Abstract:

To address the problem of high GPU memory requirements in large-scale spiking neural network simulation, a dynamic loading simulation method for large-scale spiking neural networks is proposed. This method uses data movement at the sub-network granularity and utilizes the host memory as a larger memory pool to reduce the limitation of GPU memory on the model simulation scale, enabling large-scale spiking neural network simulation on a single GPU computer. The pipeline acceleration technique is adopted to reduce the impact of data movement on simulation speed. The simulation of a million-scale neural network is achieved in a single GPU experimental environment, which solves the problem of insufficient memory during spiking neural network simulation.

Key words: brain-inspired computing, spiking neural network, neuron, synapse, simulation