and data communication bottlenecks.Today,emerging nanophotonic technology has been employed for on-chip data communication due to its CMOS compatibility,high bandwidth and low power consumption,and some recent works are also trying to apply this technology to on-chip computing.In this paper,we propose a microdisk-based photonic CNN accelerator to boost the CNN inference in datacenters.Instead of an all-photonic design,our design performs convolutions by photonic integrated circuit,and process the other operations in CNNs by CMOS circuit for high inference accuracy.We first build a photonic matrix-vector multiplier unit using microdisk-based resonators and then propose HolyLight-M based on the constructed matrix-vector multiplier.However,we found that on-chip analog-to-digital converters(ADCs)seriously limit its computing performance per Watt.We further build digital adder and shifter units by microdisk and then construct HolyLight-A without ADCs.Compared to the state-of-the-art Re RAM-based accelerator ISAAC,HolyLight-A improves the CNN inference throughput per Watt by 25× with trivial accuracy degradation..