An Optical Accelerator for Deep Neural Network Based on Integrated Nanophotonics
An Optical Accelerator for Deep Neural Network Based on Integrated Nanophotonics
复制标题
DOI：
10.1109/icrc2020.2020.00017
复制
发表时间：
2020-12
期刊：
影响因子：
--
通讯作者：
Jun Shiomi;T. Ishihara;H. Onodera;A. Shinya;M. Notomi
中科院分区：
文献类型：
--
作者：
Jun Shiomi;T. Ishihara;H. Onodera;A. Shinya;M. Notomi
纳米光子器件的出现使得设计光速极低延迟的片上光学电路成为可能。提出了一种支持光速推理的可扩展深度神经网络(DNN)的光学实现。光神经网络的关键问题是可扩展性受到面积、功率和可用波长数量的限制。由于可伸缩性，很难为大规模DNN设计全光硬件加速器。针对这一问题，本文首先提出了一种低延迟的光矢量矩阵乘法器(VMM)结构。VMM中的乘法器基于波分复用(WDM)技术高度并行化，在不牺牲超高速特性的情况下减少了面积开销。提出了一种在不牺牲超高速特性的情况下存储和处理中间VMM数据的电子数字接口，使得VMM能够在低延迟的情况下多次重用。
The emergence of nanophotonic devices has enabled to design light-speed on-chip optical circuits with extremely low latency. This paper proposes an optical implementation of scalable Deep Neural Networks (DNNs) enabling light-speed inference. The key issue in optical neural networks is the scalability limited by area, power and the number of available wavelengths. Due to the scalability, it is thus difficult to design an all-optical hardware accelerator for a large-scale DNN. To solve this problem, this paper firstly proposes an optical Vector Matrix Multiplier (VMM) structure operating with a low latency. The multipliers in a VMM are highly parallelized based on the Wavelength Division Multiplexing (WDM) technique, which reduces the area overhead without sacrificing the ultra-high speed nature. This paper then proposes the electrical digital interfaces for storing and handling intermediate VMM data without sacrificing the ultra-high speed nature, which enables to reuse the VMM multiple times with a low latency.