Project
|
|
|
|
|
|
|
PI
|
|
|
|
|
|
|
Some co-authors
|
Seo, Merolla
|
Lester, Temple, Shapiro
|
Meier, Ehrlich,
Davidson, Bruderle
|
Bishop, Schurmann,
Druckmann, Hill, Hines
|
Lin, Sridharan, Wittig,
Wen, Chandrasekraran
|
Joshi, Vogelberg,
Mallik, Hafliger, Bergh, Yu
|
Site
|
IBM Research
|
Manchester
|
Heidelberg+
|
EPFL Switzerland
|
Stanford
|
UCSD
|
Related projects
|
|
|
|
|
|
NeuroDyn
project with fine-grain bio-realism
|
Overview papers
|
|
|
|
|
|
|
Other papers
|
|
|
|
|
|
|
Neuron simulation
|
Leaky integrate-and-fire
neurons on chip implemented in digital hardware,
Hebbian or STDP learning in circuit.
|
Low-power ARM 968 CPUs
simulate 1000 neurons each. 18 ARM CPUs
with RAM on each chip.
|
Interconnected analog
neuro-morphic ASICs. Claim to emulate same
waveforms as neurons on same input. Several
hundred “Spikey” chips, each hosts about 100K
synapses.
|
Supercomputer simulates
analog processes of neurons and dendrites.
Can handle about 100K neurons with BlueGene.
|
Programmable analog
neurocore chip emulates about 1000 neurons. Also
building custom ASICs with similar architecture.
|
HiAER IFAT chips emulate
65,000 neurons. Uses dynamically programmable
conductance-like synapses.
|
Synapse simulation
|
Array of axons x neurons
w/synapses at junctions. SRAM stores
weights. AER networking between chips, not yet
specified.
|
Spikes delivered by
AER packet network both on-chip and
between chips. Simple routers store synaptic
connections in RAM tables. Programmable delays.
|
Novel wiring of uncut
wafers; intra-wafer chip connections. HICANN
“analog network chip”. Use AER packets
inter- wafer?
|
Message-passing in
software with programmed delays. STDP/etc
implemented in software as well.
|
Direct wiring to nearby
within neurocore, digital AER messaging for more
distant, synaptic addresses stored in RAM.
|
HiAER-IFAT uses local
connections on-chip, tree of FPGA routers for
AER inter-chip.
|
Comments on overall
approach
|
All digital (no analog).
Early state at this point, with 256-neuron chip.
|
Connection info in
shared RAM on chip, used for routing; also, each
CPU has 100KB RAM.
|
Led project with other
EU sites. 100K times faster than real
neurons. PyNN language to define layouts.
|
Constructing model of
10,000 neurons from rat somatosensory cortical
column on “Blue Gene” supercomputer.
|
Encouraging others to
use tools and approach.
|
HiAER-IFAT scales
w/simple neuron model, NeuroDyn is small-scale
for more bio-realism.
|
Main advantage
|
Utilizes latest IBM chip
fabrication. Identical to software
emulation, w/higher performance.
|
Uses and expands on
known silicon, architecture, and networking
technology. Easy to experiment, all software.
|
Moderately bio-realistic
at neuron level. Some innovation in
connectivity. Combines analog and digital
effectively.
|
Most bio-realistic. No
new hardware required. Software is easy to
change.
|
Combines best of analog
and digital.
|
Combines best of analog
and digital.
|
Main disadvantage
|
Axon-neuron crossbar may
not scale well? Inter-chip connectivity
TBD.
|
|
Interconnection problem
unsolved: using wires locally, “event packets”
with target address/time for longer range?
|
Expensive simulation.
|
Limited scalability to
date; degree of neuron plasticity unclear.
|
Can AER router tree
handle bandwidth at root?
|
Other comments
|
Received substantial
DARPA SyNAPSE funding.
|
Is an ARM CPU overkill
to simulate neurons? Could put 1000 simpler CPUs
on a chip?
|
Use “wafer scale
integration”: connect chips directly on wafer.
Each chip can host 8 neurons with 16K inputs, or
512 neurons with 256 inputs.
|
Using model of 10K
neurons from somatosensory cortex of 2-week-old
rat. Model is stochastic based on
statistical % connectivity, not the actual
connections.
|
Experimenting with
applications of their approach, e.g. analog VLSI
Silicon
Retina growth cones and Silicon
Cochlea.
|
The papers with Yu are
about the Neurodyn
chip.
|