Research On The Architecture Of Large-Scale Neuromorphic Computer

In the era of big data intelligence,due to the limitations of the existing computer architecture and hardware,large-scale intelligent data processing is increasingly difficult to meet.Based on the "Von Neumann" architecture of the traditional computer architecture,its storage unit and processing unit is an independent two parts,the two rely on the data transmission bus for interconnection,high energy consumption and low efficiency caused the "Von Neumann bottleneck" appeared.For the structure of the brain,the brain neural network contains approximately 100 billion neurons and 100 trillion neural synapses.They are processing units and memory units respectively.They are physically connected and do not need to be very large.Communication loss,especially after the relevant signals are excited to send out,the system is still in an excellent transmission state.These highefficiency,low-cost,and large-scale brain neural network architectures have powerful advantages and are very conducive to the processing of massive data.With the emergence of the crisis of Moore’s Law and breakthroughs in the study of pulsed neural networks,the academic community has turned the next generation of computer architecture research into the neuromorphic computer field.The current research on large-scale neuromorphic computers is in full swing,but it is actually difficult to implement.This article focuses on exploring the system architecture of ultra-large-scale neuromorphic computers,focusing on solving the problems of massive neuron pulse transmission problems,real-time processing problems,and board-level interconnection issues that are plagued by ultra-large-scale neuromorphic computers.After deeply understanding the research progress in related fields at home and abroad and absorbing beneficial essences,this paper proposes a simulation program for neurons and designs a CAM-based massive pulsed neuron packet multicast transmission mechanism and successfully solves the board-level interconnection problem.In order to verify the feasibility of this architecture,this paper designed a basic neuromorphic computer prototype and neuromorphic computer motherboard.The neuromorphic computer motherboards successfully developed have a size of 38*68cm.There are 48 multi-core CPUs integrated on the board.Each CPU contains 8 ARM cores and a total of 384 ARM cores.Neurons are modeled in software and peak-produced packets propagate through custom multicast routers.After testing,the performance of this type of neuromorphic computer is in line with expectations,has low power consumption,high performance characteristics,and can simulate millions of levels of neurons.