ScholarWorks@성균관대학교

초록

Recent developments in deep learning have significantly enhanced image classification capabilities and established a new performance standard for computer vision applications. However, these advancements are constrained by the high-energy demands of conventional von Neumann computing architectures. We propose an in-memory vision transformer (ViT) system that utilizes synaptic ferroelectric thin-film transistor (FeTFT) arrays combined with a high-mobility indium-gallium-zinc oxide (IGZO) channel to address this limitation. The in-memory ViT system facilitates parallel operations through vector-matrix multiplication (VMM) with a minimal hardware burden, thereby significantly reducing energy consumption while maintaining a high performance. The synaptic IGZO FeTFT array exhibits high mobility, precise conductance modulation, and robust endurance over extensive program/erase cycles. Precise weight-transfer capabilities and reliable VMM operations are demonstrated using synaptic IGZO FeTFT arrays. The proposed in-memory ViT system achieves an exceptional accuracy of approximately 94 % on the CIFAR-10 dataset even after more than 107program/erase cycles. A reliable and energy-efficient in-memory ViT system comprising the use of synaptic IGZO FeTFT arrays provides a viable solution for the energy limitations of advanced computer vision applications.

키워드