Design Of Efficient AI Accelerator Using Spiking Neural Network

摘要

To design and implement energy-efficient AI accelerators using Spiking Neural Networks (SNNs) through simulation, focusing on Network optimization architecture to maximize effectiveness and minimize power consumption. The developed architecture is designed for high-performance AI solutions, suitable for applications in edge computing, robotics, autonomous vehicles, and neuromorphic computing. SNNs, inspired by biological neurons, rely on spike-based communication and event-driven processing, which allows them to operate asynchronously, significantly reducing energy usage by processing only when spikes occur. The proposed SNN accelerators explores the design of custom hardware components, such as scalable neuron and synapse modules, optimized memory access patterns, and the integration of learning mechanisms like Adaptive learning through Spike-Timing-Dependent The SNN accelerator is proposed using ResNet, deep learning model a type of convolutional neural network (CNN). Previously SNN accelerators are produced using LeNet and pipelining methodology is used for storing the spike signals. Adopting ResNet in SNN includes following steps like Replacing Rectified Linear Unit (ReLU) activations with spiking neurons (Leaky Integrate-and-Fire (LIF) or Spike Response Model (SRM), Adjusting batch normalization and weights to fit spike-based computation, Encoding input data into spike trains using rate coding or temporal coding. This work adopts ResNet for processing and reservoir computing for spike signal storage. The implementation of an energy-efficient Spiking Neural Network (SNN) accelerator on the Xilinx PYNQ FPGA demonstrates significant improvements in both power consumption and computational speed. The computational speed and power efficiency is improved in proposed SNN. Compare to the previous SNN methodology there is slight increase in increase but power has been reduced to 25 % and improvement in the computational speed of 30% is achieved. The SNN architecture’s adaptability, coupled with its low power consumption and real-time learning capability, positions it as a promising solution for future intelligent systems that need to balance performance, sustainability, and scalability.