antiferromagnet/ferromagnet heterostructures as synapses ......2019/10/07 · pigeon spiking neural...
TRANSCRIPT
SPICE workshop, Mainz 2019
Antiferromagnet/ferromagnet
heterostructures as
synapses and neurons
A. Kurenkov1-3, S. DuttaGupta1-3, C. Zhang1,4,5
S. Fukami1-3,5,6, Y. Horio1 and H. Ohno1-3,5,6
1 Laboratory for Nanoelectronics and Spintronics, RIEC,
2 CSIS, 3 CSRN, 4 FRIS, 5 CIES, 6 WPI-AIMR (Tohoku University)
Spiking neural networks with CMOS
CMOS-based
synapse
CMOS-based
neuron
E. Chicca et al., Proc. IEEE 102, 9 (2014)
1 million neurons
PNAS 105, 3593 (2008)
0.001% of the human brain
1/60th of its speed
3 GHz CPU x60
100 billion neurons at full speed
=
60 × 105 × 60
=
36 million CPUs /
1011 neurons and 1015 synapses
≈
1016 transistors
≈
2 million CPUs /
2.5 million neurons
Science 338, 1202 (2012)
1 million neurons
Science 345, 6197 (2014)
“True North” by IBM
8 million neurons
IEEE Micro, 38, 1 (2018)
“Loihi” by Intel
1 million neurons
PNAS 105, 3593 (2008)
0.001% of the human brain
1/60th of its speed
3 GHz CPU x60
1995 2000 2005 2010 2015 202010
0
101
102
103
104
105
106
107
108
109
1010
1011
1012
CMOS CPU
"Performance" SNN
"Exploratory" SNN
Tra
nsis
tors
/
N
eu
ron
s
Year
Human
Pigeon
Spiking neural networks with CMOS
Why bother with number of neurons?
“Overall brain size … best predicts cognitive ability across non-human primates”
R. O. Deaner et al., Brain Behav. Evol. 70, 115 (2007)
What properties are we seeking?
Leaky integrate-and-fire
Spike-timing-dependent plasticity
New material approaches
Leaky integrate-and-fire
Spike-timing-dependent plasticity
S. Boyn et al., Nat. Commun. 8, 14736 (2017)
Ferroelectric cell
C. Zamarreno-Rames et al., Front. Neurosci. 5, 26 (2011)
Leaky integrate-and-fire
Spike-timing-dependent plasticity
S. Boyn et al., Nat. Commun. 8, 14736 (2017)
Ferroelectric cell
Mott insulatorP. Stoliar et al., Adv. Funct. Mater. 27, 11 (2017)
New material approaches
Leaky integrate-and-fire
Spike-timing-dependent plasticity
S. Boyn et al., Nat. Commun. 8, 14736 (2017)
[ Ferroelectric tunnel junction ]
P. Krzysteczko et al., Adv. Mater. 24, 762 (2012)
[ Magnetic tunnel junction ]
Y.-F. Wang et al., Sci. Rep. 5, 10150 (2015)
[ Resistive ]
D. Kuzum et al., Nano Lett. 12, 2179 (2012)
Y. Li et al., Sci. Rep.3, 1619 (2013)
[ Phase-change ]
F. Alibart et al., Adv. Funct. Mater. 22, 609 (2012)
[ Hybrid organic-nanoparticle ]
M. T. Sharbati et al., Adv. Mater. 30, 1802353 (2018)
[ Graphene ]P. Stoliar et al., Adv. Funct. Mater. 27, 11 (2017)
Pickett et al., Nat. Mater. 12, 114 (2013)
H. Lim et al. Sci. Rep. 5, 9776 (2015)
[ Mott insulators ]
S. Dutta et al., Sci. Rep. 7, 8257 (2017)
Z. Wang et al., Proc. Int. Electron Devices Meeting
(IEDM) p. 13.3.1 (2018)
[ MOSFET / CMOS ]
New material approaches
Memory
controller
The problem of uniformity
CPUMemory
SNN
A bus Buses (2.000.000.000.000.000)
Transistors Memory cells “Buses” Synapses
Equal to 285.000 Intel i9 CPUs assuming
single-transistor converters
States and Rules
Leaky integrate-and-fire Spike-timing-dependent plasticity
MemristiveBinary
States
Rules
[Co/Ni]
PtMn
In-plane magnetic field
annealing
Material system
Current, Ich
Hext = 0
S. Fukami, A. K. et al. Nat. Mater. 15, 535 (2016)
Field-free SOT switching
Current
Hall
resis
tance
Size-dependent SOT switching
5000 nm
Hall
resis
tance
Current
650 nm
Hall
resis
tance
Current
225 nm
Hall
resis
tance
Current
50 nm
In collaboration with M.
Baumgartner, G. Sala, G.
