Fault-Tolerant Computing Based On RRAM Crossbar

Neural Network is becoming more and more popular in a wide range of domains,such as data mining,image classification,face detection.Neural network algorithms always consume large amount of computing resources,resulting in substantial power efficiency.Meanwhile,the hardware should have both computing ability and low power consumption to ensure the realtime and flexibility of the neural network algorithms.However,the CMOS-based techniques meet a constraint of the scale and a ”memory wall” bottleneck.RRAM crossbar is a computing device with a bunch of ”memristor”.This device can naturally solve the matrix multiplication in the neural network applications.Compared with traditional CMOS technique,this device has high energy-efficiency.RRAM crossbar,at the same computation ability level of CMOS-based device,has a smaller chip area.However,RRAM crossbar suffers from resistance variation and stuck-at faults,which influence the robustness and the test rate of neural network applications.This paper will analysis the features of RRAM crossbar and neural network algorithms.By leveraging the inherent self-healing capability of the neural-network,we will propose an accelerator-friendly neural network training method to solve the resistance variation and stuckat fault in software level.