Neuromorphic Circuit Design Using Memristive Devices for Energy-Efficient Edge Artificial Intelligence Applications

Abstract

Edge Artificial Intelligence (AI) deployment is fundamentally constrained by the memory wall and energy inefficiencies inherent to traditional von Neumann computer architectures. As deep learning models expand in complexity, executing real-time inference and on-device learning at resource constrained edge nodes requires a radical departure from conventional digital CMOS paradigms. Neuromorphic computing, which mimics biological neural architectures by co-locating computation and storage, presents a transformative alternative. Memristive devices—including Resistive Random Access Memory (RRAM), Phase-Change Memory (PCM), and Spin-Transfer Torque Magnetic RAM (STT-MRAM)—have emerged as prime hardware enablers due to their non-volatility, high integration density, analog tuneability, and nanosecond switching kinetics. This paper presents a comprehensive review of neuromorphic circuit designs leveraging memristive devices for energy-efficient edge AI applications. We analyze fundamental memristor device physics, crossbar array architectures for matrix-vector multiplication (MVM), and bio-inspired spiking neural network (SNN) hardware implementations. Furthermore, we examine key non-ideal device effects—such as resistance variation, conductance drift, non-linearity, and sneak-path currents—and evaluate circuit-level and algorithmic mitigation strategies. Quantitative performance benchmarking demonstrates that memristive neuromorphic accelerators achieve up to 2–3 orders of magnitude higher energy efficiency (TOPS/W) compared to digital GPUs and edge TPUs. Finally, we highlight current technical bottlenecks, ethical considerations, and future avenues toward monolithic 3D integration and bio-hybrid edge AI devices.

KEYWORDS: Neuromorphic Computing, Memristor Crossbar Arrays, Edge Artificial Intelligence, Resistive RAM (RRAM), Spiking Neural Networks (SNN), Energy-Efficient Hardware, In-Memory Computing.