Energy-and Area-efficient Fe-FinFET-based Time-Domain Mixed-Signal Computing In Memory for Edge Machine Learning

In this work, for the first time, ferroelectric (FE)-FET is proposed to implement time-domain (TD) computing in memory with nonvolatility (nvCIM) for both multiply-accumulate (MAC) operation and activation function with the record highest area and energy efficiency. Benefiting from the three-terminal transistor structure, the fabricated FeFET based on 14nm-node FinFET can function as both a nonvolatile element for weight storage and a controllable switch for neural network inputs modulation simultaneously, thereby realizing local multiplication of MAC in the proposed FE delay unit with only 2T-1FeFET. The novel dual-edge operation is also demonstrated for TD MAC, enabling further energy efficiency improvement. Furthermore, by utilizing the physics of time-dependent gradual polarization switching, FeFET as the activation function for TD nvCIM with additional memory trace behavior is experimentally demonstrated with only single transistor. Based on the proposed FE-based TD nvCIM design, high-accuracy pattern recognition and accelerated deep reinforcement learning are also demonstrated with scaled voltage of 0.5V, providing a promising area- and energy-efficient mixed-signal neural network for edge AI.