[Paper] Analog AI Accelerator using Current-Programming Memory with IGZO TFT

Abstract

This paper presents an analog AI accelerator utilizing current-programming memory based on Indium Gallium Zinc Oxide (IGZO) thin-film transistors (TFTs). Leveraging current-copier pixel circuits from organic light-emitting display (OLED) technology, the architecture performs high-precision multiply-accumulate (MAC) operations while serving as analog memory. IGZO’s ultra-low off-current facilitates long-term weight retention, significantly reducing power consumption by removing the need for frequent refreshes. To address analog non-idealities like feedthrough effects, an n-channel source follower is integrated, achieving output current deviations as low as 0.39% (corresponding to an equivalent precision of ∼8 bits over full scale). Furthermore, Monte Carlo simulations confirm robust parameter mismatch immunity, and system-level simulations on the MNIST dataset using the circuit-extracted error distribution validate minimal accuracy degradation (< 0.4%, achieving 97.4%). This design provides a scalable, energy-efficient solution for sustainable edge AI computing.