NeuroMat supercomputer to simulate neuronal systems

Aug 31, 2015

The Research, Innovation and Dissemination Center for Neuromathematics (RIDC NeuroMat) should launch soon a High-Performance Computational Center at the University of São Paulo, Ribeirão Preto campus, and thus advance in the understanding of interacting neuronal systems. "Simulations provide models on which one may apply and test hypotheses that remained at an abstract level. The key challenge is to incorporate as many key features in the simulations for the results to be meaningful," said Antonio Carlos Roque, a NeuroMat principal investigator and the coordinator of NeuroMat's supercomputer cluster.

The first goal of NeuroMat’s HPCC in Ribeirão Preto is to generate a valid simulation of the cortex. The structure that is being conceived has millions of interacting neurons, organized in an architecture that combines low and high-amplitude connectivity (local and global connectivity). This structure will capture the patterns of neuronal connections in the cortex, both along the vertical direction, with columns of densely interconnected neurons, and the horizontal direction, with a patchy network of interconnected columns organized in a macroscopic network of brain areas.

Elements to compose local and global areas come from accumulated research on neural tracts and circuits. Each column in the model will have around 80,000 interacting neurons, and the expectation is that dozens of columns will be considered in the simulation, this demanding high computational capacity. The simulated structure will take into consideration other elements that are known from accumulated experimental studies on cortical networks, such as proportions of excitatory and inhibitory neurons and probabilities of connections between cell types in specific locations.

A second goal is to simulate stochastic cortical spiking, thus connecting NeuroMat’s theoretical agenda to neural simulation. NeuroMat’s HPCC will constitute one of the first attempts to implement a stochastic model of the cortex at large-scale.

Such stochastic simulations could shed new light on brain functioning both under normal and abnormal conditions, like brain injuries, which could eventually lead to advances in the understanding of operating modes of the cortex and the role of plasticity mechanisms.

NeuroMat’s HPCC is the first of two supercomputers that are expected to be placed at the RIDC. A supercomputer with specific focus on data storage and management is expected to be purchased to give support to NeuroMat’s neuroscientific open access database research.

NeuroMat’s computational center is located at the Neural Systems Laboratory (SisNe - FFCLRP - USP), and the work associated to the HPCC will bring together several NeuroMat members, including Antonio Carlos Roque (coordinator), Jorge Stolfi (scientific advisor), Mauro Coppelli and Osame Kinouchi.

Specifications

The cluster presents high speed processing capacity having four nodes, each containing eight Intel Xeon processors (E5-2650, v3 family) with ten cores. Moreover, for each processor there is a 16GB DDR4 memory and the system also has a graphic accelerator Nvdia Tesla K40 with 2880 CUDA cores, which are essential to enable fast and large numerical simulations.

In order to make efficient use of its processing capacity at multiuser level, the system is provided with specific parallelizing software for C++ and Fortran. Besides processing facility, the supercomputer will also improve NeuroMat’s numerical and experimental data storage capacity with six hard drives of 1TB each. Additionally, it has a 15 KVA APC Nobreak to provide stability and security to the simulations.

Launching event

The launching event of NeuroMat’s HPCC will occur during the VI Latin American School on Computational Neuroscience (LASCON), from 3 to 29 January 2016, when researchers and students will use the functionalities for their practical classes.

LASCON is the first and foremost school of computational neuroscience in Latin America. It is a biennial school aimed at introducing advanced undergraduate, graduate students and young researchers, mostly from Latin America, to the use of mathematical and computational methods for modeling neurons and neural networks of the brain. It is an intensive four-week school with morning, afternoon and evening activities.

The faculty is composed of researchers with large experience in computational neuroscience and the use of these programs. Besides attending the classes and doing the exercises, students have to carry out small research projects using the tools taught at the school and present them orally at the end of the school.

More information on LASCON is available here.

This piece is part of NeuroMat's Newsletter #19. Read more here

Share on Twitter Share on Facebook