Research On Network On Chip Machine Learning Mapping And Interconnect Crosstalk Technology

As an important technical support for the development of Very Large Scale Integrated Circuits(VLSI),Network on Chip(NoC)is widely used in the design of Multi-Processor System on Chip(MPSoC),due to its modular design,high bandwidth,flexible and simple design.As the most important bus technology and development direction of digital integrated circuits,NoC plays an important role in the development of digital integrated circuits.Breaking through the NoC IP(Intellectual Property)technology is an important step in the development of digital integrated circuits.Application mapping and interconnect crosstalk suppression of NoC are key technologies which have the ability to determine the development of NoC.Application mapping determines the most performance of NoC design,and interconnect crosstalk suppression determines the back-end realization of NoC.This paper focuses on the research of NoC technology,especially on the isomorphism of application mapping and interconnect crosstalk.From the perspective of basic theory,mathematical model,theoretical research and solution research are carried out on the mapping isomorphism,interconnect crosstalk error rate and fault-tolerant system.Based on the theory of isomorphism,an NoC application mapping algorithm based on optimized genetic algorithm and an NoC machine learning mapping algorithm are proposed,which optimize the performance,computing speed and stability of NoC application mapping.Based on the physical model of interconnect crosstalk,the interconnect crosstalk error rate model and the IC(Integrated Circuit)interconnect crosstalk delay simulation method are proposed.The relationship between crosstalk error and chip design parameters is established.A crosstalk delay mapping method is proposed to reduce the complexity of NoC crosstalk simulation.Based on the theory of HARQ(Hybrid Auto Repeat re Quest)fault-tolerant system,the data transmission model of HARQ fault-tolerant system is established,and two NoC HARQ fault-tolerant system algorithm are proposed,which improves the effective transmission throughput.The main contributions are as follows:(1)An NoC application mapping algorithm based on optimized genetic algorithmAiming at the problems of solving process with low stability and easily falling into suboptimal solution in the calculation process in the NoC application heuristic mapping algorithm,an NoC genetic mapping algorithm based on density direction is proposed.On the basis of summarizing the iterative process and data form of existing algorithms,the definition of mapping solution isomorphism is proposed.On this basis,the convergence problem in the iterative process of NoC genetic mapping algorithm is analyzed.The definition of density direction and a density direction conversion algorithm are proposed to eliminate the isomorphic of mapping solutions in the iterative process.The density direction transformation algorithm is used to optimize NoC application mapping algorithm based on genetic algorithm,which improves the stability and convergence speed.Comparing with random mapping,the average delay of proposed NoC application mapping algorithm based on optimized genetic algorithm is reduced by 23.48% and the average power consumption is reduced by 7.15%.(2)A machine learning mapping algorithm for NoC optimizationMachine learning is introduced into NoC application mapping.According to the characteristics of machine learning,research is carried out in data form,solution space complexity,dataset construction method and machine learning model,which is applicable to proposed data form.An NoC mapping algorithm based on multi-label machine learning is proposed.This paper further analyzes the isomorphism and data form of NoC application mapping schemes,and designs a solution data form and a dataset augmentation algorithm for NoC machine learning mapping,which can generate the dataset based on a small amount of solutions and effectively reduce the difficulty of dataset construction.According to the data set form,a multi-label machine learning model is designed,which can improve the stability of the NoC application mapping.The simulation results show that:(a)The average accuracy of the multi-label machine learning model is 99.6%;(b)The average accuracy of the NoC multi-label machine learning mapping algorithm is 96.3%.Comparing with the global optimal NoC mapping solutions,the average performance loss caused by error prediction is 28.03%,and the average performance loss of all prediction results is 1.04%.Proposed algorithm effectively improves the stability of NoC application mapping.(3)The interconnect crosstalk error rate model and delay mapping methodThe coupling three-wire crosstalk model is introduced into NoC design,and the interconnect crosstalk error rate model and delay mapping method are established.A physic model of interconnect crosstalk is established and the main factors are analyzed.With the analysis,a crosstalk error rate model is established,which can be used to obtain the influence of the interconnect length,spacing,and operating frequency on the interconnect crosstalk error rate.It provides a parameter design basis for the chip design considering crosstalk.On this basis,with comparison of interconnect crosstalk SPICE simulation delay and parasitic parameter extraction delay,a crosstalk delay mapping method is proposed,which maps the delay of key interconnect parasitic parameter extraction delay to the worst crosstalk delay.So that the chip under crosstalk can be simulate with back-end functional simulation.Comparing with SPICE simulation analysis,the standard deviation of the proposed delay mapping method is 0.574 ps.This method optimizes the crosstalk simulation process,which simplifies the crosstalk analysis process of key interconnect SPICE simulation,and greatly saves the time of crosstalk simulation.(4)HARQ fault-tolerant system algorithm for NoCAccording to the characteristics of data transmission and communication in NoC,two HARQ fault-tolerant system algorithms for NoC are proposed.Two data transmission models of NoC HARQ fault-tolerant system are established,which model the expected transmission times and expected transmission delay of HARQ fault-tolerant system respectively.On this basis,an HARQ fault-tolerant system algorithm based on effective information length sifting is proposed,which can select the optimal coding style and corresponding single transmission effective information bit length of HARQ fault-tolerant system under different error rates.Comparing with ARQ(Automatic Repeat Request)system,the average maximum effective throughput of the sifted completed linear code HARQ system is increased by 6.6% under different interconnect crosstalk error rates.In order to further improve the transmission efficiency,an HARQ fault-tolerant system algorithm based on effective bandwidth analysis is proposed.Through this algorithm,the expected transmission bits of redundant code are reduced,and the effective throughput is further improved.Comparing with(8,4)Hamming code data transmission system commonly used in the industry,the effective throughput of the HARQ fault-tolerant system proposed is increased by 36.59% under 1% crosstalk error rate.