Work

Energy-Efficient Deep Learning through Memristive Neuromorphic Synapses: A Hardware Implementation Study

Year: 2025
Type: article
Abstract: Advances in artificial intelligence and machine learning, especially in deep learning, have driven rapid adoption across various fields. However, the high computational demands and extensive data processing needs of these algorithms pose major energy efficiency challenges for traditional Von Neumannbased computing systems. These issues are compounded by the slowing scalability of semiconductor technology and the inefficiencies of parallel processing in multi-core architectures. To address these limitations, neuromorphic computing systems which unify memory and processing at the hardware level have emerged as a promising solution for energy efficient AI. Among their key components, memristive devices stand out by mimicking biological synaptic behavior with extremely low power consumption, allowing for physical representation of synaptic weights in neural networks. This study explores the hardware implementation of memristive synapses in deep neural networks. While memristive systems may have longer training times compared to software-based convolutional neural networks, they achieve competitive accuracy (up to 90%) using gradient descent optimization methods, all while consuming around 100,000 times less energy. This dramatic improvement in energy efficiency makes memristive technology a leading candidate for both current and future sustainable AI systems.
Language: English
Open Access status: diamond
OpenAlex ID: W7217113477
FWCI:
Cites: 18
Cited by:
Related to: 9