<s>
Scalable	B-Device
Link	I-Device
Interface	I-Device
(	O
SLI	O
)	O
is	O
a	O
brand	O
name	O
for	O
a	O
deprecated	O
multi-GPU	O
technology	O
developed	O
by	O
Nvidia	O
for	O
linking	O
two	O
or	O
more	O
video	B-Device
cards	I-Device
together	O
to	O
produce	O
a	O
single	O
output	O
.	O
</s>
<s>
SLI	O
is	O
a	O
parallel	B-Operating_System
processing	I-Operating_System
algorithm	O
for	O
computer	O
graphics	O
,	O
meant	O
to	O
increase	O
the	O
available	O
processing	O
power	O
.	O
</s>
<s>
The	O
initialism	O
SLI	O
was	O
first	O
used	O
by	O
3dfx	O
for	O
Scan-Line	B-Device
Interleave	I-Device
,	O
which	O
was	O
introduced	O
to	O
the	O
consumer	O
market	O
in	O
1998	O
and	O
used	O
in	O
the	O
Voodoo2	O
line	O
of	O
video	B-Device
cards	I-Device
.	O
</s>
<s>
Nvidia	O
later	O
reintroduced	O
the	O
SLI	O
name	O
in	O
2004	O
and	O
intended	O
for	O
it	O
to	O
be	O
used	O
in	O
modern	O
computer	O
systems	O
based	O
on	O
the	O
PCI	B-Protocol
Express	O
(	O
PCIe	O
)	O
bus	O
;	O
however	O
,	O
the	O
technology	O
behind	O
the	O
name	O
SLI	O
has	O
changed	O
dramatically	O
.	O
</s>
<s>
SLI	O
allows	O
two	O
,	O
three	O
,	O
or	O
four	O
graphics	B-Architecture
processing	I-Architecture
units	I-Architecture
(	O
GPUs	B-Architecture
)	O
to	O
share	O
the	O
workload	O
when	O
rendering	O
real-time	B-General_Concept
3D	O
computer	O
graphics	O
.	O
</s>
<s>
Ideally	O
,	O
identical	O
GPUs	B-Architecture
are	O
installed	O
on	O
the	O
motherboard	O
that	O
contains	O
enough	O
PCI	B-Protocol
Express	O
slots	O
,	O
set	O
up	O
in	O
a	O
master	B-Operating_System
–	I-Operating_System
slave	I-Operating_System
configuration	O
.	O
</s>
<s>
All	O
graphics	B-Device
cards	I-Device
are	O
given	O
an	O
equal	O
workload	O
to	O
render	O
,	O
but	O
the	O
final	O
output	O
of	O
each	O
card	O
is	O
sent	O
to	O
the	O
master	O
card	O
via	O
a	O
connector	O
called	O
the	O
SLI	O
bridge	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
a	O
two	O
graphics	B-Device
card	I-Device
setup	O
,	O
the	O
master	O
works	O
on	O
the	O
top	O
half	O
of	O
the	O
scene	O
,	O
the	O
slave	O
the	O
bottom	O
half	O
.	O
</s>
<s>
The	O
SLI	O
bridge	O
is	O
used	O
to	O
reduce	O
bandwidth	O
constraints	O
and	O
send	O
data	O
between	O
both	O
graphics	B-Device
cards	I-Device
directly	O
.	O
</s>
<s>
It	O
is	O
possible	O
to	O
run	O
SLI	O
without	O
using	O
the	O
bridge	O
connector	O
on	O
a	O
pair	O
of	O
low-end	O
to	O
mid-range	O
graphics	B-Device
cards	I-Device
(	O
e.g.	O
,	O
7100GS	O
or	O
6600GT	O
)	O
with	O
Nvidia	O
's	O
Forceware	O
drivers	O
80.XX	O
or	O
later	O
.	O
</s>
<s>
Since	O
these	O
graphics	B-Device
cards	I-Device
do	O
not	O
use	O
as	O
much	O
bandwidth	O
,	O
data	O
can	O
be	O
relayed	O
through	O
just	O
the	O
chipsets	B-Device
on	O
the	O
motherboard	O
.	O
</s>
<s>
However	O
,	O
if	O
