Low Power Non-volatile Memory Technology for Edge AI Application

Abstract

For the edge AI applications, emerging memories such as RRAM, MRAM are potential candidates to address these concerns. However, these memories require high write energy consumption due to a current driven switching. Recently, a hafnium oxide based ferroelectric technologies have been strongly focused. We experimentally demonstrate 1T1C FeRAM memory array for the first time.
Deep neural network (DNN) inference for edge AI requires low power operation. We demonstrated it by implementing massively parallel matrix-vector multiplications (MVM) in the analog domain on highly resistive memory array. We propose a 1T1R compute cell (1T1R-cell) using FeFET and tunneling junction of MΩ resistor (MOR) for analog in-memory computing (AiMC).