(A) framework for mitigating non-idealities of ReRAM-based DNN accelerators

Abstract

Processing in-memory (PIM) approaches have emerged to address the memory wall problem of the conventional Deep Neural Network (DNN) accelerators. Resistive RAM (ReRAM) is regarded as a proper platform for PIM. Owing to the current-sum manner of the ReRAM array, ReRAM-based DNN accelerators achieve energy-efficient computation. Despite the promising computation ability, various non-idealities hinder practical usage of the ReRAM-based DNN accelerator. To ensure reliable computation results, a framework to mitigate the non-idealities of the ReRAM-based DNN accelerator is presented in this thesis. Among several non-idealities, major non-idealities (thermal, Stuck-At-Fault, and read disturbance problems) are intensively solved. In the case of the thermal problem, weights of DNN are stored in the array in a way that increases the temperature the least. When the Stuck-At-Fault problem occurs, the weight is stored to have a minimum error from the original weight value to prevent the accuracy degradation of DNN. In addition, the read disturbance problem caused by the drift of the weight value stored in the cell is solved through a selective refresh approach. Through this paper, the reliability of the ReRAM-based DNN accelerator is improved.