Krishnaswamy, P. Gambardella
[ETH Zurich, Switzerland]
and F. Maccherozzi, S. S. Dhesi
[Diamond Light Source, UK]
Measured at Diamond Light
Sourcce (UK) end-station I06
A. Kurenkov et al., Appl. Phys. Lett. 110, 092410 (2017)
Poster by
Gunasheel
Krishnaswamy
Size-dependent SOT switching
Variation of exchange bias
A. Kurenkov et al., Appl. Phys. Lett. 110, 092410 (2017)
-4 0 4
Co
un
ts o
f d
om
ain
s
Effective field (mT)
Out-of-plane component
-20 0 20 40
Sh
are
of
do
ma
ins
Effective field (mT)
In-plane component
MemristiveBinary
Rules: local time
Leaky integrate-and-fire Spike-timing-dependent plasticity
Pulse 1
excites
a timer Pulse 2
“reads” the timer and
changes system
accordingly
Timer
evolves in time
Local time
Switching by short pulses
-4x1011
-2x1011
2x1011
4x1011
1 s
10 s
100 ns
10 ns
1 ns
RH
all (
)
jch
(A/m2)
0.3
10-7
10-5
10-3
10-1
101
-0.3
-0.2
-0.1
0.0
0.1
0.2
0.3
P (s)
1e-2
1e-3
1e-4
RH
all (
)
Integrated time (s)
V = 6.3 Volts
Effect of temperature dynamics
10-8
10-7
10-6
10-5
10-4
5
6
7
8
9
10
0
0.1
0.2
0.3
0.4
0.5
RHall
()
Puls
e a
mp
litud
e (
V)
Pulse width (s)
10-7
10-5
10-3
10-1
101
-0.3
-0.2
-0.1
0.0
0.1
0.2
0.3
P (s)
1e-2
1e-3
1e-4
1e-5
1e-6
5e-7
3e-7
1e-7R
Hall (
)
Integrated time (s)
V = 6.3 Volts
Effect of temperature dynamics
10-8
10-7
10-6
10-5
10-4
5
6
7
8
9
10
0
0.1
0.2
0.3
0.4
0.5
RHall
()
Puls
e a
mp
litud
e (
V)
Pulse width (s)
0 10 20 300
50
100
150
200
250
300
calculated from
resistance change
fit with 4.6 s by
Troom
T (1 et /
)Te
mp
era
ture
(C
)
Time (s)
-11 0 11
-0.1
0.0
0.1
Pulse
= 10 ns
Stage temp
150°C
130°C
100°C
80°C
50°C
21°C
RH
all (
)
VPulse
(V)
Temperature-based clock
0 1 2 3
-5
0
5
0 1 2 3
25
30
35
Tem
pera
ture
(C
)V
oltage (
V)
Pulse generator output
Calculated temperature dynamics
Time (μs)
3
2
Pulse amplitude
Hall
resis
tance
1
10-8
10-7
10-6
10-5
10-4
6
7
8
9
10
"cold pair"A
mp
litude
of th
e p
uls
es |V
P| (V
)
Width of the pulses P (s)
0.00
0.04
0.08
0.12
R50ns
HallR
1ms
Hall
(Ohm)
P,-V
P
P,V
P
P,-V
P
P,V
P
1 ms
21
21
50 ns
"hot pair"
Too small
Too large
Temperature-based clock
Just right
10-8
10-7
10-6
10-5
10-4
5
6
7
8
9
10
0
0.1
0.2
0.3
0.4
0.5
RHall
()
Puls
e a
mp
litud
e (
V)
Pulse width (s)
Synaptic functionality
-5 -1 0 1 5-0.2
-0.1
0.0
0.1
0.2
post
Delay (s)
R
Ha
ll (
)
175 ns
9.2V
175 ns
9.2V
pre
A. Kurenkov et al. Adv.Mater. 31, 1900636 (2019)
Spike-timing-dependent
plasticity in rat hippocampal
synapses
G. Bi, M. Poo, J. Neurosci. 18, 10464 (1998)
Synaptic learning
E. Izhikevich, Cerebral Cortex 17, 10 (2007)
«Increase of synaptic
strength by consistent
coincidental firing of the
pre- and post-neurons»
0.01 0.1 1 10
0
1
0 ns 800 ns
0 ns 800 ns
0 ns 800 ns
100
101
102
103
104
0
1
0.0 ms 1.5 ms
0.0 ms 1.5 ms
0.0 ms 1.5 ms
npulses
= 104
fin (MHz)
Sw
itch
ing
pro
ba
bili
ty
Vp 777 mV,
p 100 ns V
p 777 mV,
p 100 ns
fin = 8 MHz
Sw
itch
ing
pro
ba
bili
ty
npulses
L. Li et al. eLife 6, 23612 (2017)
Simulation-intensity-dependent firing probability in
the artificial neurons
Neuronal functionality
Stimulation-intensity-
dependent firing probability in
a mice neuron
A. Kurenkov et al. Adv.Mater. 31, 1900636 (2019)
Coincidence detection
AB
Outcome 1 if A and B happen within a time
window
Outcome 2 if A and B are more separated
-200 -100 0 100 200
0
1
(ns)
Psw
25 ns, 999 mV
100 ns, 999 mV
100 ns, 777 mV
175 ns, 777 mV
Sound localization
T. Franken et al., Nat. Neurosci. 18,
444 (2015)
Δτ
-Δτ
CD1
CD2
CD3
Conclusions
• Large-scale and efficient spiking neural networks can’t rely on CMOS
• Uniformity of new-materials synapses and neurons is crucial (because of the
dense and highly interconnected nature of spiking networks)
• AFM/FM devices provide a uniform and
scalable solution with all the benefits of
spintronics (speed, endurance, scalability,
3D stacking)
A portion of this work was supported by JSPS
KAKENHI 17H06093, 18KK0143 and 19K15428,
JST-OPERA, JSPS Core-to-Core Program and
Cooperative Research Projects of RIEC.
More details in
A. Kurenkov et al. Advanced Materials 31, 1900636 (2019)
Material requirements for a hardware SNN basis:
• Tunable switching behavior: binary and memristive
• Presence of dynamics that can be excited and read out