the bandwidth. This move was mainly
necessary due to the ever increasing take-
up of 4K and higher resolutions.
Even more recently, with the launch
of RTX, Nvidia brought over its NVLink
technology that had previously featured
only on its Quadro cards. NVLink brings a
massive leap in bandwidth, from the 2GB/
sec of HB SLI to 900GB/sec, plus it works
in a fundamentally different way. While SLI
purely sets up a master/slave relationship
between the first graphics card (which
handles all your video connections) and
any other cards, with NVLink, the data
communication is bi-directional and the
cards operate in a mesh fashion, with no
master/slave relationship.
That’s all great for computational
workloads, but it doesn’t actually translate
that well to GPU workloads where latency
is so crucial. Moreover, Nvidia simply hasn’t
enabled true NVLink compatibility on its
consumer graphics cards. While you get the
connector fingers on the top of RTX cards
and you’ll have to buy an NVLink bridge,
the connection still actually operates in the
same SLI mode as before.
This master/slave relationship is also
part of the reason why SLI (and CrossFire)
doesn’t see a doubling of your VRAM when
you add two cards together, unlike with
true NVLink. That’s because all the data in
VRAM is basically managed by the master
card and then duplicated on the other cards,
rather than being an independent data set.
Which cards will work?
As both consumers and manufacturers alike
have shown less and less enthusiasm for multi-
GPU setups, so has support for the technology
steadily dwindled. As we mentioned earlier,
AMD simply no longer supports CrossFire in its
latest GPUs, while Nvidia has limited support to
only some of its newest models.
Out of its latest raft of RTX 20-series
cards, only the 2070 Super, 2080, 2080
Super and 2080 Ti support SLI. That
means the likes of the RTX 2070 and
2060 aren’t supported, and neither are
any of the company’s Turing-based
GTX cards. On a certain level, this makes
a lot of sense. If you’re seeking better
performance when you only have a GTX
1660, it’s unequivocally a better bet to just
upgrade to a faster card. There are plenty
of much faster cards available that aren’t
ludicrously expensive.
It’s only when we reach the upper
echelons of performance that multi-GPU
setups start to make sense. Here, there’s
either nowhere to go in terms of faster
single cards if you already have an RTX
2080 Ti. Or, if you have an RTX 2080
Super, that’s just one rung down from the
best in terms of performance, and doubling
up this card’s performance would put it
significantly above the fastest single card.
There are occasionally circumstances
where you’re rocking a three or four-
The evolution of the SLI
connector. From left to
right, an Asus HB-SLI
connector, an EVGA
NVLink bridge and the
very first SLI connector