Abstract Processing wireless RF signals using analog electromagnetic (EM) wave-based neural networks enables energy efficiency and parallelism by integrating sensing, memory, and computation, avoiding analog-to-digital conversions (ADCs) and the von-Neumann bottleneck. Yet, a notable challenge remains: the absence of compact and programmable building blocks for EM wave-based neural networks. To overcome this limitation, we propose a piezoelectric surface acoustic wave (SAW) memristor that integrates an Ag/SiO 2 /Au memristor with an acoustoelectric phase shifter. Operating at shorter wavelengths than EM waves, it offers a compact footprint and encodes tunable neural network parameters via nonvolatile programmability and the acoustoelectric effect. A proof-of-concept SAW memristor neural ne…
Abstract Processing wireless RF signals using analog electromagnetic (EM) wave-based neural networks enables energy efficiency and parallelism by integrating sensing, memory, and computation, avoiding analog-to-digital conversions (ADCs) and the von-Neumann bottleneck. Yet, a notable challenge remains: the absence of compact and programmable building blocks for EM wave-based neural networks. To overcome this limitation, we propose a piezoelectric surface acoustic wave (SAW) memristor that integrates an Ag/SiO 2 /Au memristor with an acoustoelectric phase shifter. Operating at shorter wavelengths than EM waves, it offers a compact footprint and encodes tunable neural network parameters via nonvolatile programmability and the acoustoelectric effect. A proof-of-concept SAW memristor neural network was experimentally demonstrated on a vector classification task, achieving 91.7% accuracy on par with software while reducing footprint by 10 5 times versus EM systems and energy consumption by 37 times versus digital systems. This work paves the way for compact, energy-efficient RF signal processing at the edge.