Larger GPU-accelerated brain simulations with procedural connectivity.

Area of Science:

• Computational Neuroscience
• Neuroscience

Background:

• Large-scale spiking neural network models are crucial for understanding brain function.
• Previous simulation methods required substantial memory for synaptic connections, necessitating high-performance computing systems.

Purpose of the Study:

• To introduce a novel, memory-efficient simulation method for large spiking neural network models.
• To make complex brain simulations more accessible to researchers.

Main Methods:

• Developed a 'procedural connectivity' approach, generating connectivity and synaptic weights dynamically during simulation.
• Implemented optimizations for graphical processing units (GPUs) and GPU code generation.

Main Results:

• Successfully simulated a large macaque visual cortex model (4.13 × 10^6 neurons, 24.2 × 10^9 synapses) on a single GPU.
• Demonstrated significant memory reduction compared to traditional storage methods.

Conclusions:

• Procedural connectivity offers a viable and efficient alternative for simulating large-scale neural networks.
• This method enhances the accessibility of advanced brain modeling research by leveraging common GPU hardware.