Novel Negative-Feedback Method for Writing Variation Suppression in FeFET-Based Computing-in-Memory Macro

In this work, a novel one-shot negative-feedback writing method is proposed for suppression of writing variation in ferroelectric FET (FeFET) based analogy computing-in-memory (CIM) macro for high-accuracy artificial neural network (ANN). By utilizing the source-voltage negative-feedback mechanism and the voltage and time dependent multi-domain switching dynamics, a source-follower writing structure with an FeFET and a NMOS is proposed and simulated to reduce variation of target programming V TH for FeFET with even multi-level state, revealing significant decrease of FeFET synaptic conductance variation. Furthermore, based on the proposed FeFET writing variation suppression method, the CIM macro of FeFET array is demonstrated to improved accuracy of image recognition ANN by 40% compared with direct writing, showing great potential for high-accuracy neural network system.