Research On Optical Network-on-Chip For Neuromorphic Computing

With the increasing applications of machine learning in recent years,massive amounts of data and computational operations have placed higher demands on the capabilities of data processing,interconnection,and machine learning for computer processors,which has promoted the innovation of computing architecture and calculation methods of the processors.In the past development of processor technology,reducing the size of transistors and increasing the operating frequency are the main methods to improve the performance of single-core processors.However,due to physical limitations,it has become more and more difficult to further reduce the size of transistors,and excessively high operating frequencies will also lead to a reduction in processor energy efficiency.Network-on-Chip(No C)transplants the idea of networking into the chip design,and efficiently interconnects the numerous on-chip IP(Intellectual Property)cores through the network.It has a series of advantages such as excellent communication performance,reusability,scalability and parallelism.As the feature size of the VLSI process is further reduced,the system performance improvement of the on-chip network using electrical interconnection is severely restricted due to the parasitic effects of metal wires,delay time,signal crosstalk,and energy consumption.Optical Network-on-Chip(ONo C)uses optical interconnection to effectively solve the problems of electrical interconnection,with higher bandwidth density,smaller communication delay,lower system power consumption and smaller network crosstalk.The widely used von Neumann computing system now has a ”memory access wall” problem caused by the communication delay between separate storage and computing units,which limits the further development of future processor computing capabilities.Neuromorphic computing is inspired by mechanisms of both cranial nerve and cognitive behavioral,learning from biological brain processing mechanisms to make the machine adaptive and selflearning.It breaks through the traditional von Neumann architecture ”memory access wall”bottleneck,increases the information processing rate and reduces power consumption significantly at the same time.And it becomes a hot spot in the field of artificial intelligence and computer science in recent years.The Optical Network-on-Chip for neuromorphic computing uses the multi-dimensional parallel capabilities of optical signals and high energy-efficiency processing capabilities to construct an on-chip optical Spiking Neural Network.It has the advantages on adaptability,robustness and rapidity,effectively avoiding both the noise accumulation of traditional digital optical computing chip and the integration problems of analog optical computing chip,so that becoming a breakthrough direction for the new generation of machine learning hardware.This article aims to design an Optical Network-on-Chip for neuromorphic computing.It aims at the energy efficiency ratio,connectivity,scalability,and intelligence of existing artificial intelligence implementation methods,combining the low power consumption,high bandwidth and high rate as well as other advantages of on-chip optical computing and onchip optical interconnection.Our research is on the construction the on-chip optical neuron,design for on-chip optical interconnect architecture,optimization of on-chip optical Spiking Neural Network training algorithm.By theoretical analyze,strategy design,experimental verification with device-level simulation,network-level simulation,prototype tape-out testing and functional experiment,we verify and evaluate the performance of the study,and build a low-energy,high-parallel,large-scale,and easy-to-train Optical Network-on-Chip for neuromorphic computing.The main work and research results of the thesis are as follows:(1)The research progress on optical computing for machine learning is comprehensively overviewed,starting from neural network models such as deep learning,probability graphs,and spiking neural networks.The existing design schemes are sorted and summarized,with emphasis on the optical realization of machine learning.The relevant theoretical basis,technical route,research content are introduced,the existing challenges of artificial intelligenceoriented optical computing are summarized,and the new development of silicon-based optoelectronic integration technology is proposed to improve the optical neural network.(2)In view of the insufficient scalability of the current on-chip artificial optical neurons and the integration difficulty of off-chip integrated lasers in the neuromorphic network,we use the In As / Ga As Quantum Dot Laser for large-scale silicon-based optoelectronics on-chip integration,To characterize the energy level transition by rate equation through a numerical model,the performance of the quantum dot laser is optimized and matched to the LIF neuron model.The simulation results show that the on-chip optical neuron based on quantum dot laser has the ability of artificial neuron pulse generation,weighting,integration,threshold processing,reset and reset,and the pulse time resolution reaches the picosecond level.In addition,using the constructed artificial neuron based on quantum dot lasers,a four-input on-chip neuron was modeled and simulated,and the output pulse signal amplitude was configured according to different weights to verify the QDL-based neuron.The meta-stimulation and response functions lay the foundation for the further construction of large-scale on-chip networks.(3)Aiming at the problems on the power consumption of the on-chip network photoelectric conversion for electrical signal injection and the low utilization rate of the on-chip network for coherent light injection,combined with the basic architecture of the incoherent optical signal injection type on-chip network,a biologically-inspired optical neuromorphic Network-on-Chip is designed.The on-chip network architecture is specifically divided into five levels: neurons,neuron groups,nerve clusters,cellular functional areas,and regional coordination systems.Through hierarchical and regional scheduling and coordination,wavelength overhead is reduced,and wavelength division multiplexing and space division are used.Multiple methods such as multiplexing improve the parallelism of the system,and overcome the problems of low connectivity and poor scalability between the neurons of the existing artificial neural network.Based on the structural characteristics of the current spiking neural network that can be realized by two layers under the classic model,a universal construction method of the on-chip fully-connected structure supporting wavelength division multiplexing between the two layers of neurons is designed,and the tape-out verifies the realizability and functionality of the fully connected structure LACE,providing a physical basis for building a highly parallel,large-scale optical neuromorphic No C architecture.(4)In view of the poor interpretability of the current Spiking Neural Network training algorithm and the hardware loading problem of the on-chip optical components in the large-scale integration,based on the coherence effect model for on-chip optical synaptic based on the microring resonator and the measured hardware loss data of the on-chip fully connected structure LACE,a weight correction method based on the Spike Timing Dependent Plasticity of the Spiking Neural Network is proposed.The Spike Neural Network is based on our biologically-inspired optical neuromorphic Network-on-Chip architecture,oriented to a two-layer neural network with an excitatory neuron layer and an inhibitory neuron layer,combined with the mathematical representation of the MR weight realization range to optimize the STDP learning algorithm.With optimization of training parameters,our optical weight correction method of Spiking Neural Network for neuromorphic computing reach91.54% recognition accuracy rate of Mixed National Institute of Standards and Technology database.