A Novel Ferroelectric FET based Multibit Content Addressable Memory with Dynamic and Static Modes for Energy-Efficient Training and Inference

In this work, a novel multibit content addressable memory (CAM) based on complementary ferroelectric FET (FeFET) with split-gate, enabling both dynamic and static modes, is proposed and experimentally demonstrated for the first time. For dynamic CAM (DCAM), by utilizing the complementary polarity engineering and the non-destructive readout of n/p-type FeFET, only single storage capacitor and single search signal are required for dynamic entry storage and dynamic query searching, enabling reduced hardware cost and power consumption. Moreover, by further utilizing the splitgate engineering of FeFET, the proposed dynamic CAM supports additional in-situ multibit expansion for high-accuracy training. For static CAM (SCAM), by utilizing the non-volatile multilevel ferroelectric polarization for static entry storage, the proposed CAM can also support static binary and multibit CAM modes with single static search signal. Based on the proposed dual-mode CAM design, where the DCAM mode with update-frequent capability for training and the SCAM mode with long retention for inference, the image recognition tasks are demonstrated with high accuracy and significant improvement in speed and energy-efficiency over GPU, along with the lowest write and search energy consumption compared with other state-of-the-art CAM designs, showing its great potential for area-and energy-efficient machine learning.