Memristive synapses as building blocks of neuromorphic artificial intelligence (AI) hardware

Abstract

The rapid advancement of artificial intelligence (AI) has significantly expanded the use of machine learning and deep learning, yet traditional Von-Neumann architecture-based computers struggle with the energy demands of these technologies. This challenge is making worse by the limitations of semiconductor miniaturization and the inefficiencies of multicore architectures. Neuromorphic computing, leveraging hardware-based neural networks, offers a potential solution by integrating computation and memory processes more efficiently. Recent developments in memristive devices, which emulate biological synapses with reduced power consumption, have shown promise in addressing these issues. This paper explores the use of memristive synapses as components in AI hardware for deep neural networks (DNNs). Despite longer training latencies compared to conventional CNN software, memristive models achieve up to 9 2 % accuracy using the Adam optimization method, while consuming approximately 100,000 times less energy. This significant reduction in energy usage highlights the potential of memristive devices for energy-efficient AI applications, making them a compelling alternative for current and future computing needs.