Binary Convolutional Neural Network Based on 3d Field-Programmable Ferroelectric Diode Array

Abstract

Ubiquitous artificial intelligence (AI) applications call for increasing computing power and large communication bandwidth, which are challenging to meet given the diminishing benefits from traditional planar CMOS scaling. Three-dimension (3D) integration of integrated circuits (IC) promises to achieve further performance improvements at reduced power and smaller footprint. Here, we report a 3D array of 8 layers of field-programmable ferroelectric diode (FPD) for binary convolutional neural network (BCNN) in energy-efficient AI applications. The convolutional kernels are duplicated and stored in the pillars in 3D FPD array to providing high computation parallelism. An AND-based logic using FPD is proposed to realize the XNOR function in BCNN. As a proof-of-concept demonstration, a BCNN is constructed based on the 3D FPD array for MNIST test and qualified to present the advantages of our proposal in 28nm logic process design kit.