Experimental Demonstration of Resonant Adiabatic Writing and Computing in Ferroelectric Capacitive Memory Array for Energy-Efficient Edge AI

In this work, a novel resonant adiabatic (RA) capacitive computing-in-memory (C-CIM) with non-volatility is proposed and experimentally demonstrated based on 3D stackable ferroelectric (FE) capacitor (FeCap) array, enabling both writing and multiply-accumulate (MAC) computing with record-high energy efficiency. For the first time, by leveraging the capacitive nature of hafnia-based FeCap, both capacitance state writing and non-destructive readout for computing are implemented with incorporated resonant energy recovery and adiabatic charging in FeCap based C-CIM. The experimentally measured power consumption indicates that the proposed ferroelectric RA C-CIM can significantly reduce dynamic power loss inherent in conventional pulse-based C-CIM, along with suppressing resistive heating dissipation. Further co-designing FeCap-based RA C-CIM with layer normalization algorithm, high-accuracy pattern recognition is demonstrated with $35\times/8.95\times$ energy efficiency improvement compared with conventional resistive and capacitive CIM respectively, providing a promising solution for energy-efficient edge AI.