25
Neuromorphic Computing in the European Human Brain Project Karlheinz Meier @brainscales Ruprecht-Karls-Universität Heidelberg NICE 2016, Berkeley

Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Embed Size (px)

Citation preview

Page 1: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Neuromorphic Computing in the

European Human Brain Project

Karlheinz Meier @brainscales

Ruprecht-Karls-Universität Heidelberg

NICE 2016, Berkeley

Page 2: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

1.  Neuroinforma-csPla3ormAggregateneurosciencedata,deliverbrainatlases

2.  MedicalInforma-csPla3ormAggregateclinicalrecords,classifybraindiseases

3.  BrainSimula-onPla3ormDevelopsoDwaretools,runclosedloopbrainsimula-ons

4.  HighPerformanceCompu-ngPla3orm

DevelopandoperateHPCsystemsop-mizedforbrainsimula-ons

5.  NeuromorphicCompu-ngPla3ormDevelopandoperatenovelbrainderivedcompu-nghardware

6.  Neurorobo-csPla3ormDevelopvirtualrobo-csystemsforclosedloopcogni-veexperiments

Publiclauncheventatendoframp-upphase–March,30th

The6ICTPla+ormsinHBP

Page 3: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Temporal Scales and Strong Scaling

Computa7onalComplexity

Memory

Requirement

1MB

10GB

1TB

100TB

100PB

CellularNeocor-calColumn

CellularMesocircuit

CellularRodentBrain

CellularHumanBrain

1Gigaflops 1Teraflops 1Petaflops 1Exaflops

SingleCellularModel

Subcellulardetailandplas-cityrequireadvancesinstrongscaling!

Glia-Cell/VasculatureO(1-10x)

Reac-on-DiffusionO(100-1,000x)

MolecularDynamicsO(>1,000,000,000x)

Plas-cityO(1-10x)

LearningO(10-100x)

DevelopmentO(100-1000x)

Page 4: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

TheONLYwaytoevermakeuseof

ar-ficialneuralcircuitsderivedfrom

biologyistomakethemadap$ve

Connec-vity–Synapses–Neurons

byclosed-loopinterac-onwithdata

Page 5: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Click to edit Master title style

• Click to edit Master text styles

–  Secondlevel•  Third level

–  Fourth level

Previous work is essential

BrainScaleS

SpiNNaker

FACETS/BrainScaleS 2005-2015 SpiNNaker 2005-2015

8-10 years from chip design to system !

Requires roadmap and sustained funding

- Not developed

in HBP -

Page 6: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project
Page 7: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

• 18ARM968Coresperchip• IntegerArithme-c

• 200MHzProcessorClock

• SharedsystemRAMondie

•  128MbyteDRAMstackedondie

•  EachChip6bi-direc-onallinks•  6millionspikes/s/link

•  RealTimeSimulator

SpiNNaker

Group(+HBP)

Page 8: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

HBP SpiNNaker Machine Generations (Manchester Site)

103 104 105

Page 9: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Ra-onalesfortheBrainScaleSPhysicalModelSystem

Ø  Mixed-Signal(Localanalogcomputa-on,binaryspikecommunica-on)

Ø  Drivenbyarchitecture,notdevices(180nmCMOS)

Ø  HighNeuronInputCount(>10.000)

Ø  Configurability(cellparameters,connec-ons)->Universality

Ø  Scalability:ChipScale(105)->WaferScale(108)->Systems(>109)

Ø  Accelera-onx10.000,consistent-meconstants(1daycompressedto10seconds)

Ø  Short-termundlong-termPlas-city

Ø  Upgradabilitywithunchangedsystemarchitecture

Ø  HybridOpera-on,closedloopexperiments

Ø  Non-ExpertUserAccess

Objec-ve:Exploitconfigurabilityandaccelera-on

-rapidexplora-onoflargeparameterspaces

-covershortandlong-mescalecircuitdynamics

-performcompu-nginthepresenceofspa-alandtemporalnoise

Page 10: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

HiCANNHigh

InputCount

Analogue

NeuralNetwork

Chip

Millner,S.,Grübl,A.,Meier,K.,Schemmel,J.andSchwartz,M.-O.,AVLSIImplementa-onoftheAdap-veExponen-alIntegrate-and-FireNeuronModel

AdvancesinNeuralInforma-onProcessingSystems(NIPS)(2010)

Page 11: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project
Page 12: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Physical Model, local analogue computing,

binary continuous time communication

Wafer-Scale Integration of 200.000 neurons and 50.000.000 synapses on

a single 20 cm wafer

Short term and long term plasticity, 10.000 faster

than real-time

Wafer-scaleintegra$onofanalogneuralnetworks,J.Schemmel,J,FieresandK.Meier

In:ProceedingsofIJCNN(2008),IEEEPress,431

Page 13: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Click to edit Master title style

