In terms of the balance of resources
in Navi10, the first implementation
of RDNA, AMD has made astute
choices. The ratio of ALU to texture
operations remains solid at 16:1, which
is fine and in line with other competitive
designs from other vendors. The ratioofALU
to back-end pixel export resources, which
blend and then move pixels out of thecore
into the GDDR6 memory at the end ofpixel
shading, is down to 40:1, from 64:1 in Vega.
That means there’s a higher
back-end export rate to service
finished pixels and get them
out of the core, so that more
can be flushed through. That
leads to less back pressure
in the memory system, and
less stalling as the hardware
waits for export grants to be given so that
old pixels can make way for new ones.
Lastly, that focus on the new front-end
geometry performance in RDNA has
resulted in a 320:1 ALU:triangle ratio, down
from 1,024:1 in Vega. That means there’s
comparatively much more front-end
performance to feed the new WGP-based
shader core with useful work to do, keeping
the machine busier and better utilised for,
again, common gaming graphics workloads.
NAVI10 BY THE NUMBERS
Navi10 is basically a well-balanced 40 WGP
part. Each WGP has a new dual 32-wide
SIMD structure, and they share access to
a pair of beefy scalar ALUs, plenty of local
data shares, and a big shared VGPR store.
Each group of ten WGPs also shares a big
512KB L1 cache, well-provisioned back-
end export hardware, a Primitive Unit and
a rasteriser:together,theyallprovide
thebasicbuildingblockofanRDNA
GPU.Indeed,thatbuildingblockcould
reasonablystandaloneasitsownGPU.
Navi10takesfourofthosebuildingblocks
andamalgamatesthemintotwoShader
Engines,allsharinga massive4MBL2cache
andthemaincommandandcontrollogic
neededtodrivethedesignandtalktothe
CPU.Theheadlinestatsthereforerunto
2,560SIMDlanes, 80 scalarALUs, 40 high-performancefour-sample-per-clocktexture
units, and 64 pixels per clock of export rate.REMEMBERING HOW
TO BUILD A GPU
So what’s the end result of this newly
balanced GPU? The tested performance
of the first RDNA-based GPU, Navi10, and
the Radeon RX 5700-series products
that it powers, show a solid performance
increase at the same power as comparable
Vega products, and an even larger increase
in measured performance for the same
die area, even though some of that
benefit comes from the new 7nm process
technology AMD is using to make it.
Mostly importantly, Radeon RX
5700-series products compete with Nvidia’s
equivalents in a way that we haven’t seen
from AMD for quite a while, and at the
right price points.AMD’sdesignchoicesinRDNA,the
configurationputtogetherforNavi10,and
theeconomicsit getsforit fromTSMC’s
7nmprocess,plusthepairingwithGDDR6
insteadofHBM2,meanyougeta strongly
competitiveproductbasedona reworkedand
reimaginedmicroarchitecture,onwhichAMD
cancontinuetobuildandkeepcompeting.
Therearenonewrenderingfeaturesin
RDNAcomparedwithVega,
forwhichsomepeoplehave
criticisedthecompany.There’sno
hardwareraytracing,novariable
rateshading,andnomeshshader
support.However,theseomissions
arelikelydeliberate– enabling
AMDtotreatRDNAandNavi10
similarlytothewayit treatedZen1 andthe
first Summit Ridge-based Ryzen 1000 Series
processors – you get strongly competitive
products out there with a great technical
base, nail the basics in the microarchitecture
again, use advanced manufacturing
technology to get an extra edge, and then
power on towards future products that
can really put some pedal to the metal.
That looks like AMD’s goal with
RDNA: remember how to build great
GPUs to lay the foundations for the
future, and bring more performance
and features to the table next time.
Not having to do everything at once paid
dividends for the Zen team, as we’ve all
seen with Zen2 and Ryzen 3000-series
products. RDNA and Navi10 is hopefully
similar for the graphics team at AMD; more
competition in any semiconductor market
can only be good for consumers.AMD’s Radeon RX 5700-series cards might
not have ray-tracing hardware, but their
rasterisation performance is competitiveTHERE’S COMPARATIVELY MUCH
MORE FRONT-END PERFORMANCETO FEED THE NEW WGP-BASED
SHADER CORE