Simulating Spiking Neural Networks Based on SW26010pro

The spiking neural network (SNN) simulators play a significant role in modeling neural systems and the study of brain function. Currently, many simulators using CPU or GPU have been developed. However, these simulators usually show low efficiency, resulting from the random synaptic connections, random spiking events, and random synaptic delay and plastic properties in the SNN models. To overcome the problem of random memory access etc., a new simulator named SWsnn is developed based on a new Chinese processor, SW26010pro. SW26010pro consists of six core groups (CGs), and each CG has 16 MB of local direct memory (LDM) (similar to L1/L2 cache), which is enough to store neuron data for a long time. By rearranging the synaptic and neuron data, SWsnn ensures that most of the random memory access occurs in the neuron data, and the reusability of the LDM is improved obviously. The results illustrate that the proposed SWsnn runs faster than other GPU-based simulators.