Implementation of Discrete Fourier Transform using RRAM Arrays with Quasi-Analog Mapping for High-Fidelity Medical Image Reconstruction

Abstract

In this paper, we report the first experimental implementation of discrete Fourier transform (DFT) on analog resistive switching random-Access memory (RRAM) arrays with computing-in-memory (CIM). Considering the features of transform matrix, we developed a novel conductance mapping strategy, namely quasi-Analog mapping (QAM), to realize high-precision mapping by taking advantage of the analog switching characteristics of our RRAM. Based on the RRAM-based DFT models, high-fidelity medical image reconstruction was further demonstrated, achieving a software-comparable peak signal-To-noise ratio (PSNR) of 26.1 dB. Compared to the commonly used quantized mapping (QM), QAM enhanced the image reconstruction quality and showed strong robustness to RRAM read noise. RRAM-implemented DFT also achieved ~128× higher energy efficiency than CPU. This work provides a general strategy for using RRAM array with CIM feature to accelerate signal processing algorithms.