GPUS: IT’SLOVELACE, RDNA 3
& ALCHEMY, OH MY!
2022 TRIPLE WHAMMY
IT’S NO SECRET that the world of PC graphics has had problems
delivering sufficient quantities of cards ever since the current
generation Ampere RTX 30 -series and RDNA 2 RX 60 00-series
parts launched in late 20 20. Nevertheless, both AMD and Nvidia
should be forging ahead with their next-generation RDNA 3 and
Lovelace architectures in 2 022, and they will be joined by Intel’s
Arc Alchemist. Will you be able to buy a card? Let’s hope the
miners find somewhere else to dig.
Officially, neither company has said much of note about their
next-gen GPU architectures. We expect AMD to move from TSMC
N7, which it has used on the past two generations of RDNA and
RDNA 2 GPUs, to TSMC N5, keeping the latest generation CPUs
and GPUs on the same node.
As with CPUs, N5 should enable a lot more transistors in the
same area, along with better power and performance. RDNA 2
saw some impressive gains from the large Infinity Cache, and
there’s no reason to expect RDNA 3 won’t keep that big cache.
We could see a move to a stacked die holding the cache, but that’s
potentially more expensive and with the larger transistor budget
likely to be unnecessary.
Navi 21 currently has a modest (for a halo GPU) die size of
520mm square, with nearly 27 billion transistors. RDNA 3 and
Navi 31 (or whatever it eventually gets named) could keep the
same die size and increase to almost 50 billion transistors. A 38 4-
bit memory interface and 120 CUs seem possible. Architectural
improvements to increase ray tracing and general shader
performance should yield a GPU that’s up to 50 percent faster
than the RX 6900 XT. It’s also possible AMD could go with a multi-
chip approach and with that a truly massive bump in performance
(see boxout, above right).
Nvidia meanwhile has something to prove with Lovelace.
It’s not that Ampere isn’t good—sometimes, it’s too good when
it comes to mining performance—but efficiency took a big step
back compared to previous designs. Also, the reference model
Founders Edition cards can run incredibly hot thanks to the
GDDR6X memory. Both deficiencies should be addressed with
the RTX 40-series Lovelace cards.
Lovelace is also expected to move to TSMC’s N 5 process,
joining AMD and Apple on the cutting edge. Nvidia perhaps got
a bit complacent with Ampere, figuring that Samsung’s 8 N could
stay competitive with TSMC’s N7. Supply and prices may have
been lower on Samsung Foundry as well, but going from a 260 W
RTX 2080 Ti to a 35 0W RTX 3080 Ti basically scaled power use
with performance. Jumping straight to N 5 should bring drastic
reductions in power.
It’s not just power, of course. 5 nm should provide for substantial
increases in transistor density and performance. GA1 02 was a
large 628mm square die—20 percent larger than AMD’s Navi 21,
despite only having a billion more transistors. Some of that is due
to Infinity Cache, but we expect Nvidia to go after a 6 00mm square
die on N5, which would mean over 50 billion transistors.
MONOLITHIC DIES VS. CHIPLETS
Most graphics cards have traditionally used monolithic
dies—everything built into a single chip. It makes signal
and data routing faster, but it does result in large chips.
AMD may choose to go a different route with RDNA 3, and
Nvidia and Intel may do the same with future GPUs.
With planar scaling of semiconductors slowing down,
opening up the third dimension with stacked dies and
multiple chiplets could be the solution. Making smaller
chips and linking them together also reduces the cost and
the time it takes to get to market.
There are rumors that AMD will use its proven chiplet
approach from CPUs with RDNA 3, with smaller chiplets
interlinked by the Infinity Fabric. Done properly, this could
allow AMD to scale from budget to extreme GPUs without
the need to make three or more different chip designs.
We also know AMD is using a chiplet approach with its
MI250X data center solution, and Intel’s data center Xe
HPC uses multiple ‘tiles’ as well, plus 3 D chip stacking.
We may not see those used in RDNA 3 or Alchemist, but
don’t be surprised if the next-generation products from
all three companies opt for multi-chiplet designs.
AMD’s Zen architecture should march forward in
2022, with the next generation Zen 4 potentially
delivering up to 25 percent more performance
JAN 2022 MAXIMUMPC 19
©^
AM
D
