Design of Ferroelectric FET-Based Capacitive-Coupling Computing-In-Memory For Binary Neural Networks

Ferroelectric FETs (FeFETs) have been applied to computing-in-memories (CIMs) for binary neural networks (BNNs) because of the advantages of ultra-low energy cost, while suffering from large output variation or write disturb. This work proposes a novel FeFET-based CIM for BNNs, reducing the output variation by exploiting charge-domain scheme and avoiding write disturb by exploiting additional access transistors. Simulation results show that compared with conventional FeFET-based current-domain design, the proposed design can improve accuracy by 20% on MNIST by reducing the output variation. Compared with previous SRAM-based and RRAM-based CIM for BNNs, the proposed design can reduce energy cost by 54% and 98%, showing its great potential for edge AI applications.