The reason for this bigchangeis rather
given away with the newGeForceRTX
naming scheme: for theveryfirsttimein
any consumer GPU, Turinghassupportfor
hardware accelerated real-timeraytracing.
To cover the history of real-timeraytracing
would take a feature allofitsown,butthere’s
no hyperbole in saying thatitsarrivalis a
momentous event in GPUhistory.
Any key advance in graphics
microarchitecture isn’t somethingthatjust
happens in hardware though.Software
support is also key, andsureenough,
Nvidia has also investedheavilyinthis
area, working with Microsofttoaddray-
tracing support to DirectX 12 viaDirectX
Ray Tracing (DXR). It hasalsopartneredwith
several of the biggest gamestudiosinthe
world to ensure some majorgamesinclude
a form of ray-tracing capability.ThoseDXR
titles only run on Turing-poweredGeForce
RTX products today, sinceevena yearlater
no other vendor has support.
Ray tracing isn’t the onlyfeatherinTuring’s
cap though. Nvidia has alsoaddedhardware
support for some otherdeveloper-friendly
graphics pipeline changesthatarelessvisible
thanraytracing,butnolessimportanttohow
games are rendered on Turing and future
GPUs from all vendors.
There’s a real sense, then, that the Turing
architecture is Nvidia’s bleeding-edge vision
of what should be possible in a GPU. Let’s
start with the products themselves, before
we dive into the details.
THE PRODUCT LINE-UP
The benefit of writing about Turing around
a year after the initial introduction is that we
now have a very clear picture not just of the
Turing GPUs and product line-up built on
them, but how that line-up has been shaped
over time by the competition Nvidia has
faced from AMD.
Turing has been implemented in various
forms in five separate GPUs to date, from
the biggest TU102 to the smallest TU117, and
that quintet underpins 11 separate desktop
GPU products from the headlining Titan RTX
totheGeForceGTX 1650 —that’sright,
therearestillGTX,ratherthanRTX,named
productsinthefamily.
ThelargestTuringvariantisTU102,which
isa 754mm²chipmadeupof18.6billion
transistorsbuiltusinga specialNvidia-only
variantofTSMC’s12nmprocesstechnology
called12FFN.Thatmightsoundhuge,and
it reallyisinsemiconductorterms,butit’s
actuallyslightlysmallerthanthe815mm²
GV100Volta-architectureGPUusedinthe
prior-generationTitanV.TU102implementssixTuring-class
GraphicsProcessingClusters(GPCs)each
with 12 64-coreStreamingMultiprocessors
(SMs)fora totalof4,608Turing-class
CUDAcores,andhasa huge6 MBL2
cacheconnectedtoa 384-bitwideGDDR6
memorybuswithover 650 GB/secof
memorybandwidthinthefastestproduct
it powers.Thatgiantchiprunsitsstream
processorsata boostclockofalmost
1.8GHzoutofthebox,withstaggeringpeak
performancenumbers.
TU102wasoriginallyfoundsolelyinthe
GeForceRTX 2080 Ti,whichlaunchedin
September 2018 andstillhasn’tfallenbelow
£1,000a yearlater.It ‘only’had4,352CUDA
coresenabledoutof4,608,though,and
had11 GB of GDDR6 memory connected to a
352-bit memory bus. However, a while later,
the full TU102 was unleashed in the £,2500
Titan RTX.ThisGPUisconnectedto 24 GBofGDDR6
memoryona full-width384-bitbus.It may
havelaunchedjustintimeforChristmaslast
year,butatthateye-wateringpricewedon’t
thinkSantaputthemundertoomanytrees!
OnestepdownfromtheTU102is the
TU104,whichpowersthelikesoftheGeForce
RTX2080.Measuring545mm²,andagain
builton12FFN,it’sstilla heftyGPU,byany
standardsotherthancomparisontoNvidia’s
verylargestchips.Inside,therearestillsix
GPCs, but each only has eight SMs compared
to the 12 in TU102. However, there’s still a
huge 4MB of L2 cache for those SMs, and up
to 8 GB of GDDR6 connected to its 256-bit
memory bus.
Three desktop products use TU104. The
GeForce RTX 2080 was there at launch in
September 2018 and uses a slightly cut-
down TU104 (46 of 48 total SMs) running
at 1.71GHz. Then, to spoil AMD’s Navi launch
in late July this year, Nvidia discontinued that
first config and introduced the GeForce RTX
2080 Super that uses a fully enabled TU104
with a 1.815GHz boost clock and 10 per cent
faster GDDR6 memory. Finally there’s the
GeForce RTX 2070 Super, which removes a
whole GPC, making a total of five GPCs and
40 SMs, while the core clock runs to 1.77GHz.
TU106 is the smallest ‘full’ Turing chip,
at 445mm², but again, that’s still pretty
massive for a mainstream GPU. Three
desktop products use TU106, starting with
the GeForce RTX 2070 that came out a few
weeks after the RTX 2080 and RTX 2080 Ti,
and featured a full-fat TU106 configuration
of three GPCs with 12 SMs each running at a
boost clock of 1.62GHz.TheTitanRTXis thefastest
consumergraphicscardbased
ontheTuringarchitectureTHE TURING ARCHITECTURE IS NVIDIA’S
BLEEDING-EDGE VISION OF WHAT SHOULD
BE POSSIBLE WITH A GPU