Fast Simulations of Highly-Connected Spiking Cortical Models Using GPUs.

Area of Science:

• Computational Neuroscience
• Neuroscience Engineering
• High-Performance Computing

Background:

• Growing interest in simulating large-scale spiking neural networks.
• GPU-accelerated systems offer cost-effective and versatile solutions.
• CUDA-C/C++ programming enables efficient parallelization.

Purpose of the Study:

• Introduce NeuronGPU, a GPU library for large-scale spiking neural network simulations.
• Evaluate NeuronGPU's performance on benchmark neural network models.
• Demonstrate state-of-the-art simulation speeds for complex neural network architectures.

Main Methods:

• Developed NeuronGPU using C++ and CUDA-C++ with a novel spike-delivery algorithm.
• Implemented parallel fifth-order Runge-Kutta method with adaptive step-size control for AdEx models.
• Tested performance on a cortical microcircuit model (LIF neurons) and balanced networks (AdEx/Izhikevich neurons).

Main Results:

• NeuronGPU achieves state-of-the-art performance in simulation time per second of biological activity.
• A cortical microcircuit model (77,000 neurons, 3e8 connections) simulated near real-time on a single GPU.
• A balanced network of 1,000,000 AdEx neurons simulated at ~70s per second of biological activity.

Conclusions:

• NeuronGPU library enables highly efficient, large-scale spiking neural network simulations on GPUs.
• The library's performance is suitable for simulating complex brain models.
• Accelerates research in computational neuroscience through advanced simulation capabilities.