quickstart
10 MAXIMU MPC MAR 2022
© AMD, INTELAMD HAS RELEASED the first of its budget-
orientated graphics cards sporting
RDNA2 silicon, the RX 6500 XT, which was
quickly followed by the RX 6400. Both use
the new Navi 24, which was designed with
laptops in mind. This helps explain some
of the missing features, including media
coding and decoding, and only two display
controllers. The RX 6500 XT sports 4GB of
GDDR6 RAM, and costs $199.
The GPU packs 16 Compute Units,
1,024 Stream processors, 16MB of AMD’s
Infinity cache, and a ‘game’ clock of
2,610MHz. That’s half of what you get on an
RX 6600 XT, and the quoted fill rates and
FLOPS numbers neatly match that. There
are only four PCIe lanes, as the rest aren’t
wired-up. That wouldn’t be too bad if you
had a system with PCIe 4.0, but for anyone
running PCIe 3.0, it will hurt as anything
that hits the main memory will be pushed
for bandwidth. The RX 6400 is the same,
but with GPU power cut by another 25
percent. It’s expected to sell for $149.
When testers got hold of the card, they
weren’t impressed, and things got spiky.
At 1080p gaming, the RX 5600 XT is barely
a match for AMD’s RX 580, which had a
similar price at launch five years ago. At
1440p, it gets worse. Last spring, AMD
claimed 4GB was “evidently not enough for
today’s games”. It was right then and it’s
still right now. Given more PCIe lanes, you
might get away with it, but not here. The
RX 6500 XT appears to be an improvised
design, coupled with dubious cost-cutting.
The ray-tracing cores are next to
useless, vital features are missing,
and by every metric it’s mediocre. The
recommended price may be regarded as
budget by the market, but it isn’t cheap.
AMD has delivered a cheaper RDNA2 card,
but not the one we wanted. Of course,
given the current market, where any card
is a good card, the initial shipments sold
out almost immediately. –CLCut-down card cut down
RADEON’S
RDNA2 GOES
BUDGET
PCIe to double
data rate again
The goal of PCI Special Interest
Group is to double the data rate
of PCIe every three years and it’s
been true to this since version 1.
in 2003. Version PCIe 6.0 has been
finalized and data rates top out at
8GB/s per lane, meaning a heady
128GB/s on a fully-wired x16 slot.
It will take a while to filter down
to the desktop, where we’ve just
got used to PCIe 4.0. It has been
achieved using PAM4 encoding
(Pulse Aptitude Modulation), with
four signal levels rather than two,
so you can carry twice the data
on the same signal frequency.
PAM comes from the world of
high-speed networks with fewer
physical wires and little headroom
to increase the frequency. –CLIntel Alchemist
enters the fray
If there was ever a time for a new
player to start shipping decent
graphics cards, it’s now. The wait
for Intel to join the party has been
long, but we’re finally nearly there.
The company’s Arc Alchemist
was due to land this quarter, and
now ‘leaked’ benchmarks have
appeared with good news. The top
card uses the DG2-512 die, with
512 execution units (comparable to
4,096 Stream processors), and 32
ray-tracing units.
Run against a $600 GeForce
3070 Ti, the Alchemist holds its
own. We don’t know the clock rates
yet, but there’s always room for
tweaking. The hardware is good,
what about production capacity?
We’re going to need a lot of them. –CLIntel’s new mega-fab
INTEL HAS ANNOUNCED that it’s opening a huge new chip fabrication plant in Ohio.
Construction on the 1,000-acre site is due to start this year, and the first phase
will see an investment of $20 billion into two fabs. It will then be expanded into a
‘mega’ site with up to eights fabs and a total investment of around $100 billion. Once
complete, it will be one of the world’s biggest semiconductor manufacturing plants,
and though it’s not yet clear which process nodes each fab will build, we expect
Intel’s 20A and 18A process (yes, we have reached angstrom measurements). Intel’s
CEO, Pat Gelsinger, said it will “help build a more resilient supply chain”. Fabs aren’t
simple to build though and we won’t see any silicon from the site for three years.
The new plant is part of Intel’s Integrated Device Management 2.0 initiative, which
aims to mix its own foundry capacity with selected outsourcing to regain leadership
in the process technology. Intel struggled with 10nm chips and lost out to Samsung
and TSMC as things shrunk further. Intel’s efforts suffered from high failure rates
as it tried to get multi-layered patterning techniques to work, while its rivals pushed
ahead with EUV (extreme ultraviolet) lithography and the switch in techniques
delayed everything. Intel believes it is back on top form, although it still needs to go
to TSMC for its 3nm chips, and will continue to outsource some other chips. Intel has
not been shy about its ambition to be the world’s leading designer and manufacturer
of silicon chips by 2025. That’s a tall order, but the new plant should help. –CLWill Intel’s plant help
the company realize
its ambitious plans?