Professional Documents
Culture Documents
Neuromorphic platforms
Neuromorphic platforms objective is to enable simulation of Spiking Neural
Networks (SNN) in real-time
The guiding philosophy is not to simulate the brain in complete detail, but to search
for organizing principles that can be applied in practical devices
Real-time (or faster) simulation performance is a desirable feature
Cognitive robotics need to interact with the environment through
asynchronous sensors
analog
ASIC
1 billion neurons
Digital
pla=orm
20-wafer
system:
four
million
neurons
and
one
billion
synapses
Programming models
What are suitable programming models for neuromorphic
chip/platforms?
How to describe algorithms
How to map networks on the platform
Objective:
Maximize efficiency neuromorphic computing systems
Scale-up neural max network size
Improve simulation reliability
Communication optimization
Allocate neurons to minimize chip-to-chip communication packets generated by
spikes
Approach: Put in the same core densely interconnected neurons through spectral
analysis of the neuron network graph (*)
A Biological Benchmark
NEURONS
L5E
L5I
L6E
L6I
TIME
Andrea Acquaviva (andrea.acquaviva@polito.it)
Mapping Impact
Baseline
Op+mized
Cor+cal
Microcircuit
Op+mized
Baseline
EXTERNAL
DUMPED
EXTERNAL
5%
242 K
No
6 M
10%
10%
992 K
No
RUNTIME ERROR
10%
9%
923 K
No
25 M
15%
15%
2 M
No
RUNTIME ERROR
5%
DUMPED
Increasing
network
size
(*)
Barchi,
Urgese,
Macii,
Acquaviva,
OpKmizing
Network
Trac
for
Spiking
Neural
Network
SimulaKons
on
Densely
Interconnected
Many-Core
Neuromorphic
Pla=orms,
IEEE
TransacKons
on
Emerging
Topics
in
CompuKng,
2016
Computation optimization
Optimize synaptic input computation by splitting synaptic computation
from neuron computation: Synapse-centric approach(*)
Synapses vs neuron simulation allocation of different processors
Post-synapKc
neurons
computaKon
is
divided
among
cores
SynapKc
input
currents
are
computed
by
synapKc
processors;
Neuron
processors
sum
the
input
currents
for
each
neuron
Ecient
in
case
of
sparse
networks
such
as
TCM
(*)
Knight,
Furber,
Synapse-Centric
Mapping
of
CorKcal
Models
to
the
SpiNNaker
Neuromorphic
Architecture,
FronKers
in
Neuroscience,
2016
Mapping impact
While the synapse-centric approach requires more cores in general, it
allows the network to be simulated in real time in all configurations (*)
Increasing
network
size
(*) Knight, Furber, Synapse-Centric Mapping of CorKcal Models to the SpiNNaker Neuromorphic Architecture, FronKers in Neuroscience, 2016
10
11
THANK YOU
12
How-to
13
Methodology
14
How-to
d
15
Talamo-Cortical microcircuit
16
How-to
Clustering
17
18
19
20
Some results
Impact
on
simulaKon
size
Op+mized
Baseline
EXTERNAL
24.5
Million
of
packets
723
Dumped
packets
EXTERNAL
16.5
Million
of
packets
0
Dumped
packets
21