Julie Grollier

CNRS/Thales lab,

Palaiseau, France

1

Electronic Nanodevices

for Bio-inspired Computing

bioSPINspired

Impressive progress in artificial intelligence but in terms of power efficiency, the brain is the

winner

2

Brain: 20W

AlphaGo: 150 kW

Computing with low energy requires entangling memory with processing

Power consumption in current computers is high because of Von Neumann’s bottleneck: separation between memory and processing

3

proces- sing

memory

Digital computer

synapse neuron

neuron memory

processing

Brain

100 W/cm2 20 W in total !

Recent progresses with CMOS technology : example of IBM’s TrueNorth chip

• Highly parallel, colocalized memory and processing

• Low power consumption 20 mW/cm2 (processor 100 W/cm2)

• Cannot learn

4 Merolla et al, Science 345, 668 (2014)

5.4 billion transistors

256 million synapses

5000 neurons

(1 million neurons

w. time-multiplexing)

Relying on current technology (CMOS) alone is not a long-term solution

• A transistor is nanoscale but it is just a switch

• CMOS does not provide memory (volatile)

5

10-100 µm

SRAM banks

plasticity

Schemmel et al., IJCNN (2006)

CMOS neuron

CMOS synapse 10 µm

Merolla et al, Science 345, 668 (2014)

To build smart chips, novel nanoscale devices are needed to emulate neurons and synapses

6

Hundreds of millions of neuron-like and synapse-like devices in a 1 cm2 chip Each device << 1 µm2

• Brain : 1011 neurons, 1015 synapses

• AlphaGo: millions of neurons and synapses

• Visual cortex: 500 millions of neurons

Ingredients needed for neural networks: non-linearity, memory and plasticity

7

Cat

Dog

Cat

Dog

• Synapses: analog valves (weights w)

• Neurons: non-linear

+1

-1

x1

x2

x3

y

w1

w2

w3

y = f ( w1 x1 + w2 x2 + w3 x3)

• Nano-Synapses

• Nano-Neurons

• Variability and noise

8

Many options ! I will focus on electronic devices

9

• Nano-Synapses

• Nano-Neurons

• Variability and noise

Memristors tunable nano-resistors with memory implement the fundamental ingredients of synapses

10

Red-Ox Phase change

Chua, IEEE Trans.

Circuit Theory (1971)

Yang et al., Nat. Nano. (2013) Kuzum et al, Nanotechnology (2013)

Memristors that do not involve large ionic/atomic displacements are interesting for

improved endurance and speed

11

André Chanthbouala, JG et al, Nat. Mat. 11, 860 (2012)

Spintronic

Steven Lequeux, JG et al, Sci. Rep. 6:31510 (2016)

S. Fukami et al, Nat. Mater. 15, 535 (2016)

Ferroelectric

P. Wadley et al, Science 351, 587 (2016)

Vth

Memristors emulate electronic synapses: the weight is their tunable conductance G

12

Ui < Vth : calculating mode Ui > Vth : learning mode

G1

G2

Gi

Kirchhoff’s law

Current = S Gi Ui

memristor

One challenge being currently tackled is building large arrays of memristors

• 3D Xpoint, Intel/Micron Optane Lenovo 32 Gbits

13

http://www.theregister.co.uk/2017/01/04/optane_arrives_at_ces/

memristor

selector

Recent progress has been achieved towards the commer-cialization of binary memories made of memristor arrays

Neuromorphic computing with memristors is a nascent field but it opens the path to real-time

on-chip learning

14

IBM experiments: Handwritten digit recognition with 165 000 synapses (phase change with selector)

Burr et al, IEEE IEDM (2014)

Supervised learning, back-propagation

TiO2

Prezioso et al, Nature 521, 61 (2015)

Biological neurons are more than non-linear functions

• Spikes

• Integration (memory)

• Non-linear oscillator

15

0 5 10 15 20

-70

-35

0

35

Time (ms)

Voltage (

mV

)

Spikes allow learning without teacher(s)

16

pre-synaptic neuron

post-synaptic neuron -600 -400 -200 0 200 400 600

-2

-1

0

1

2

G

(µ

S)

t (ns)

Spike timing plasticity

Jo et al, Nanoletters 10, 1297 (2010)

Zamarrenos-Ramos et al, Frontiers in Neuroscience 5, 26 (2011)

Sören Boyn, JG et al, Nature Com. 8, 14736 (2017)

Memristor conductance can evolve autonomously

17

Spikes allow learning without teacher(s)

Unsupervised learning

Presented input

Presented input

18

Spikes allow learning without teacher(s)

Unsupervised learning

Presented input

19

Spikes allow learning without teacher(s)

Unsupervised learning

Presented input strengthened synapse

weakened synapse

20

Spikes allow learning without teacher(s)

Unsupervised learning

Recent small scale demonstration: Pedretti, Ielmini et al, Sci Rep 2017

Large scale simulations: Querlioz, Bichler et al, Neural networks 2012

Perspectives

• Memristors can enable online learning for deep networks

• Issues: noise and non-linearity

• Improve: energy consumption, cyclability

• The future: unsupervised learning

21

22

• Nano-Synapses

• Nano-Neurons

• Variability and noise

Biological neurons are more than non-linear functions: can we reduce the number of neurons needed for computing by giving them dynamical properties ?

