Design and Evaluation of Neuromorphic Computing Hardware for Energy-Efficient Edge AI and Advanced Electronics

Abstract

To overcome this problem of the traditional digital processor in terms of energy and latency, neuromorphic computing developed as a brain-based architecture to process the edge artificial intelligence (AI). The paper proposes a design and analysis of neuromorphic hardware spiking neural network proposals in CMOS and memristor technologies towards real-time, low-power, AI inference. Performance benchmarks were simulation along circuit and system levels (e.g. Cadence, LTSpice, Brian2) and in terms of energy consumption, latency, accuracy rate of classification and chip area. Data results indicate that the neuromorphic systems are capable of realizing an energy-saving up to 60 percent and latency improvement of more than 50 percent greater than the traditional convolutional neural networks (CNNs), with an insignificant accuracy compromise of about 5-7 percent only. Works based on memristors demonstrated a better energy efficiency and integration density, although CMOS-based works were more stable in time. These results represent the feasibility of using neuromorphic hardware in the next generation edge field of autonomous sensing, robotics and embedded electronics. The paper ends off with future direction towards scalable integration, on chip learning capabilities and fabrication progressions.