there	O
are	O
two	O
high-end	O
graphics	B-Device
cards	I-Device
installed	O
and	O
the	O
SLI	O
bridge	O
is	O
omitted	O
,	O
the	O
performance	O
will	O
suffer	O
severely	O
,	O
as	O
the	O
chipset	B-Device
does	O
not	O
have	O
enough	O
bandwidth	O
.	O
</s>
<s>
Two-way	O
,	O
three-way	O
,	O
and	O
four-way	O
SLI	O
uses	O
two	O
,	O
three	O
,	O
or	O
four	O
individual	O
graphics	B-Device
cards	I-Device
respectively	O
.	O
</s>
<s>
Two	O
GPUs	B-Architecture
on	O
one	O
graphics	B-Device
card	I-Device
.	O
</s>
<s>
This	O
configuration	O
has	O
the	O
advantage	O
of	O
implementing	O
two-way	O
SLI	O
,	O
while	O
only	O
occupying	O
one	O
PCI	B-Protocol
Express	O
slot	O
and	O
(	O
usually	O
)	O
two	O
expansion	O
I/O	O
slots	O
.	O
</s>
<s>
The	O
LED	O
bridges	O
can	O
only	O
function	O
at	O
the	O
increased	O
pixel	O
clock	O
if	O
the	O
GPU	B-Architecture
supports	O
that	O
clock	O
.	O
</s>
<s>
This	O
analyzes	O
the	O
rendered	O
image	O
in	O
order	O
to	O
split	O
the	O
workload	O
equally	O
between	O
the	O
two	O
GPUs	B-Architecture
.	O
</s>
<s>
For	O
example	O
,	O
in	O
a	O
scene	O
where	O
the	O
top	O
half	O
of	O
the	O
frame	O
is	O
mostly	O
empty	O
sky	O
,	O
the	O
dividing	O
line	O
will	O
lower	O
,	O
balancing	O
geometry	O
workload	O
between	O
the	O
two	O
GPUs	B-Architecture
.	O
</s>
<s>
Each	O
GPU	B-Architecture
renders	O
entire	O
frames	O
in	O
sequence	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
a	O
two-way	O
setup	O
,	O
one	O
GPU	B-Architecture
renders	O
the	O
odd	O
frames	O
,	O
the	O
other	O
the	O
even	O
frames	O
,	O
one	O
after	O
the	O
other	O
.	O
</s>
<s>
Ideally	O
,	O
this	O
would	O
result	O
in	O
the	O
rendering	O
time	O
being	O
cut	O
by	O
the	O
number	O
of	O
GPUs	B-Architecture
available	O
.	O
</s>
<s>
While	O
AFR	O
may	O
produce	O
higher	O
overall	O
framerates	O
than	O
SFR	O
,	O
it	O
also	O
exhibits	O
the	O
temporal	O
artifact	O
known	O
as	O
micro	B-General_Concept
stuttering	I-General_Concept
,	O
which	O
may	O
affect	O
frame	O
rate	O
perception	O
.	O
</s>
<s>
This	O
is	O
a	O
standalone	O
rendering	O
mode	O
that	O
offers	O
up	O
to	O
double	O
the	O
antialiasing	B-Algorithm
performance	O
by	O
splitting	O
the	O
antialiasing	B-Algorithm
workload	O
between	O
the	O
two	O
graphics	B-Device
cards	I-Device
,	O
offering	O
superior	O
image	O
quality	O
.	O
</s>
<s>
One	O
GPU	B-Architecture
performs	O
an	O
antialiasing	B-Algorithm
pattern	O
which	O
is	O
slightly	O
offset	O
to	O
the	O
usual	O
pattern	O
(	O
for	O
example	O
,	O
slightly	O
up	O
and	O
to	O
the	O
right	O
)	O
,	O
and	O
the	O
second	O
GPU	B-Architecture
uses	O
a	O
pattern	O
offset	O
by	O
an	O
equal	O
amount	O
in	O
the	O
opposite	O
direction	O
(	O
down	O
and	O
to	O
the	O
left	O
)	O
.	O
</s>
<s>
This	O
mode	O
is	O
not	O
intended	O
for	O
higher	O
frame	O
rates	O
