FEATURE / ANALYSIS
frames, creating inefficiency. Meanwhile, for
SFR, there’s ever more overhead in managing
the splitting and restitching of frames. Quad SLI
did introduce an alternative rendering mode that
had each pair of GPUs work on alternate frames,
with each frame then split using SFR between
those to GPUs. However, this is all by the by, as
quad SLI is no longer supported.
What do you need for SLI?
Since its launch, SLI has remained a
remarkably unchanged technology, at least in
terms of its fundamentals. You require two or
more graphics cards that have identical GPUs
and memory configurations, although they
can be a mix of card vendors. You’ll also need
a hefty enough power supply with sufficient
cables to power all the cards. With SLI now
limited to only the highest-end GPUs (from
the RTX 2070 Super upwards), you’ll likely be
looking at needing either two each of 6-pin
and 8-pin graphics power connections, or
four 8-pin connections.
Next, you need an SLI-compatible
motherboard, which are usually based on
the latest high-end chipsets from AMD or
Intel, although not all of these boards support
SLI. Look for an SLI badge on the box, or
check the manual or online specs list of your
board first. You'll also want two appropriately
positioned 16x PCI-E slots on your board for
SLI operation, with at least space for three
slots between them. That way, there’s still
room for a dual-slot card to get air through its
cooling system.
There were also considerations for PCI-E
bandwidth back when three and four-card
SLI was an option, but with a maximum of
two cards, it’s no longer an issue. Generally,
a mainstream desktop board will split its two
16x PCI-E slots into two eight-lane slots for
SLI operation, while some HEDT boards will
let you get the full 16 lanes of bandwidth in
both slots. It rarely makes much of a difference
though – eight PCI-E 3 lanes is still enough for
SLI, even on a top-end RTX card.One reason why PCI-E bandwidth is
less of an issue than you might expect
is because much of the communication
between cards isn’t done over PCI-E but
the SLI bridge. This is a separate physical
connection between cards, which plugs into
the top ridge of the cards. These bridges
have taken various forms over the years,
including generic flexible ribbon cables,
more rigid PCBs and even illuminated
plastic doohickeys, but fundamentally
they’ve remained very similar, allowing SLI
bridges to span multiple generations of
GPU, at least until recently.HB-SLI and NVLink
While the fundamentals of SLI have
remained very similar over the years, there
has been a couple of recent tweaks to the
formula. For a start, with the GTX 10-series
Pascal line-up, Nvidia introduced a new
high-bandwidth SLI bridge, which doubled
the width of the bridge and in turn doubledIT’S ONLY WHEN WE REACH THE UPPER ECHELONS
2)3(5)250$1&(7+$708/7,ˎ*386(7836
START TO MAKE SENSEGPU1
Frame
nFrame
n+1Frame
n+2Frame
n+3Frame
n+4Frame
n+5GPU2
AFR mode simply sends each frame to alternate GPUs SFR mode sees each GPU working on the same frames together
GPU1
Frame
nFrame
n+1Frame
n+2Frame
n+3Frame
n+4Frame
n+5GPU2