Fault Tolerance-Driven Synapse Mapping On Memristive Crossbars

The emerging memristive crossbars are promising in building Neuromorphic Computing Systems(NCS)due to the advantages of non-linearity,non-volatility,low power consumption,and high integration.But the most neural networks are sparse,which contradicts the high-density connections provided by memristive crossbars.Additionally,due to the immature manufacturing process,the currently available memristive crossbars are small.And there is a certain proportion of defects and faults in the memristive crossbars.Consequently,this thesis proposes a fault-tolerant clustering framework for the memristive crossbars-based large-scale neural networks.The main contributions of this thesis are as follows.This thesis designs a fault-tolerant mapping for the memristive crossbar-based neuromorphic computing systems.The mapping includes a fault-tolerant-driven neuron spectral clustering algorithm,in which 0-1 matrices are used to represent the synaptic connection relationship between neurons.Through the neuron distances defined based on Hamming distance,we guide the spectral clustering to form relatively sparse synaptic sub-matrices with maximized Hamming distances between rows,which increase the probability of avoiding faulty memristors by rearranging the matrix rows.The bipartite matching-based heuristic is used to complete the mapping of the synaptic matrix on the defect memristive crossbars.To achieve versatility on a set of memristive crossbars with fixed dimension size,we design a fault-tolerant clustering method for large-scale neural networks.The synaptic connections are grouped into a set of clusters with similar dimension size through a fault-tolerant driven clustering based on a classic partitioning method,which greatly increases the sizes of the neural network that can be processed within an acceptable runtime.We maximize the synapse connection sub-matrices under a certain degree of redundancy by considering the probability of successful mapping of the synaptic sub-matrix on a fixed-shape memristive crossbar during clustering,which improves the utilization of the memristive crossbars.A half-transpose method is proposed to reduce the gap between rows and columns of an asymmetric matrix,which increases the size of the sub-matrix that can be accommodated in a given memristive crossbar,and reduce the number of memristive crossbars.Finally,a Monte Carlo simulation experiment was carried out on the frame,and floor planning was carried out according to the clustering and mapping results.The experimental results under the fixed failure rate of 10.79%show that the FTNCS framework can achieve a mapping success rate of 96.69%,and the FTTCS-NCS mapping success rate is as high as 99.52%.For a large-scale neural network with millions of nodes,FTDC-NCS reduces the running time from more than 6h to about 1500 seconds,while improving hardware utilization.