,	O
and	O
can	O
actually	O
lower	O
performance	O
,	O
but	O
is	O
instead	O
intended	O
for	O
games	O
which	O
are	O
not	O
GPU-bound	O
,	O
offering	O
a	O
clearer	O
image	O
in	O
place	O
of	O
better	O
performance	O
.	O
</s>
<s>
When	O
enabled	O
,	O
SLI	O
antialiasing	B-Algorithm
offers	O
advanced	O
antialiasing	B-Algorithm
options	O
:	O
SLI	O
8×	O
,	O
SLI	O
16×	O
,	O
and	O
SLI	O
32×	O
(	O
for	O
quad	O
SLI	O
systems	O
only	O
)	O
.	O
</s>
<s>
GeForce	O
Boost	O
allows	O
the	O
rendering	O
power	O
of	O
an	O
integrated	O
graphics	B-Architecture
processor	I-Architecture
(	O
IGP	O
)	O
and	O
a	O
discrete	B-Device
GPU	I-Device
to	O
be	O
combined	O
in	O
order	O
to	O
increase	O
performance	O
.	O
</s>
<s>
The	O
setup	O
consists	O
of	O
an	O
IGP	O
as	O
well	O
as	O
a	O
GPU	B-Architecture
on	O
MXM	B-Architecture
module	O
.	O
</s>
<s>
The	O
IGP	O
would	O
assist	O
the	O
GPU	B-Architecture
to	O
boost	O
performance	O
when	O
the	O
laptop	O
is	O
plugged	O
to	O
a	O
power	O
socket	O
while	O
the	O
MXM	B-Architecture
module	O
would	O
be	O
shut	O
down	O
when	O
the	O
laptop	O
was	O
unplugged	O
from	O
power	O
socket	O
to	O
lower	O
overall	O
graphics	O
power	O
consumption	O
.	O
</s>
<s>
Hybrid	O
SLI	O
is	O
also	O
available	O
on	O
desktop	O
Motherboards	O
and	O
PCs	O
with	O
PCI-E	O
discrete	O
video	B-Device
cards	I-Device
.	O
</s>
<s>
NVIDIA	O
claims	O
that	O
twice	O
the	O
performance	O
can	O
be	O
achieved	O
with	O
a	O
Hybrid	O
SLI	O
capable	O
IGP	O
motherboard	O
and	O
a	O
GeForce	O
8400	O
GS	O
video	B-Device
card	I-Device
.	O
</s>
<s>
HybridPower	O
was	O
later	O
renamed	O
as	O
Nvidia	B-Application
Optimus	I-Application
.	O
</s>
<s>
In	O
May	O
2016	O
Nvidia	O
announced	O
that	O
the	O
GeForce	B-Device
10	I-Device
series	I-Device
would	O
feature	O
a	O
new	O
SLI	O
HB	O
(	O
High	O
Bandwidth	O
)	O
bridge	O
;	O
this	O
bridge	O
uses	O
2	O
SLI	O
fingers	O
on	O
the	O
PCB	O
of	O
each	O
card	O
and	O
essentially	O
doubles	O
the	O
available	O
bandwidth	O
between	O
them	O
.	O
</s>
<s>
Currently	O
,	O
only	O
GeForce	B-Device
10	I-Device
series	I-Device
cards	O
support	O
SLI	O
HB	O
and	O
only	O
2-way	O
SLI	O
is	O
supported	O
over	O
this	O
bridge	O
for	O
single-GPU	O
cards	O
.	O
</s>
<s>
Tests	O
with	O
a	O
GTX	B-Device
1080	I-Device
GPU-board	O
model	O
showed	O
that	O
the	O
improvements	O
in	O
gaming	O
performance	O
is	O
quite	O
marginal	O
.	O
</s>
<s>
Not	O
all	O
motherboards	O
with	O
multiple	O
PCI-Express	O
x16	O
slots	O
support	O
SLI	O
.	O
</s>
<s>
On	O
August	O
10	O
,	O
2009	O
Nvidia	O
announced	O
that	O
Intel	O
and	O
other	O
leading	O
motherboard	O
manufacturers	O
including	O
ASUS	O
,	O
EVGA	O
,	O
Gigabyte	O
and	O
MSI	O
have	O
all	O
licensed	O
Nvidia	B-Device
SLI	I-Device
technology	O
for	O
inclusion	O
on	O
their	O
Intel	O
P55	O
Express	O
