Bloomberg Businessweek - USA (2019-06-10)

BloombergBusinessweek|SoonerThanYouThink June 10, 2019

it couldusetheenergysavings tocrunchmore

numbers;it’ssortoftheequivalentofa humanbrain

shiftingfromcalculating the exact GPS coordinates of

a restaurant to just remembering its name and neigh-

borhood. “If we built a processor more attuned to this

sortofwork,wecouldincreasetheperformancebya

factorofa thousand,”Knowlessays.Youngandoth-

ersweresoimpressedthatKnowlesandToondecided

theyhadtostartGraphcore.Theybeganraisingcap-

italtodeveloptheideaasearlyas 2013 andrevealed

thecompanytotheworldin2016.

Thesemiconductorindustryiscurrentlydebating

thesustainabilityofMoore’slaw,anobservationdat-

ingbacktothe1960sthatsaysthenumberoftransis-

torsona chip—andthus,itspriceperformance—will

doubleabouteverytwoyears.Graphcore’sleaders

areinsteadconcernedwitha relatedconcept,called

Dennardscaling,whichstatedthatastransistorden-

sityimproved,powerdemandswouldstayconstant.

Buttheprinciplenolongerapplies,andaddingmore

transistorstochipsnowmeansthechipstendtoget

hotterandmoreenergy-hungry.Tomitigatethisissue,

somechipmakersdesigntheirproductssotheydon’t

usealltheirprocessingpoweratonce—unusedareas

of thechiparecalled“darksilicon”—andinsteadrun

onlythepartsnecessarytosupportanapplication.

KnowlesandToonsaytheheatproblemespeciallywill

stopphonesandlaptopsfromgettingmuchfasterin the

yearsaheadunlesscircuitscanberadicallyredesigned

forefficiency.“Iwasgivena blanksheetofpaperto

start,whichneverhappensin chipdesign,”saysDaniel

Wilkinson,whoworksonGraphcore’schiparchitecture.

Theychallengedtheteamofa fewdozenengineers,

mostlycastoffsfromtheirpaststartups,todesigna chip

thatcouldharnessallitsprocessinghorsepoweratonce

whileusinglessenergythana state-of-the-artGPU.One

of thebiggerenergystressesin siliconinvolvesmoving

andretrievingdata,buthistoricallyprocessorsarekept

separatefrommemory.Transportingthatdatabackand

forthbetweenthesecomponentsis “veryenergyexpen-

sive,”Knowlessays.Graphcoresetouttodesignwhat

he callsa more“homogeneousstructure”that“inter-

mingles”a chip’slogicwithmemory,soit doesn’thave

to expendasmuchpowertotransportdatatoanother

pieceofhardware.

Overthreeyears,theysimulatedcomputertestson

hundredsofchiplayouts,eventuallysettlingona design

with1,216processorcores,whichKnowlesreferstoas

“lotsoftinyislandsofprocessorsthatsplitupenergy

resources.”TheresultingIPU,firstmanufacturedin 2018,

is a sleekchipthesizeofa WheatThinthathasalmost

24 billiontransistorsandis abletoaccessdatafora frac-

tionofthepowerofa GPU.“Eachofthesechipsruns

at 120watts”—about the same as a bright incandescent

lightbulb—“soabout0.8ofa voltandabout 150 amps,”

Toonsays,standingina messyelectronicslabatthe

Bristolheadquarters,slidinghisthumboveranIPU’s

mirrorlikefinish.

Totestoutitsprototype,theteamfedit astandard

data-trainingmodelofmillionsoflabeledimagesof

commonobjects(fruits,animals,cars).Anengineer

thenqueriedtheIPUwitha photoofhisowncat,Zeus,

andwithinanhourthecomputernotonlyidentifiedit

correctlybutcorrectlydescribedZeus’coat.“TheIPU

wasabletorecognizeit asa tabby,”Knowlessays.

Sincethatfirsttest,theIPUhasspedupandcannow

recognizemorethan10,000imagespersecond.The

goalisforthechiptobeabletodigestandascertain

farmorecomplexdatamodels,tothepointthesys-

temwouldunderstandwhata catis onsomemorefun-

damentallevel.“Wedon’ttellthemachinewhattodo;

wejustdescribehowit shouldlearnandgiveit lots

ofexamplesanddata—itdoesn’tactuallyhavetobe

supervised,”hesays.“Themachinesarefindingout

whattodoforthemselves.”

O

n thefifthfloorofGraphcore’soffices,hulking

industrialairconditionersblastcoolairintothe

company’sdataserverroomandflapwindowshades

toandfro,lettinginsomeofBristol’sunusualmid-

Maysunlight.Asenergy-efficient asthechipsare,

plantedinserversstackedtogetherinfridge-sizecas-

ings,themachinesstillgeneratea hellishamountof

heat.TheseIPUserverracksarepotentenoughtoper-

form 64 petaflopsofcomputing,theprocessingequiv-

alentof183,000iPhoneXsworkingsimultaneouslyat

maxspeed.KnowlesandToonnicknamedtheirIPU

“Colossus,”aftertheworld’sfirstelectronicprogram-

mablecomputer,whichtheBritishgovernmentdevel-

opedto crackencrypted messagesfrom Germany

duringWorldWarII.

Graphcorehasraised$328millionfrominvestors,

includingBMW,Microsoft,andSamsung,andwaslast

valuedinDecemberat$1.7billion.It declinedtocom-

mentonspecificapplicationsforitschips,citingnondis-

closureagreements,butgivenitsinvestors,someseem

obvious—self-drivingcars,Siri-likevoiceassistants,and

cloudserverfarms.ButKnowlesis mostexcitedabout

humanity-altering applications, such as the impact IPUs

could have on the complex analysis that scientists need

for research in climate change and medicine.

“It doesn’t actually have to

be supervised. The machines

are findingoutwhat to do

forthemselves”

62