Intra-array Non-Idealities Modeling and Algorithm Optimization for RRAM-based Computing-in-Memory Applications
Chenyang Zhao,Jinbei Fang,Jingwen Jiang,Zhiwang Guo,Xiaoyong Xue,Xiaoyang Zeng
DOI: https://doi.org/10.1109/asicon52560.2021.9620289
2021-01-01
Abstract:In recent years, Computing-in-Memory (CIM) has shown attractive advantages over CPU/FPGA/ASIC in terms of area, energy efficiency and latency for neural network acceleration in edge applications. Among them, RRAM-based CIM accelerators has gained lots of attentions with fast read access, and low leakage power. However, the inherent non-idealities of RRAM devices, such as read non-linearity and read variation will directly impact the hardware performance. So, based on the mainstream RRAM mathematical model, we first establish 9 models of RRAM cells with different LRS/HRS combinations. Then we analyze the circuit simulation results to model the non-idealities mentioned above. A non-ideality-aware algorithm is proposed based on these models, which are brought into the training process of the neural network, to improve the reliability of the accelerator. A multi-layer perceptron (MLP) with one input layer, two hidden layers and one output layer is used to perform recognition of MNIST for evaluation purpose. Given the read voltage ranged from 0 to 1.0V, read variation σ =20%, LRS/HRS = 200KΩ/2MΩ, the proposed algorithm can improve the accuracy by 5.42%.