neuromorphic computing with magnetic nano-oscillators julie.pdfneuromorphic computing with magnetic...

Neuromorphic computing with magnetic nano-oscillators

Julie Grollier1

Philippe Talatchian1, Miguel Romera1, Sumito Tsunegi2, Flavio Abreu Araujo1, Mathieu Riou 1, Jacob Torrejon 1, Vincent Cros1, Alice Mizrahi 1, Paolo

Bortolotti1, Juan Trastoy1, Kay Yakushiji2, Akio Fukushima2, Hitoshi Kubota2, Shinji Yuasa2, Maxence Ernoult1,3, Damir Vodenicarevic3, Tifenn Hirtzlin3,

Nicolas Locatelli3, Damien Querlioz3

1CNRS/Thales, France 2AIST, Japan 3C2N, France

bioSPINspired

Neural networks run on unoptimized hardware

2

0011011

GPUs, TPUs, FPGAsneurons neurons

synapses

Entangling memory and processing allows for fastand energy efficient computing

3

proces-sing

memory

Digital computer

synapse

neuron neuronmemory

processing

Brain

20 W in total !

processing

100 W/cm2

CPUs, GPUs, TPUs, FPGAs

Can we build small neuromorphic chips that rundeep neural networks ?

4

Hundred millions of neurons and synapses in a 1 cm2 chip Each device smaller than 1 µm2

Nanoneurons

NanoneuronsNano-synapses

5Moritz Helmstaedter lab, retina flight 2013

Future AI will be massively interconnected

Brain: 104 synapses/neurons

Main trend : CMOS neurons + Memristive synapses

6

HP labsmemristor

CMOS

V1V2Vi…

G1

G2

Gi

I = S Gi Vi

10 000 synapses per neuron ?

Strukov and Williams, PNAS 106, 20155 (2009)

Wireless deep learning through radio-frequency (RF) communications

7

Magnetic nano-oscillators are non-linear nano-radios

8

CoFeB

FeB

MgO

spin torque

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

magnetic tunnel junction

compatible with CMOS

10-100 nm

Nanoscale, fast (GHz), non-linear and easily measurable

Same structure as magnetic memories

current0 10 20

-10

-5

0

5

10

Time (ns)

Voltage (

mV

)

)

0 2 4 6 8 10

0

5

10

15

Voltag

e a

mplit

ude (

mV

)

dc current (mA)

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

9J. Torrejon, M. Riou, F. Abreu Araujo et al, Nature 547, 428 (2017)

Spoken digit recognition through reservoir computing

Magnetic nano-oscillators have a high tunability : theyare radio-receivers

10

AC signals

Enhanced sync ranges

0.0 0.5 1.0 1.5 2.0

3.4

3.6

3.8

4.0

fre

que

ncy (

GH

z)

dc current (mA)0 25 50 75 100 125

2.5

3.0

3.5

4.0

4.5

fre

que

ncy (

GH

z)

magnetic field (Oe)

A. Slavin and V. Tiberkevich, IEEE TM 45, 1875 (2009)

The oscillators ability to mutually interact opens the path to RF on-chip communication between neuronlayers

11

…

Same frequency: strong synapse

Different frequencies: weak synapse

i j

Vowels classification with spin-torque oscillator neural network

Inputs: vowels

MagneticNano-Oscillators

Outputs: synchronized states

fA fB

M. Romera, P. Talatchian et al,Nature 563, 230 (2018)

13320 340 360 380

10-4

10-2

100

Spectr

al pow

er

density (

µW

/MH

z)

Frequency (MHz)

I1 I2 I3 I4

Response of the neural network without inputs

320 340 360 380

10-4

10-2

100

Spectr

al pow

er

density (

µW

/MH

z)

Frequency (MHz)

fA fB

I1 I2 I3 I4

The inputs modify the oscillator responses:oscillator 4 is sync to source B

fA fB

15

fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380

We summarize all thesemeasurements in a map

where the differentsynchronization states have different colors

4B

1A

3A

4B

16

Microwave sources~ 300 MHz

MagneticNano-Oscillators~ 300 MHz

Outputs: synchronized states

fA fB

Hillebrands Michingandatabase

Input vowels~ 1 kHz 12 formant frequencies + 1 bias

fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380

For classification, all the points

corresponding to one vowel should fall in a single synchronization

region

We train the network by tuning the currents throughthe oscillators according to an online learning rule

18

I1 I2 I3 I4

Oscillators

1-100

-80

-60

-40

-20

350 370

am

plit

ude (

dB

m)

frequency (MHz)330

2 3 4

fA fB

ex: « iy » 4B

Experimental set-up

Computer(Learning algorithm)

oscillator frequencies (MHz)

dc currents in oscillators(mA)

fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380

recognition rate (%)After 1 step:

0

20

40

60

80

100

4

6

8

10

0 20 40 60 80

340

350

360

370

Number of training steps



fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380

recognition rate (%)After 3 steps:

0

20

40

60

80

100

4

6

8

10

0 20 40 60 80

340

350

360

370




fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380


0

20

40

60

80

100

4

6

8

10

0 20 40 60 80

340

350

360

370


4

6

8

10



0 20 40 60 80

340

350

360

370


fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380


0

20

40

60

80

100



0 20 40 60 80

340

350

360

370


fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380


0

20

40

60

80

100

4

6

8

10

4

6

8

10



0 20 40 60 80

340

350

360

370


fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380


0

20

40

60

80

100



0 20 40 60 80

340

350

360

370


fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380


0

20

40

60

80

100

4

6

8

10

4

6

8

10



0 20 40 60 80

340

350

360

370


fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380


0

20

40

60

80

100

4

6

8

10



0 20 40 60 80

340

350

360

370


fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380

After 40 steps:

0

20

40

60

80

100

recognition rate (%)

4

6

8

10



0 20 40 60 80

340

350

360

370


fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380

After 44 steps:

0

20

40

60

80

100

recognition rate (%)

4

6

8

10

0 20 40 60 80

340

350

360

370




fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380


0

20

40

60

80

100

0

20

40

60

80

100

4

6

8

10

0 20 40 60 80

340

350

360

370




fB(MHz)

380

370

360

350

340

330

fA (MHz)330 340 350 360 370 380

After 86 steps: recognition rate (%)

85 %

34

What’s next ?

Deep neural networks that communicate throughmicrowaves

35

…

RF Nano-neurons

RF Nano-neuronsRF Nano-

Synapses

neuromorphic computing with magnetic nano-oscillators julie.pdfneuromorphic computing with magnetic...

Documents