josephson junction based neuromorphic computingtalk3... · 2017. 12. 20. · josephson junction...
TRANSCRIPT
-
Josephson Junction Based Neuromorphic Computing
Mike Schneider, Stephen Russek, Christine Donnelly, Burm Baek, Matt Pufall, Ian Haygood, Pete Hopkins, Paul Dresselhaus, Sam Benz, Bill Rippard
-
12/14/2017 2
Artificial Intelligence
“Intelligence” exhibited by machines
Machine Learning
Field of study that gives computers the ability to learn
without being explicitly programmed
Bio – Inspired
Drawing inspiration from the human brain
Neuromorphic Hardware
Deep Neural Networks
eyeriss.mit.edu
-
12/14/2017 3
Software neural networks are mainstream
• Apple: A11 bionic SoC, Siri
• Google: Search results (RankBrain), translate, street view, …
• Microsoft: Computational Network Toolkit (speech recognition)
• Amazon: from product recommendation to robotic picking routines
• Facebook Ranking, Language translation, …
M. Nielsen, “Neural Networks and Deep Learning” (2015)
-
12/14/2017 4
Neural networks dominate modern image recognition
2010 2011 2012 2013 2014 2015 2016 20170
5
10
15
20
25
30
Cla
ssific
atio
n E
rro
r (%
)
Year
Deep Convolutional Neural Network (Trained with GPUs)
Human Accuracy ~ 5% error rate
Feature based algorithms
ImageNet Competition
-
12/14/2017 5
What interaction is being modeled in neural networks?
Wikimedia neuronimage
x0
axon
w0
synapse
w0 x0
x1w1
Cell body
w1 x1
𝑓
𝑖
𝑤𝑖𝑥𝑖 + 𝑏output
…
-
Neuromorphic Single Flux Quantum
6
Information transfer: quantized pulse trains
Long distance “lossless” pulse transmission
3D architecture
Memory/ plasticity
Neural Single Flux Quantum
axons Josephson/passive transmission lines
Magnetic JJsynapse
-
• Barrier = insulator, semiconductor, normal metal, ferromagnet, …
7
Josephson junction (JJ)
sincs II
1
101
ie 2202
ie
-
𝜑0
2𝜋𝐶 ሷ𝜃 +
𝜑0
2𝜋𝑅𝑛
ሶ𝜃 + 𝐼𝑐𝑠𝑖𝑛 𝜃 − 𝐼𝑏 = 0
Ib
Circuit Model of a Josephson Junction
8
𝑈 =𝐼𝑐𝜙02𝜋
(−cos 𝜃 −𝐼𝑏𝐼𝑐𝜃)DampingMass
Vse
h 150 1003.2
2 where
-
Interaction of order parameters
9
S F
Phase Progression
Exchange Field
Ph
ase
FM Thickness|C
riti
cal C
urr
ent
Den
sity
|
Jc > 0 Jc < 0(-JJ)
Jc > 0
S-F-S
– eiQ∙R – e-iQ∙R
= ( – ) cos(Q∙R)+ i( + ) sin(Q∙R)
Singlet(S = 0)
Singlet + Triplet(S = 0) (S = 1)
Superconducting Spintronics Reviews
Buzdin, Rev Mod Phys 2005Bergeret, Rev Mod Phys 2005Eschrig, Phys Today 2011Blamire, J of Phys Cond Matt 2014Linder and Robinson, Nature Phys 2015
Fulde-Ferrell-Larkin-Ovchinnikov(FFLO) state
S
-
Magnetic nanoclusters in a Josephson junction
10
Thin Films
Si
Mn
0 100 200 300
0
30
60
90 10 mT
5 mT
0.5 mT
0 mT
Mom
ent (n
Am
2)
Temperature (K)
ordered
disordered
Nb
Nb
-
Changing magnetic order with fast electrical pulses in devices
12/14/2017 11
H
H
Small applied magnetic field
250 psElectrical pulse
Only the small applied magnetic field
250 psElectrical pulse
+
H=0
-
The basic neuro-inspired cell
12
-
Magnetic Josephson junction neural network simulation
12/14/2017 13
Pre-synaptic Neurons
Synaptic Weights
Post-synaptic Neurons
• 9 input pixel (JJs)• 27 weights (MJJs)• 3 outputs (MJJ SQUIDs)
-
Real time (~ 3ns) recognition
14
100111001
Prezioso et al. Nature 2015
z
v
n
-
Real time (~ 3ns) recognition
15
100111001
0 20 40 60 80 100
0
20
40
60
80
100
120
Time (ns)
Offse
t V
olta
ge
(
V)
Prezioso et al. Nature 2015
-
Real time (~ 3ns) recognition
16
000111001
0 20 40 60 80 100
0
20
40
60
80
100
120
Time (ns)
Offse
t V
olta
ge
(
V)
Prezioso et al. Nature 2015
-
Large scale image recognition
• ARGUS
• 1.8 billion pixels at 12 fps
• resolution of 6 inches over 10 square miles
12/14/2017 17
• Currently pre-processed on board using kilowatts
• Full stream processed on a supercomputer
-
Scale to a large network
12/14/2017 18
• Recent ImageNet winners use ~ 1010 MA-operations / image
• ~ 1017 MA-operations /sec to process ARGUS real time
• JJ neural network would take ~ 3 watt (including cooling)
( 1 spike / MA-operation)
-
Benchmarks
12/14/2017 19
• Energy• Operational energy < 10-18
• Training energy < 10 -17
• Cooling overhead ~ 103
• (Human brain ~ 10-15)
• Speed• Operational (device) ~ 100 GHz• Operational (circuit) ~ 10 GHz
• Size • Demonstrated 1.5 x 3 µm• Demonstrated for MJJ 100 nm
• Scalability• Demonstrated ~ 101
• There is a lot of scaling to go…
-
Conclusions
12/14/2017 20
• Josephson junctions are a very promising neuromorphic computing technology
• Magnetic Josephson junctions are a cryogenic non-volatile memory
• Magnetic Josephson junctions can be used for the synaptic function
• There is a lot of scaling needed for Josephson junction based neuromorphic computing