Accelerated Optimization for Simulation of Brain Spiking Neural Network on GPGPUs

Abstract

As the application scenarios for large-scale spiking neural networks (SNN) increase, efficient SNN simulation becomes more essential. However, simulating such a large-scale network faces expensive overhead in terms of computation and communication, especially for high firing rates. To address this problem, we propose an effective accelerated optimization method for simulating SNN on GPGPUs, which simultaneously takes into account workload balancing and communication overhead. We design a workload-oriented network partition algorithm to minimize the number of external synapses and ensure workload balance. Additionally, we propose spike synchronization optimization by incorporating fine-grained scale, data compression, and full-duplex communication. This optimization aims to achieve lower communication overhead and better performance improvement. Furthermore, to avoid thread warp divergence, we assign an entire thread block for each neuron without collecting information on fired neurons in the spike propagation phase, which simplifies the execution flow and enhances performance. Experimental results demonstrate that our simulator can achieve up to $1.31{\times }{\backsim }6.74{\times }$ speedup for SNN with different configurations, and the efficiency is improved by $40.21\%{\backsim }51.11\%$ compared with the state-of-the-art methods.