Nonlinear Weight Quantification for Mitigating Stress Induced Disturb Effect on Multilevel RRAM-Based Neural Network Accelerator

The RRAM-based array is one of the most promising core functional primitives to accelerate the inference process of neural networks. However, the stress-induced disturbance can cause a significant accuracy drop during inference process where input vectors with different voltage levels are fed to the device. This kind of disturb can hardly be avoided by optimizing the fabrication process. Here, we investigate this phenomenon based on TaOx-based devices with different electrodes. The results indicate that the stress-induced disturb mainly appears in the intermediate resistance states when the voltage is applied on the device. The simulation result by COMSOL reveals the relationship between read disturb and electric field. Therefore, we propose a nonlinear weight quantification method to mitigate read disturb effect on inference accuracy by reducing the number of devices in intermediate resistance states. The simulation results based on the fully-connected networks for MNIST recognition indicate that stress disturb phenomenon can be well suppressed by nonlinear weight quantification compared with the conventional linear quantification method, which will advance the application of the RRAM-based accelerator.