A Low Power Spiking Neural Network Accelerator on FPGA for Real-Time Edge Computing

Machine learning (ML) has become a cornerstone for innovative applications in the Internet of Things (IoT), such as intelligent security and autonomous systems. However, cloud-based ML solutions introduce latency, high power consumption, and privacy risks, making them unsuitable for real-time, resource-constrained IoT applications. Spiking Neural Networks (SNNs), with their low power consumption and biologically inspired computational efficiency, offer a promising alternative to traditional artificial neural networks (ANNs) in edge computing environments. Despite their potential, current SNN hardware accelerators are limited by scalability and power inefficiencies. This paper presents a co-optimization approach for both software and hardware to enhance the performance of SNNs in IoT applications. On the software side, we quantize weights to 8-bit fixed-point values and merge batch normalization layers to minimize computational overhead. On the hardware side, we design a frame-based SNN accelerator based on Leaky Integrate-and-Fire (LIF) neuron model, which reuses computational units and memory to improve resource efficiency. We also introduce a four-dimensional parallel computing architecture to enhance data processing speed and apply a bitmap-based sparse encoding method for weight storage, reducing memory and bandwidth requirements. Experimental results demonstrate that the proposed design significantly improves both the energy efficiency and real-time processing capabilities, which is more suitable for deployment in resource-constrained IoT devices.