Efficient Large Scale Neural Network Acceleration With 3-D FeNOR-Based Computing-in-Memory Design

In this work, we introduce and experimentally demonstrate a 3-D stacked ferroelectric nor (FeNOR) memory, featuring a back-end-of-line (BEOL) zinc oxide (ZnO) channel, and a metal-ferroelectric–metal-insulator5 semiconductor (MFMIS) unit cell. The main contributions of this work are as follows: 1) enhanced memory window (MW) and high on/off ratio: The MFMIS architecture in 3-D FeNOR enables a tunable and large MW (~4 V), as well as an on/off ratio ( ${I}_{\text {on}}$ / ${I}_{\text {off}}$ ) of six orders of magnitude; 2) low operation voltage and high endurance: The integration of ferroelectric materials allows for low operation voltages (~4 V) and excellent endurance ( $10^{{7}}$ cycles); 3) efficient neural network implementation: Leveraging the 3-D FeNOR structure, we further develop VGG-16 and ResNet-50 convolutional neural networks that achieve high prediction accuracy, decent area efficiency, and low power consumption. The emergence of 3-D FeNOR technology positions ferroelectric devices as a highly promising candidate for computing-in-memory (CIM) applications.