Optimized Programming Scheme Enabling Symmetric Conductance Modulation in HfO₂ Resistive Random-Access Memory (RRAM) for Neuromorphic Systems

The asymmetric conductance modulation of resistance random access memory (RRAM) is one of the crucial obstacles of the application as artificial synapses for online training. In this work, an optimized programming scheme is proposed to overcome the obstacle. In the optimized programming scheme, the RRAM is grounded through the array parasitic capacitor during the depression process. By harnessing the charging effect of parasitic capacitor, the voltage applied on the RRAM-based synapse is self-regulated according to its conductance. Meanwhile, the linearity of depression process can be tuned by adjusting the operation pulse shape to realize symmetric conductance modulation. The proposed programming scheme is verified by experiment and simulation with a 4 kb RRAM array. Finally, the optimized programming characteristics are applied in the simulation of the training of a VGG-like neural network for the classification of CIFAR-10 dataset. The results show that with the linearity of 6, 88.93% training accuracy is achieved, showing 17.3% accuracy improvement without any overhead either in fabrication process or hardware design cost.