Brian2CUDA: Flexible and Efficient Simulation of Spiking Neural Network Models on GPUs

Area of Science:

• Computational neuroscience
• Scientific computing

Background:

• Graphics processing units (GPUs) offer significant computational acceleration but require specialized low-level programming.
• The Brian simulator facilitates high-level definition of spiking neural networks (SNNs) for efficient CPU execution.
• Access to GPU acceleration for SNN simulations is limited for researchers without low-level coding expertise.

Purpose of the Study:

• To introduce Brian2CUDA, an open-source software extending the Brian simulator with a GPU backend.
• To enable efficient simulation of SNNs on GPUs for researchers lacking low-level coding skills.
• To leverage the parallel processing capabilities of GPUs for accelerating SNN simulations.

Main Methods:

• Developed a GPU backend for the Brian simulator, named Brian2CUDA.
• Implemented efficient code generation for numerical integration of neuronal states and synaptic event propagation on GPUs.
• Utilized the massively parallel arithmetic capabilities of GPUs.

Main Results:

• Brian2CUDA accelerates SNN simulations by up to three orders of magnitude compared to Brian's CPU backend.
• It supports Brian's full feature set on GPUs, including arbitrary neuron/synapse models, plasticity, and heterogeneous delays.
• Performance comparisons with Brian2GeNN show comparable speedups, with Brian2CUDA being faster for larger networks.

Conclusions:

• Brian2CUDA combines the Brian simulator's flexibility with GPU simulation speed.
• It significantly reduces the effort required for efficient SNN simulation.
• This software democratizes GPU computing for a broader neuroscience community.