Research And Implementation Of Neuromorphic Computing Workload Mapping Method Based On SNN

At present,driven by massive annotation data and increasing computing power,artificial intelligence represented by deep learning has developed rapidly,but behind its high accuracy rate,there are also weak levels of general intelligence and higher computing power dependence.Neuromorphic computing represented by the third-generation artificial neural network-Spike neuron network(SNN)draws on the computing characteristics of the brain with high efficiency and low energy consumption,which is considered to be an important way to solve artificial intelligence problems.Due to the fast and even real-time simulation requirements of SNN and the obvious characteristics of distributed computing,large-scale distributed clusters are used as the main way of constructing neuromorphic platforms.For distributed neuromorphic platforms,if there is a workload mismatch between the SNN case and the platform,the computing energy efficiency of the dedicated neuromorphic system will even be lower than that of the general-purpose computer system.Therefore,how to quickly complete the analysis of the load characteristics of the SNN case based on the brain-inspired platform,and realize the perfect matching and reasonable mapping between the case and the platform,has become one of the thorny issues that the current research and optimization of the brain-inspired architecture need to face.To this end,this article built 100 PYNQ-Z2 development boards to form a distributed cluster as the experimental hardware platform,and selected the SNN simulation software NEST as the load research object.Three parts of research work were carried out: SNN load characteristics analysis,SNN workload automatic mapping,and load evaluation and analysis for the energy-efficient PYNQ cluster platform.(1)Analyzed the SNN workload characteristics and established a load model for it,and further instantiated SNN's memory,calculation and communication load models for the NEST simulator,and used specific load data to illustrate the computing platform that carries the SNN case Load situation.(2)Based on the load instantiation model,a Workload Automatic Mapper for SNN(SWAM)is designed and implemented.SWAM contains three parts of design: quantitative design,mapping special script design and automation design.Quantitative design realizes one-time quantification of memory and time parameters at a low cost through MPI quantization program and benchmark quantization program;On the basis of retaining the original NEST network construction habits,a special script format for mapping was developed,and a simple function control interface design was used to help the realization of the automation design;The automation design includes three automation parts: network parameter collection,load forecasting and mapping.The overall automation design is an integrated realization,which greatly shortens the overall process of load forecasting and mapping.(3)For the energy-efficient PYNQ brain cluster platform based on ARM+FPGA,SWAM is used to provide three load services for it: load evaluation,accelerated judgment and call,load prediction and node mapping.Through three load services,to help the platform to always maintain high energy efficiency,high stability and high performance when carrying SNN cases.By running SNN typical cases on PC,pure ARM and ARM+FPGA three different cluster computing platforms,and comparing SWAM,LM(Levenberg-Marquardt)algorithm fitting and actual mapping results.The experimental results show that the average mapping accuracy of SWAM reaches 98.833%;compared with the LM method and the measured mapping,SWAM has an absolute time cost advantage.Through experimental analysis,it is proved that SWAM has good compatibility,can cover and predict the load situation of different SNN cases on different computing platforms,and quickly and effectively derive the reasonable mapping results of the entire workload on the computing platform.It ensures the stable and highperformance operation of the computing platform,avoids the extremely time-consuming operation and trial of the complete workload,and has practical significance for the research and optimization of the neuromorphic architecture.