Symmetric n/p Schottky Barrier Modulation for Precision-Configurable Neural Network.

Abstract

Achieving both high precision and efficiency in edge devices presents a notable challenge in neuromorphic computing. Conventional neuristors typically operate with fixed computational precision, forcing a trade-off between accuracy and efficiency when addressing tasks of varying complexity. To overcome this limitation, we propose a Schottky barrier neuristor that combines high-efficiency nonlinear logic with high-precision linear operations within a single device. A distinctive global bottom gate modulates the Schottky barrier, maintaining a linear relationship between gate voltage and transconductance. Furthermore, electrostatic doping-induced image force effects, alongside an optimized source-drain work function, enable uniform and symmetric n-/p-type modulation, enhancing the driving capability. This innovative design supports the development of reconfigurable digital-analogue units, requiring only one-fifth the number of devices needed for nonlinear functions compared to silicon. Simulations demonstrate that an accelerator based on this device achieves 98.3% accuracy and an energy efficiency of 1359.62 TOPS/W.