Transcript Saighi-Basics_of_Neurmorphic_Design

Basics of Neuromorphic Design
Sylvain Saïghi
IMS Lab – University of Bordeaux
Outline
 Integrated Circuit Design
 Models for neuromorphic implementation
 Main types of neuromorphic systems
GDR BioComp – Colloque 2015 – St Paul de Vence
2
How to design an Integrated Circuit?
 IC technology providers for prototyping and low volume
production
 Circuits Multi-Projets
 Europractice
http://cmp.imag.fr/
http://www.europractice-ic.com/
 Same kind of providers in North America
GDR BioComp – Colloque 2015 – St Paul de Vence
3
How to design an Integrated Circuit?
 Schedule
GDR BioComp – Colloque 2015 – St Paul de Vence
4
How to design an Integrated Circuit?
 Schedule and Prices…
GDR BioComp – Colloque 2015 – St Paul de Vence
5
How to design an Integrated Circuit?
 Second step: the schematic
 Design
GDR BioComp – Colloque 2015 – St Paul de Vence
6
How to design an Integrated Circuit?
 Second step: the schematic
 Design
 Simulation
 Monte Carlo to
check the reliability
against the parameters
mismatch
GDR BioComp – Colloque 2015 – St Paul de Vence
7
How to design an Integrated Circuit?
 Third step: the layout
 Mask drawings are sent
to the IC provider
 The provider gathers all
chips from different
designers
GDR BioComp – Colloque 2015 – St Paul de Vence
8
How to design an Integrated Circuit?
 Third step: the layout
 Mask drawings are sent
to the IC provider
 The provider gathers all
chips from different
designers
 You will receive 25 samples
3 or 4 months later
 Total duration about 12-18 months
GDR BioComp – Colloque 2015 – St Paul de Vence
9
Another solution for digital design
 VHDL: Very High-Level Design Language
 Descriptive language
 Behavioral simulation
 Choice of target
• IC (Founder library)
• FPGA (Field
Programmable
Gate Array)
 Physical simulations
GDR BioComp – Colloque 2015 – St Paul de Vence
10
A neuron for neuromorphic designers
Synaptic terminals
Soma
Node of Ranvier
Axon
Myelin sheath
 A few figures




Pre-synaptic
cell
Soma diameter ~ 4 to 100 mm
Axonal time propagation ~ 1 to 30ms
Membrane voltage from ~ -120mV to 50mV
Ionic currents ~ 10 nA
GDR BioComp – Colloque 2015 – St Paul de Vence
Post-synaptic
cell
11
A neuron for neuromorphic designers
 Spiking Neuron → time dependent
SYNAPTIC
CURRENTS
Integrator
Threshold
detector
GDR BioComp – Colloque 2015 – St Paul de Vence
Spike
generator
12
Learning in Neuromorphic Systems
 Hebbian rule: Spike Timing Dependent Plasticity (STDP)
Presynaptic neuron
Nj
Causality
Wji
Postsynaptic neuron
Ni
Increasing of
Synaptic Weight
(potentiation LTP)
No causality
Decreasing of
Synaptic Weight
(depression LTD)
13
Threshold = 2
1/3
1/3
1/3
1/3
1/3
1/3
14
Threshold = 2
2/3
2/3
2/3
2/3
2/3
2/3
15
Threshold = 2
1/3
1/3
1/3
1
1
1
16
Threshold = 2
0
0
0
2/3
1
1
17
Threshold = 2
0
0
0
1/3
1
1
18
Threshold = 2
0
0
0
0
1
1
19
First neuromorphic design
 Carver Mead – Caltech
1989
 The term neuromorphic was coined by Carver Mead, in the late
1980s to describe Very-Large-Scale Integration (VLSI) systems
containing electronic circuits that emulate the bio-physics of
neural systems using the physics of silicon.
GDR BioComp – Colloque 2015 – St Paul de Vence
20
2 main types of neuromorphic design
 Neuromimetic
 Rebuild, understand the life
 Biological Real Time (99% of case)
Analog
Digital
?
?
?
?
 Could be connected to life cells
 Bio-Inspired
 Engineering issues
 Event based computation
GDR BioComp – Colloque 2015 – St Paul de Vence
21
Neuromimetic design
NeuroGrid (Standford)
 Simulate the brain
in real-time
 Specifications:
 sixteen Neurocores
each of which has
256 x 256 silicon neurons in an 11.9 mm x 13.9 mm chip
 Analog core and Digital communication bus
“Neurogrid simulates a million neurons connected
by billions of synapses in real-time, rivaling a
supercomputer while consuming a 100,000 times
less energy—five watts instead of a megawatt!”
GDR BioComp – Colloque 2015 – St Paul de Vence
22
Neuromimetic design
SpiNNaker (Manchester University) – Human Brain Project
 Simulate the brain in real time
 Specifications
 a million-core computing
engine
 massively-parallel
 fully digital design
GDR BioComp – Colloque 2015 – St Paul de Vence
23
Neuromimetic design
Heidelberg’s system (Heidelberg University) – Human Brain Project
 Simulate the brain 104 faster than real-time
 Specifications
 Based around wafer-scale VLSI
 Analog core and Digital
communication bus
 Each 20-cm-diameter silicon wafer
contains 384 chips, each of which
implements 128,000 synapses and up to 512 spiking neurons.
Total of around 200,000 neurons and 49 million synapses per wafer.
GDR BioComp – Colloque 2015 – St Paul de Vence
24
Bio-Inspired design
 TrueNorth – IBM
 Video treatment
 Specifications
 Fully digital
GDR BioComp – Colloque 2015 – St Paul de Vence
25
Bio-Inspired design
 TrueNorth – IBM
TrueNorth video
GDR BioComp – Colloque 2015 – St Paul de Vence
26
Bio-Inspired design
 Zeroth – Qualcomm
 Neural Processing Units
which learn your abs
 Specifications
 Fully digital
GDR BioComp – Colloque 2015 – St Paul de Vence
27
Qualcomm video
GDR BioComp – Colloque 2015 – St Paul de Vence
28
Other neuromorphic research groups
(non-exhaustive)
 Europe
•
•
•
•
Institute of Neuroinformatics (Zurich)
Instituto Microelectronica (Sevilla)
University of Bordeaux
…
 North America
•
•
•
•
Georgia Tech (Atlanta)
Johns Hopkins University (Baltimore)
University of California (San Diego)
…
 Asia
• University of New South Wales (Sydney)
• …
GDR BioComp – Colloque 2015 – St Paul de Vence
29
Thanks for your attention