Fast Simulations of Highly-Connected Spiking Cortical Models Using GPUs

Over the past decade there has been a growing interest in development of parallel hardware systems for simulating large-scale networks spiking neurons. Compared to other highly-parallel systems, GPU-accelerated solutions have advantage relatively low cost and great versatility, thanks also possibility using CUDA-C/C++ programming languages. NeuronGPU is GPU library simulations neural network models, written C++ CUDA-C++ languages, based on novel spike-delivery algorithm. This includes simple LIF (leaky-integrate-and-fire) neuron models as well several multisynapse AdEx (adaptive-exponential-integrate-and-fire) with current or conductance synapses, user definable different devices. The numerical solution differential equations dynamics performed through implementation, CUDA-C++, fifth-order Runge-Kutta method adaptive step-size control. In this work we evaluate performance simulation cortical microcircuit model, neurons current-based balanced excitatory inhibitory neurons, conductance-based synapses. On these will show that proposed achieves state-of-the-art terms time per second biological activity. particular, single NVIDIA GeForce RTX 2080 Ti board, full-scale cortical-microcircuit which about 77,000 $3 \cdot 10^8$ connections, can be simulated at speed very close real time, while 1,000,000 1,000 connections was 70 s