link to homepage

Institute of Neuroscience and Medicine

Navigation and service

Talk by Bruno Golosio

Associate Professor of Applied Physics, University of Cagliari, Italy

15 Feb 2017 15:00
15 Feb 2017
INM-6, Bldg. 15.22, Seminar Room 3009, 1. OG

A GPU library for simulation of large-scale networks of biological neurons

The human brain is an extremely complex system, being made up of about 100 billion neurons that exchange signals through hundreds of trillions of synaptic connections. This massive level of parallelism makes the simulation of significant proportions of the brain a difficult challenge even on the most powerful supercomputers. General purpose computing on graphical processing units (GPGPU) can significantly reduce processing time compared to multi-core CPU systems, because a single GPU can perform thousands of core computations in parallel. In this talk I will present a library for simulating large-scale networks of biological neurons on GPU architectures. The simulations rely on the adaptive-exponential-integrate-and-fire (AdEx) neuron model, which in many cases can provide a realistic description of neuronal signal and is much faster than biologically detailed models. I will illustrate the results of simulations of balanced networks of excitatory and inhibitory neurons, with up to 1 million neurons in a single GPU device,

interconnected by up to 100 thousands connections per neuron. I will show that simulations on a single GPU architectures can be more than 20 times faster than those on a multi-core CPU. I will also present the results of simulations made on MPI-GPU clusters and, finally, I will discuss about the potential interest of hybrid simulations, with the GPU used for fast simulation of a large number of AdEx neurons and the CPU used for simulating a smaller number of instances of more

biologically detailed neuron models.


Prof. Dr. Markus Diesmann