Chipset-based	O
motherboards	O
designed	O
for	O
the	O
upcoming	O
Intel	O
Core	O
i7	O
and	O
i5	O
processor	O
in	O
the	O
LGA	B-Device
1156	I-Device
socket	O
.	O
</s>
<s>
Older	O
motherboards	O
using	O
the	O
P55	O
's	O
predecessors	O
Intel	B-Device
P35	I-Device
or	O
Intel	B-Device
P45	I-Device
do	O
not	O
support	O
SLI	O
.	O
</s>
<s>
Recent	O
motherboards	O
as	O
of	O
October	O
2017	O
that	O
support	O
it	O
are	O
Intel	O
's	O
Z	O
and	O
X	O
series	O
chipsets	B-Device
(	O
Z68	O
,	O
Z77	O
,	O
Z87	O
,	O
Z97	O
,	O
Z170	O
,	O
Z270	O
,	O
Z370	O
,	O
X79	O
,	O
X99	O
and	O
X299	O
)	O
along	O
with	O
AMD	O
's	O
990FX	O
,	O
X370	O
and	O
X399	O
chipsets	B-Device
.	O
</s>
<s>
Earlier	O
chipsets	B-Device
,	O
such	O
as	O
the	O
Intel	B-Device
X58	I-Device
,	O
could	O
support	O
2-way	O
SLI	O
over	O
16	O
lane	O
PCI-e	O
.	O
</s>
<s>
In	O
order	O
for	O
motherboards	O
of	O
that	O
generation	O
to	O
support	O
more	O
than	O
two	O
GPUs	B-Architecture
they	O
were	O
required	O
to	O
implement	O
Nvidia	O
nForce	B-Device
chipsets	B-Device
.	O
</s>
<s>
However	O
,	O
they	O
must	O
be	O
of	O
the	O
same	O
GPU	B-Architecture
series	O
(	O
e.g.	O
,	O
8600	O
,	O
8800	O
)	O
and	O
GPU	B-Architecture
model	O
name	O
(	O
e.g.	O
,	O
GT	O
,	O
GTS	O
,	O
GTX	O
)	O
.	O
</s>
<s>
Another	O
exception	O
is	O
the	O
GTS	O
250	O
,	O
which	O
can	O
be	O
paired	O
with	O
the	O
9800	O
GTX+	O
,	O
as	O
the	O
GTS	O
250	O
GPU	B-Architecture
is	O
a	O
rebadged	O
9800	O
GTX+	O
GPU	B-Architecture
.	O
</s>
<s>
This	O
is	O
also	O
true	O
for	O
AMD	O
's	O
CrossFire	O
,	O
as	O
the	O
problem	O
is	O
inherent	O
in	O
multi-GPU	O
systems	O
.	O
</s>
<s>
When	O
using	O
SLI	O
with	O
AFR	O
,	O
the	O
subjective	O
framerate	O
can	O
often	O
be	O
lower	O
than	O
the	O
framerate	O
reported	O
by	O
benchmarking	O
applications	O
,	O
and	O
may	O
even	O
be	O
poorer	O
than	O
the	O
framerate	O
of	O
its	O
single-GPU	O
equivalent	O
.	O
</s>
<s>
This	O
phenomenon	O
is	O
known	O
as	O
micro	B-General_Concept
stuttering	I-General_Concept
and	O
also	O
applies	O
to	O
CrossFire	O
since	O
it	O
is	O
inherent	O
to	O
multi-GPU	O
configurations	O
.	O
</s>
<s>
With	O
the	O
new	O
RTX	O
20xx	O
series	O
of	O
graphics	B-Device
cards	I-Device
as	O
launched	O
in	O
2018	O
the	O
interconnect	O
is	O
no	O
longer	O
SLI	O
HB	O
.	O
</s>
<s>
As	O
of	O
now	O
only	O
two	O
GPU	B-Architecture
cards	O
can	O
be	O
connected	O
with	O
NVLink	O
;	O
three-way	O
,	O
four-way	O
and	O
quad	O
are	O
not	O
possible	O
using	O
NVLink	O
bridges	O
even	O
if	O
NVLink	O
by	O
principle	O
is	O
a	O
very	O
versatile	O
interface	O
.	O
</s>
<s>
Eventually	O
,	O
for	O
the	O
40	O
series	O
;	O
this	O
port	O
was	O
also	O
dropped	O
,	O
officially	O
marking	O
the	O
end	O
of	O
the	O
era	O
of	O
multi-GPU	O
solutions	O
in	O
the	O
consumer	O
field	O
.	O
</s>
