% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@INPROCEEDINGS{Kurth:889327,
author = {Kurth, Anno and Finnerty, Justin and Terhorst, Dennis and
Pronold, Jari and Senk, Johanna and Diesmann, Markus},
title = {{S}ub {R}ealtime {S}imulation of a {F}ull {D}ensity
{C}ortical {M}icrocircuit {M}odel on a {S}ingle {C}ompute
{N}ode},
reportid = {FZJ-2021-00218},
year = {2020},
abstract = {The local cortical microcircuit is a building block of the
mammalian brain. Its prominent characteristics are similar
across species and cortical areas. This dual homogeneity
raises hopes that principles of cortical computation can be
discovered. The tissue below a 1 mm2 patch of cortex
comprises about 100,000 neurons. It is the smallest network
in which both a realistic number of about 10,000 synapses
per neuron and a connection probability of 0.1 are realized
simultaneously. Potjans and Diesmann [1] compiled a
prototype network model of the microcircuit. In this model
the spatial structure is neglected and replaced by cell-type
specific random connectivity. Each layer is represented by
an excitatory and inhibitory population of
integrate-and-fire model neurons. The circuit has become a
benchmark network for neuromorphic computing systems: its
natural size renders questions of downscaling irrelevant
[2], it can routinely be simulated by present systems [3,4],
and it marks an upper bound from which neuronal networks in
nature are necessarily less densely connected.To apply
neuronal networks with natural densities in neurorobotics,
their simulation needs to become faster. The same holds true
for the investigation of the long time scales of
system-level learning. Achieving this is a promise of
neuromorphic computing. The first intermediate objective is
real-time performance, accomplished for the microcircuit
only recently [4]. However, these results have to be
evaluated in the light of continuously advancing mainstream
architectures as they provide more flexibility at
potentially lower costs.In this contribution we show
performance data for the microcircuit model on two recent
AMD EPYC Rome 128 core compute nodes coupled by a
point-to-point Infiniband interconnect. The software is the
NEST 2.14 [5] (including cherry picked bug fix
726f9b04bbd47c) simulation code providing double precision
numerics and weight resolution. On a single compute node we
measure sub realtime performance. With two nodes the
simulation is 1.7 times faster than realtime. For the single
node the energy per synaptic event is 0.26 μJ, and for the
fastest configuration using two nodes 0.39 μJ. These values
are in the same order of magnitude as the lowest reported
for a simulation of the microcircuit model so far [6]. Our
study exposes present bottlenecks and can guide the design
of future software and hardware systems.},
month = {Sep},
date = {2020-09-29},
organization = {Bernstein Conference 2020, Online
(Germany), 29 Sep 2020 - 1 Oct 2020},
subtyp = {Other},
cin = {INM-6 / IAS-6 / INM-10},
cid = {I:(DE-Juel1)INM-6-20090406 / I:(DE-Juel1)IAS-6-20130828 /
I:(DE-Juel1)INM-10-20170113},
pnm = {574 - Theory, modelling and simulation (POF3-574) /
Advanced Computing Architectures $(aca_20190115)$ / HBP SGA3
- Human Brain Project Specific Grant Agreement 3 (945539)},
pid = {G:(DE-HGF)POF3-574 / $G:(DE-Juel1)aca_20190115$ /
G:(EU-Grant)945539},
typ = {PUB:(DE-HGF)24},
url = {https://juser.fz-juelich.de/record/889327},
}