RICHARD SWINBURNE / VIEW FROM TAIWAN
OPINION
C
hipengineershavebeenyappingonabout‘theend
ofMoore’sLaw’forwhatseemslikeforever,butnow
thattheendofscalingisinsight,astransistorgates
reacha dozenatomsofsilicon(at0.11nmeach),itlookslike
thecost,notphysics,willpreventusfromgettingthetiniest,
fastestchips.ArecentarticleinSemiconductorEngineering
claimsthatthecostofdesigninga single3nmchipwillbe
overa billionUSdollars.Itspecificallycites‘acomplexGPU
atNvidia’ascosting$1.5billionUStodevelop.Yowzers.
Incontrast,a typical12/14/16nm-generationchipcostsjust
~$100millionUStodevelop,whilethelatest7nmdesigns,
suchasAMD’sRyzen3000-seriesCPUsorits
RadeonRX5700-seriesGPUs,triplesthecost
to~$300millionUS.Developmentofnext-
generation5nmchips,whicharedueat
TSMCnextyear,pushupthatcostto$
millionUS.
Bearinmindthatthiscostisforeachchip
design,soifacompanyneedsmultiple
productsforseveralmarkets,itwillneedtomultiplythat
figureaccordingly.Forsomemarkets,suchasCPUs,a single
designcouldbeusedforthewholetop-to-bottomofthe
productstack,withcore-binningandfeature-choppingbeing
usedtocreateseparateproducts.AMD’schipletapproachis
anotherwaytokeepdowncosts–its7nmocta-corechipis
appliedacrosstherangefromentry-leveltoserverproducts,
withvaryingdesignsofthecheaper14nmI/Ochipbeingused
tomeetdifferentmarketneeds.
GPUs,however,requirethreeorfourchipdesignsforeach
generation.Forexample,Nvidiawillhaveseveralchipsbased
onitsnextGPUarchitecture,likelytobenamed02,04, 06 and
08,coveringthewholemarketfromentry-levelGPUs to the
Titan.Thismeansthateachchipneedstosellinhugevolumes
inordertocoveritscost,orit willhavetooverlapintodifferent
markets,suchasserver,supercomputerorindustrialuse.
Thankfully,thePCgamingmarketislarge,butsoon three
companieswillbevyingforyourgraphicsupgrademoney as
Intelentersthefray.It’snota stretchtodrawthedotsand see
howwecouldendupwithjusttwomainGPUmakers again
withina coupleofchipgenerations.WithNvidiareaping
70-80percentoftheGPUmarket,andIntelhavinghuge sums
ofmoneyatitsdisposal,it’snota goodoutlookforAMD.
It doesn’tlookasthoughthevariouscosts
ofchipmakingcouldcomedownintime
either.Oncewehit3nm,we’llneeda change
oftransistordesignawayfromFinFETto
moreexoticstyles,suchasGateAllAround
(GAAFET),nanowires,nanosheetsorother
approachesthatareunderconsideration.
Wemayevenhavea situationwhereall
theseapproachescoexistwithdifferentperformance
characteristicsfordifferentapplicationsneeds,withoptions
forlowpower,highdensity,highperformanceandsoon.This
addedcomplexitywilladdtothecost,ofcourse.Otherissues,
suchasverifyingthattensofbillionsoftransistorsareall
workingcorrectly,arealsoa partofthereasonwhyevery
stageofchipmakingisgettingmoreexpensive.
ThePChardwareandgamingindustriesarelargeenough
forustoexpectbusinessasusualforthenextthreetofive
years,butbeyondthattimeframe,theeconomicsofbuilding
chipsmightcompletelytransformthewayourhardware
ismade.VERY EXPENSIVE CHIPS
As the move to smaller transistors continues, the cost of developing chips
is skyrocketing. Richard Swinburne analyses the future of PC chip makingRichard has worked in tech for over a decade, asa UKjournalist,onAsus’ROGteamandnowasanindustryanalystbasedinTaiwan @ricswi
For CPUs, a single design
could be used for the
whole top-to-bottom
of the product stack