Analog In-memory Computing in FeFET-based 1T1R Array for Edge AI Applications
D. Saito, Toru Kobayashi, H. Koga, N. Ronchi, Kaustuv Banerjee, Y. Shuto, J. Okuno, Kazuo Konishi, Luca Piazza, A. Mallik, Jeroen Van Houdt, Masahiko Tsukamoto, K. Ohkuri, T. Umebayashi, Tohru Ezaki
Abstract
Deep neural network (DNN) inference for edge AI requires low-power operation, which can be achieved by implementing massively parallel matrix-vector multiplications (MVM) in the analog domain on a highly resistive memory array. We propose a 1T1R compute cell (1T1R-cell) using a ferroelectric hafnium oxide-based FET (FeFET) and TiN/SiO <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> tunneling junction of MΩ resistor (MOR) for analog in-memory computing (AiMC). The MOR exhibited a tunneling current behavior and MΩ resistance. A 1T1R-cell array-level evaluation was also performed. A random access for writing with low write disturbance scheme was confirmed from the summation-DC-current output, and binaries were successfully classified into “T” and “L.” Based on the experimental results of our proposed 1T1R-cell, we obtained a state-of-the-art energy efficiency of 13700 TOPS/W including the periphery. Furthermore, we confirmed that a high inference accuracy can be obtained with our low-resistance-variability 1T1R-cell with a properly trained model.