JAMES GORBOLD / HARDWARE ACCELERATED
James Gorbold has been building, tweaking and overclocking PCs ever since the 1980s. He now helps Scan Computers to develop new systems.
OPINION
W
henitcomestoprocessors,beitCPUs,GPUsorco-
processors,youcanneverhavetoomuchperformance.
Afterall,nobodywouldwillinglychoosetoplaya
gameat15fpsorwaitlongerthannecessaryfora renderor
encodetocomplete.ButnotallpartsofthePC’sarchitecture
are createdequal,soimprovementstosomeareascanhavea
greaterorlesserimpactonperformancethanothers.
TakeRAM,forexample.Formanyyears,successivegenerations
of premiumIntelandAMDprocessorshavehadhighbandwidth,
quad-channelmemorycontrollersasopposedtothelower-
bandwidthdual-channelmemorycontrollersofmainstream
CPUs.However,multiplestudieshavedemonstratedthat,in
gamesatleast,quad-channelmemoryis
not appreciablyfasterthandual-channel
- theonlyquantifiablebenefitisseen
in complexcontentcreationworkflows,
suchasvideoeditingandhigh-endCAD.
ThePCI-Ebus,whichis usedtoconnect
CPUstocomponentssuchasGPUs,storage
controllers, SSDs, I/O controllers and NICs, has an equally
chequered impact on performance. The latest incarnation,
PCI-E 4, is an interesting case in point, as it makes the world of
difference to SSDs and NICs. NVMe SSDs now have up to 7.9GB/
sec of bandwidth available, unlocking much higher read and
write speeds, as demonstrated by Corsair’s MP600 (see Issue
193, p26). However, while these sorts of transfer speeds are
amazing on paper, in my experience, there’s no quantifiable
speed difference compared with a PCI-E 3 SSD in games.
That said, PCI-E 4 is a godsend for workstations and servers,
particularly for deep learning and other AI workloads. Deep
learning isn’t only incredibly compute-intensive; for systems
with multiple GPUs, feeding those GPUs with data quickly
enoughisa seriouschallenge.WiththerapidrateofGPU
development,theindustryisnowscrabblingtobuildfaster
andfasterstoragearrays.
WithSATASSDsandharddrivesoutofthequestion, the
onlysensibleoptionis a PCI-ENANDorOptane-basedSSD. The
issuehereisthattheSSDscan’tusetoomanyPCI-Elanes, as
thatstarvesothercomponentsofbandwidth.Simplyadding
morePCI-Elanesisn’tviableeither,asit hasa veryhighcost in
termsofprocessorrealestate,andmakesmotherboardlayout
extremelycomplexandcostly.Forthesereasons,it’sfarmore
efficienttoprovide7.9GB/secofbandwidthfromfourPCI-E 4
lanesthanfromeightPCI-E3 lanes.PCI-E4 is alsoa blessing for
networking,enablingNVMeoverFabric
configurationsrunningatover200Gb/
secfroma singleNIC.
Allthatsaid,I can’tseePCI-E4 having
nearlysucha largeimpactonGPUs.
Steppingdownfrom 16 lanestoeight
PCI-E3 lanesalreadyhasbarelyanyeffect
on gaming or rendering performance. And in deep learning,
which is very bandwidth-intensive, Nvidia already has a
workaround for the limitations of PCI-E, NVLink. In its most
basic form, it runs at 50GB/sec, much faster than the 31.5GB/sec
that 16 PCI-E 4 lanes will provide. Plus, in some configurations,
NVLink already scales to 200GB/sec.
While this talk of deep learning might seem a bit off-topic in
Custom PC, it’s becoming extremely influential in component
design, with manufacturers queuing up for a slice of the
market. Just look at how much of a song and dance Nvidia
made about the Tensor cores in the Turing architecture, which
was designed first and foremost for deep learning, and later
repurposed for gaming.Stepping down from 16 to eight
PCI-E 3 lanes already has
barely any effect on gamingPRACTICAL PERFORMANCE
James Gorbold loves PCI-E 4 and quad-channel memory, but claims
they’re unnecessary in terms of gaming performance