SWsnn: A Novel Simulator for Spiking Neural Networks

Area of Science:

• Computational neuroscience
• Artificial intelligence
• High-performance computing

Background:

• Spiking Neural Network (SNN) simulators are crucial for modeling brain functions and advancing AI.
• Existing CPU and GPU-based simulators face challenges with memory access time due to the nature of SNN simulations.
• The Sunway SW26010pro processor offers high-speed local data memory (LDM) potentially beneficial for SNN simulations.

Purpose of the Study:

• To develop a novel Spiking Neural Network (SNN) simulator, SWsnn, optimized for the Sunway SW26010pro processor.
• To address the memory access time limitations of traditional SNN simulators.
• To enable larger-scale SNN simulations through efficient hardware utilization.

Main Methods:

• Developed the SWsnn simulator utilizing the Sunway SW26010pro processor and its 16 MB Local Data Memory (LDM) per core group.
• Implemented a simulation computation strategy based on the Sunway processor's large shared model.
• Created a multiprocessor version of SWsnn to facilitate large-scale simulations.

Main Results:

• SWsnn demonstrated superior performance compared to mainstream GPU-based simulators for specific neural network scales.
• The simulator effectively utilized the high-speed LDM of the Sunway SW26010pro processor for SNN simulation tasks.
• The multiprocessor version successfully achieved larger-scale Spiking Neural Network simulations.

Conclusions:

• The SWsnn simulator offers a significant performance advantage for Spiking Neural Network simulations on the Sunway SW26010pro architecture.
• This development paves the way for more extensive and complex neural network modeling and brain research.
• The efficient use of local data memory is key to overcoming performance bottlenecks in SNN simulation.