• Click to edit Master text styles

–  Secondlevel•  Third level

–  Fourth level

HBP : Neuromorphic Computing Platform

THEPHYSICALMODELSYSTEM

Localanaloguecompu-ngwith4Millionneurons

and50Millionsynapses–binary,asynchronous

communica-on–runningatx10000real--me

Loca-on:Heidelberg(Germany)

Offering : Access to a unique set of 2 complementary, highly configurable neuromorphic machines for modelling neural microcircuits and applying brain-like principles in machine learning and cognitive computing

THEMANY-COREDIGITALPROCESSORSYSTEM

0.5–1MillionARMprocessors–address-based,smallpacket,

asynchronouscommunica-on–runningatreal--me

Loca-on:Manchester(UK)

1

5

67

23

4

8

8

Page 14: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

500.000core

machine

Loca-on:

Manchester(UK)SeetalkbySteveFurber

Page 15: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

20Wafermodule

machine

Loca-on:

Heidelberg(GE)

SeetalkbyJohannesSchemmel

Page 16: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Click to edit Master title style

• Click to edit Master text styles

–  Secondlevel•  Third level

–  Fourth level

Remote Access ready for users

Slide16

Heidelberg

PyNN

Job queue server

Model description

Experiment description

Data storage

Mapping

Calibration data HAL

Authentication Validation Notification

Manchester

PyNN

PACMAN

SpinnMan

See demo by Eric Mueller

Page 17: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

HBPNeuromorphicCompu-ngGuidebook–Con-nuousUpdates

Comprehensiveopenaccessdocumenta-on:

Hardware,systems,firmware,low/highlevelsoDware

Benchmarks(neuroscience,machinelearning)

Tutorials,smallsystemsdescrip-on

hnp://electronicvisions.github.io/hbp-sp9-guidebook/

Page 18: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

NeuromorphicLaptopAdd-on

PlugsintoUSB,fullsoDware

supportanddocumenta-on

498neurons

100.000plas-csynapses

100.000fasterthanreal--me

[email protected]

SeedemobyEricMueller

Page 19: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Increasingnumberofusecasesandapplica-ons

coveringawidespectrumofnetworktypes

Exploi-ngSubstrateUNIVERSALITY–selec-onofpublishedwork:

-  Canonicalcircuits(synfirechains,WTA,aOractorcircuits)-  Balancedrandomnetworks-  Liquidcompu$ng,temporalpaOerniden$fica$on-  MinicolumnLayer2/3circuits-  Closed-loophybridcontrolsystems-  Mul$variatedataclassifica$on-  Phasedetec$on,applyingSTDP-  Decorrela$onthroughinhibitoryfeedback-  Stochas$cinferencethroughneuralsampling-  BayesiannetworksasBoltzmannmachinesofLIFneurons-  Impleme$ngdeeplearningwithspikingneurons-  Implemen$ngHTMwithspikingneurons

SeetalkbyMihaiPetroviciandposterbyLuziweiLeng

Page 20: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

2023 Roadmap details in FPA document

Page 21: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Click to edit Master title style

•  Click to edit Master text styles

–  Secondlevel•  Third level

–  Fourth level

From SpiNNaker to SpiNNaker2

Feature SpiNNaker SpiNNaker2

technology 130nm 28nm

cores 18 68

core frequency 200MHz >400MHz

external memory 128MByte (1 Gbyte/s) 2GByte (>10 Gbyte/s)

power 1W 1W

power management no yes

floating point support no yes

vector processing no yes

true random numbers no yes

biological realtime operation yes yes

no. of neurons / chip 16k 128k

no. of synapses / chip 16M 128M

energy/synaptic event 10-8J 10-9J

≈10ximprovementatconstantpower

Page 22: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

PhysicalModel:TargetsinHBPfor2023

PrototypesintheLab

StructuredNeurons

Ac-vedendriteswithspa-al

structure:Neuronsascomplex

panerndetectors(e.g.

hierarchicaltemporalmemory

Plas-cityProcessor

400PowerPCprocessorsper

wafer:Re-wiringonthefly,data

drivenaccelerateddevelopment,

slowandfastcircuitsdynamics

SeedemobyEricMueller

Page 23: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Observables Controls Synapseevalua-on

Popula-onrates

Arbitraryinternal

parameters

Weights

Connec-vity

Rewiring

Neuronparameters

Homeostasis

S-mulusgenerators

Externalrewards

andcontrols

Essen-al:Any-mescale>100µs(bio)isaccessible

Page 24: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

Wafer-PCBLamina-onforlargescale

highdensitymanufacturing

Page 25: Neuromorphic Computing - Helen Wills Neuroscience …neuroscience.berkeley.edu/wp-content/uploads/2016/05/Karlheinz... · Neuromorphic Computing in the European Human Brain Project

neuromorphic.eu