Fault tolerance in memristive crossbar-based neuromorphic computing systems

Qi Xu, Song Chen, Hao Geng, Bo Yuan, Bei Yu, Feng Wu, Zhengfeng Huang

Abstract: Highlights•Given a sparse neural network, a fault tolerance-aware hierarchical clustering method is proposed to partition weight connections into a set of clusters.•For connection matrix of each cluster, a non-linear programming is proposed to determine suitable size of the mapped memristive crossbar, considering both hardware cost and successful mapping rate.•Fault-tolerant mapping is formulated as an integer linear programming (ILP) to map connection matrix to memristive crossbar.•An efficient matching-based heuristic algorithm is further proposed to speed-up the ILP process.•A Monte Carlo simulation is exploited to evaluate the performance of the fault tolerant synapse mapping framework on different benchmarks.