• Spikes

• Integration (memory)

• Non-linear oscillator

23

0 5 10 15 20

-70

-35

0

35

Time (ms)

Voltage (

mV

)

Non-linear dynamics in the brain has inspired many computing models

24

Synchronization

periodic

eeg s

ign

al

t (s)

brain waves

Complex transients

synapse: neuron: oscillator coupling neuron: oscillator

Jaeger et al, Science (2004), Hoppensteadt et al, PRL (1999), Aonishi et al, PRL (1999)

25

Many attemps to compute through sync with nanoscale oscillators: NEMS, memristive osc.

R

t

Unipolar Memristor

R

C V

NDR

R

V 0

VO2 NbO2 TaOx …

Pickett et al, Nat. Mater. (2013), N. Shukla et al, Sci. Rep. (2014)

A. Sharma, et al, IEEE J. Explor. Solid-State Comput. Devices Circuits (2015)

Learning through synchronization requires controlling and tuning the coupled dynamics of

oscillators

26

this is

an ae

inputs

ae

synchronized

ah

synchronized

this is

an ah

E. Vassilieva et al, IEEE Trans.

Neural Networks, 22, 84 (2011)

Magnetic oscillators are nanodevices with well controlled dynamics

27

CoFeB

FeB

MgO

spin torque

Nicolas Locatelli, V. Cros and J. Grollier, Spin-torque building blocks, Nat. Mat. 13, 11 (2014)

magnetic tunnel junction

compatible with CMOS

10-100 nm

Nanoscale, fast (GHz) and easily measurable

Same structure as magnetic memories

0 5 10 15 20-20

-10

0

10

20

Time (ns)

Voltage (

mV

)

current

315 3200.0

0.4

0.8

1.2

Pow

er

(W

)

Frequency (MHz)

Thanks to its stability and non-linearity, a single magnetic oscillator can emulate an assembly of neurons and perform neuromorphic computing

28

J. Torrejon, M. Riou, F. Abreu Araujo, JG et al, Nature 547, 428 (2017)

Spoken digit recognition through reservoir computing

Spin-torque nano-oscillators have an unusually high tunability compared to other oscillators

29

AC signals

W. H. Rippard et al., PRL. 95, 067203 (2005)

R. Lebrun, JG et al, PRL 115, 017201 (2015)

Enhanced sync ranges

0.0 0.5 1.0 1.5 2.0

3.4

3.6

3.8

4.0

frequency (

GH

z)

dc current (mA)0 25 50 75 100 125

2.5

3.0

3.5

4.0

4.5

frequency (

GH

z)

magnetic field (Oe)

180 200 220 240

200

205

210

215

Oscill

ato

r fr

eq

uency (

MH

z)

Source frequency (MHz)

A. Slavin and

V. Tiberkevich,

IEEE TM 45,

1875 (2009)

30

AIST Tsukuba Japan: Sumito Tsunegi, Kay Yakushiji, Akio Fukushima, Hitoshi

Kubota, Shinji Yuasa

IEF Orsay France: Damir Vodenicarevic,

Nicolas Locatelli, Damien Querlioz

CNRS/Thales: Philippe Talatchian, Miguel Romera, Flavio Abreu Araujo, Paolo Bortolotti, Vincent Cros, Julie Grollier

Collaboration: implementing and training a spin-torque neural network

31

• Nano-Synapses

• Nano-Neurons

• Variability and noise

Novel nanodevices: interesting features but….

32

→ High device variability → High sensitivity to noise

→ Stochastic behavior

Difficult to achieve boolean computing but what about neuromorphic computing ?

(Memristors,

Hewlett-Packard)

(Carbon nanotube) (Nanomagnetic logic,

Cowburn et al.) (Spin transfer

oscillator,

CNRS/Thales)

Thanks to their plasticity, neural networks are highly resilient against nanodevice variability !

33

simulations

Damien Querlioz et al, IEEE Trans. Nano., vol. 12, num. 3, p. 288 (2013)

Biological synapses and neurons are noisy: the brain seems to operate at the thermal limit

to minimize its power consumption

34

Computing at low power with stochastic components is possible

Neural spikes in response to the same input recorded 50 times

Can we realize reliable computations with stochastic devices ?

Stochastic computing: sampling multiple times, averaging

35

0 2 4 6 8

Re

sis

tan

ce

Time (s)

RAP

RP

Time

Voltage emitted by neuron

Can we even leverage noise for computing ?

36

Low noise Optimal noise High noise

Randomness can be useful - Example: stochastic resonance

Gammaitoni et al, Reviews Of Modern Physics, 70(1), 223–287 (1998)

Locatelli, Mizrahi, JG et al, Phys. Rev. Appl. 2014

Conclusion

• Memristors can emulate synapses

• Magnetic nano-oscillators can emulate neurons

• Novel materials, nanodevices and physics open new perspectives for neuromorphic computing

37

• Will be hiring Ph.D students and post-docs next year

Spintronics vs neuromorphic computing

1- Why neuromorphic computing ? What are we trying to achieve ?

2- What are the major challenges in neuromorphic computing ?

3- What is the problem with all-CMOS approaches ?

4- What are the competitors for spintronics ?

5- What are the disadvantages of spintronics ?

6- What are the advantages of spintronics ?

38

Spintronics: a toolbox for neuromorphic computing

39

J. Grollier et al, PIEEE 104, 2024 (2016)

S. Lequeux, JG et al, Sci. Rep. 6:31510

(2016)

A. Mizrahi, JG et al, arXiv 1610.09394

(2016)

J. Torrejon, JG et al, Nature 547, 428

(2017)

Hynix/Toshiba

IEDM 2016