Maximum PC - USA (2022-06)

It’s easy to get carried away

looking at the top-performing

cards and thinking it’s time

for an upgrade. Resist the

temptation, at least until the

games and applications you’re

using warrant the expense.

Even then, make sure that

your ‘upgrade’ actually

represents a true leveling up

of performance.

The oft-quoted rule of

thumb is that a previous

generation GPU represents

one step down in performance,

but that’s not entirely

accurate. For Nvidia, the RTX

2080 lands midway between

the 3060 and 3060 Ti, the GTX

1080 performs about on par

with the RTX 2060, and the

GTX 980 lines up with the

GTX 1060. That’s two steps

down, but if you look at the

–70 cards, it’s maybe one to

one and a half steps down, and

among the –60 series chips it’s

sometimes less than a single

step, albeit muddied by the

Turing GTX 16-series GPUs.

AMD has swapped through

various naming paradigms so

making comparisons is a bit

trickier, but the RX 5700 XT

matches up with the RX 6600

XT, Vega 64 takes on the 5600

XT, and Fury X matches up

with the RX 580.

If you’re sitting on a GTX

1060, which still rates as the

most popular GPU according

to Steam’s hardware survey,

it’s probably worth thinking

about upgrading. Our advice

is to try and get a graphics

card that’s at least double the

performance of your current

GPU if you want to feel a big

improvement. The RTX 3060

gets you that, as does the RX

6600, but if you can splurge on

an RTX 3070 or RX 6700 XT you

get triple the performance.

The other thing to consider

is when new GPU architectures

and hardware will launch.

Upgrading right before the

new stuff comes out is usually

the worst approach while

being an early adopter can

mean paying higher prices.

Then again, anyone who paid

scalper prices of $1,100 for

an RTX 3080 after that card

launched probably felt good

about the decision when prices

shot up to over $2,000 four

months later. Let’s hope that

never happens again.

WHEN TO UPGRADE

If your old PC looks

like this, you might

need more than just a

new graphics card.

RAY TRACING, ROUND THREE

We need significantly faster hardware

if we want to plow through all these ray

tracing effects while maintaining high

framerates. Thankfully, that should be

coming soon with Nvidia’s Ada and AMD’s

RDNA 3 GPUs—and also the Intel Arc

wildcard (see sidebar), but we’re skeptical

it will be competitive. They will probably

be the RTX 40-series and RX 7000-series,

respectively, but whatever the names,

we expect both companies to cram in

even more transistors. That’s thanks to

a switch to TSMC’s 5nm process nodes—

maybe even “4nm” 4N, though the process

names are more marketing these days.

For Nvidia, this will be the third

architecture with ray tracing support.

Ampere added a second ray/triangle

intersection pipeline to its RT core, along

with some extra hardware to help with

more complex ray-tracing calculations

like motion blur. Ada will also be the

fourth generation of Tensor cores, and we

know from the datacenter Hopper H100

(see MPC203 Tech Talk) that Nvidia can

create large GPUs with a ton of compute.

We don’t think the consumer-oriented Ada

chips will be quite as extreme as Hopper,

but Nvidia could potentially double RT

and Tensor core performance yet again.

Alternatively, it might put in more RT

hardware and focus less on further boosts

to deep learning and AI performance.

There was an Nvidia hack this year that

leaked some details about the upcoming

Ada GPUs. The biggest chip, AD102, has

up to 144 SMs—71 percent more than the

Ampere GA102. Even if Nvidia keeps the

same relative performance per GPU core

and SM, stuffing in that many SMs should

provide a huge boost to performance,

regardless of workload. The second tier

AD103 chip has the same 84 SM maximum

as GA102, but it should be smaller and

presumably less expensive to produce,

which could lead to some compelling

high-end offerings. Take this with a pinch

of salt, but we’re looking forward to Ada.

AMD hasn’t been hacked by LAPSU$,

so we don’t know as much about the

upcoming RDNA 3 architecture. The

data center Aldebaran chips powering

the MI250X are massive and have matrix

cores similar to Nvidia’s Tensor hardware,

but so far, AMD has been hesitant to

embrace machine learning hardware

on consumer products. If AMD skips the

matrix hardware on RDNA 3, it could

potentially cram in more CUs (compute

units, similar to Nvidia’s SMs). Or it might

decide to add matrix cores and look at

supporting Intel’s XeSS algorithm—

stranger things have happened.

With settings maxed

out, Cyberpunk 2077

averages just 23fps

at 4K on the 3090 Ti.

JUN 2022 MAXIMU MPC 35

© GETTYIMAGES/ TASCHA RASSADORNYINDEE / EYEEM