Solving All Eigenpairs With Resistive Memory-Based Analog Matrix Computing Circuits

doi:10.1109/TCSI.2025.3628178

Return

Solving All Eigenpairs With Resistive Memory-Based Analog Matrix Computing Circuits

2025-11-11

PRE

DOI：10.1109/TCSI.2025.3628178

Original

Original request for help

Save

Abstract

En 中文

Eigenpair computation is fundamental in machine learning and signal processing but faces $\mathbf {O(n^{3})}$ complexity and data movement limits on digital platforms. Resistive random-access memory (RRAM)-based analog matrix computing (AMC) offers a low-latency, in-memory alternative, enabling $\mathbf {O(1)}$ matrix operations by encoding matrices as conductance and exploiting circuit laws. Previous AMC eigenvector circuit can compute the dominant eigenvector but requires prior knowledge of the corresponding eigenvalue, and is incapable of solving for all eigenpairs. This work introduces an approach that decomposes eigenpair computation into matrix inversion task and eigenvalue determination by sweeping, enabling two RRAM-based AMC circuit designs: the modified-INV circuit for dominant eigenpair computation and the modified-GINV circuit for all eigenpairs computation. Applied to principal component analysis, the modified-GINV circuit generated principal components for the Iris dataset and extracted eigenfaces for image reconstruction —both closely matching the theoretical values. Moreover, we investigated the impact of non-ideal factors, including RRAM device variation, operational amplifier input offset voltage, and interconnect resistance, as well as power consumption, providing valuable guidance for evaluating circuit performance.

Keywords:

Analog computing

eigenvalue

eigenvector

resistive memory

principal component analysis