Fault-Tolerant-Driven Clustering for Large Scale Neuromorphic Computing Systems

Memristive crossbar-based neuromorphic computing systems (NCS) have been extensively investigated and applied to the neural networks due to the fast computation and low design cost. In most applications, neural networks are large and sparse, which violate the size limitations and high-density connections provided by the memristive crossbars. Besides, stuck-atfaults (SAFs) in the memristor devices significantly degrade the computing accuracy of NCS. In this paper, we propose a faultdriven clustering framework for NCS based on a set of unique size memristive crossbars, with consideration of both hardware cost and mapping success rate. First, in order to group the input neurons connected to different output neurons, we design a METIS-based clustering method by redefining the distance metric, to speed up the large-scale neural network partitioning and improve the fault tolerance of the memristive crossbar-based NCS, then map the synapses to a set of unique size crossbars. Second, a half transposition method is developed to address the extremely asymmetric clusters. The simulation results show that the proposed fault tolerance-aware clustering algorithm not only improves the mapping success rate and the hardware cost but also achieves speed-up. For example, for a large-scale neural network with four million synapses, the proposed framework can complete the algorithm in one hour.