<s>
Deep	B-Device
learning	I-Device
super	I-Device
sampling	I-Device
(	O
DLSS	B-Device
)	O
is	O
a	O
family	O
of	O
real-time	B-General_Concept
deep	O
learning	O
image	O
enhancement	O
and	O
upscaling	B-Algorithm
technologies	O
developed	O
by	O
Nvidia	O
that	O
are	O
exclusive	O
to	O
its	O
RTX	B-Algorithm
line	O
of	O
graphics	B-Device
cards	I-Device
,	O
and	O
available	O
in	O
a	O
number	O
of	O
video	O
games	O
.	O
</s>
<s>
The	O
goal	O
of	O
these	O
technologies	O
is	O
to	O
allow	O
the	O
majority	O
of	O
the	O
graphics	B-General_Concept
pipeline	I-General_Concept
to	O
run	O
at	O
a	O
lower	O
resolution	B-General_Concept
for	O
increased	O
performance	O
,	O
and	O
then	O
infer	O
a	O
higher	O
resolution	B-General_Concept
image	O
from	O
this	O
that	O
contains	O
the	O
same	O
level	O
of	O
detail	O
as	O
if	O
the	O
image	O
had	O
been	O
rendered	O
at	O
this	O
higher	O
resolution	B-General_Concept
.	O
</s>
<s>
This	O
allows	O
for	O
higher	O
graphical	O
settings	O
and/or	O
frame	O
rates	O
for	O
a	O
given	O
output	O
resolution	B-General_Concept
,	O
depending	O
on	O
user	O
preference	O
.	O
</s>
<s>
As	O
of	O
September	O
2022	O
,	O
the	O
1st	O
and	O
2nd	O
generation	O
of	O
DLSS	B-Device
is	O
available	O
on	O
all	O
RTX	B-Algorithm
branded	O
cards	O
from	O
Nvidia	O
in	O
supported	O
titles	O
,	O
while	O
the	O
3rd	O
generation	O
unveiled	O
at	O
Nvidia	O
's	O
GTC	O
2022	O
event	O
is	O
exclusive	O
to	O
Ada	B-General_Concept
Lovelace	I-General_Concept
generation	O
RTX	B-Algorithm
4000	O
series	O
graphics	B-Device
cards	I-Device
.	O
</s>
<s>
Nvidia	O
has	O
also	O
introduced	O
Deep	O
learning	O
dynamic	O
super	O
resolution	B-General_Concept
(	O
DLDSR	O
)	O
,	O
a	O
related	O
and	O
opposite	O
technology	O
where	O
the	O
graphics	O
are	O
rendered	O
at	O
a	O
higher	O
resolution	B-General_Concept
,	O
then	O
downsampled	O
to	O
the	O
native	O
display	B-General_Concept
resolution	I-General_Concept
using	O
an	O
AI-assisted	O
downsampling	O
algorithm	O
to	O
achieve	O
higher	O
image	O
quality	O
than	O
rendering	O
at	O
native	O
resolution	B-General_Concept
.	O
</s>
<s>
Nvidia	O
advertised	O
DLSS	B-Device
as	O
a	O
key	O
feature	O
of	O
the	O
GeForce	B-Device
RTX	I-Device
20	I-Device
series	O
cards	O
when	O
they	O
launched	O
in	O
September	O
2018	O
.	O
</s>
<s>
At	O
that	O
time	O
,	O
the	O
results	O
were	O
limited	O
to	O
a	O
few	O
video	O
games	O
(	O
namely	O
Battlefield	B-Application
V	I-Application
and	O
Metro	B-Application
Exodus	I-Application
)	O
because	O
the	O
algorithm	O
had	O
to	O
be	O
trained	O
specifically	O
on	O
each	O
game	O
on	O
which	O
it	O
was	O
applied	O
and	O
the	O
results	O
were	O
usually	O
not	O
as	O
good	O
as	O
simple	O
resolution	B-General_Concept
upscaling	B-Algorithm
.	O
</s>
<s>
In	O
2019	O
,	O
the	O
video	O
game	O
Control	B-Application
shipped	O
with	O
ray	B-Algorithm
tracing	I-Algorithm
and	O
an	O
improved	O
version	O
of	O
DLSS	B-Device
,	O
which	O
did	O
not	O
use	O
the	O
Tensor	O
Cores	O
.	O
</s>
<s>
In	O
April	O
2020	O
,	O
Nvidia	O
advertised	O
and	O
shipped	O
an	O
improved	O
version	O
of	O
DLSS	B-Device
named	O
DLSS	B-Device
2.0	O
with	O
driver	B-Application
version	O
445.75	O
.	O
</s>
<s>
DLSS	B-Device
2.0	O
was	O
available	O
for	O
a	O
few	O
existing	O
games	O
including	O
Control	B-Application
and	O
Wolfenstein	B-Application
:	I-Application
Youngblood	I-Application
,	O
and	O
would	O
later	O
be	O
added	O
to	O
many	O
newly	O
released	O
games	O
and	O
game	O
engines	O
such	O
as	O
Unreal	B-Operating_System
Engine	I-Operating_System
and	O
Unity	B-Application
.	O
</s>
<s>
Despite	O
sharing	O
the	O
DLSS	B-Device
branding	O
,	O
the	O
two	O
iterations	O
of	O
DLSS	B-Device
differ	O
significantly	O
and	O
are	O
not	O
backwards-compatible	O
.	O
</s>
<s>
The	O
first	O
iteration	O
of	O
DLSS	B-Device
is	O
a	O
predominantly	O
spatial	O
image	O
upscaler	O
with	O
two	O
stages	O
,	O
both	O
relying	O
on	O
convolutional	B-Architecture
auto-encoder	B-Algorithm
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
The	O
first	O
step	O
is	O
an	O
image	O
enhancement	O
network	O
which	O
uses	O
the	O
current	O
frame	O
and	O
motion	O
vectors	O
to	O
perform	O
edge	B-Algorithm
enhancement	I-Algorithm
,	O
and	O
spatial	B-Algorithm
anti-aliasing	I-Algorithm
.	O
</s>
<s>
The	O
second	O
stage	O
is	O
an	O
image	B-Algorithm
upscaling	I-Algorithm
step	O
which	O
uses	O
the	O
single	O
raw	O
,	O
low-resolution	O
frame	O
to	O
upscale	O
the	O
image	O
to	O
the	O
desired	O
output	O
resolution	B-General_Concept
.	O
</s>
<s>
Using	O
just	O
a	O
single	O
frame	O
for	O
upscaling	B-Algorithm
means	O
the	O
neural	B-Architecture
network	I-Architecture
itself	O
must	O
generate	O
a	O
large	O
amount	O
of	O
new	O
information	O
to	O
produce	O
the	O
high	O
resolution	B-General_Concept
output	O
,	O
this	O
can	O
result	O
in	O
slight	O
hallucinations	B-General_Concept
such	O
as	O
leaves	O
that	O
differ	O
in	O
style	O
to	O
the	O
source	O
content	O
.	O
</s>
<s>
The	O
neural	B-Architecture
networks	I-Architecture
are	O
trained	O
on	O
a	O
per-game	O
basis	O
by	O
generating	O
a	O
"	O
perfect	O
frame	O
"	O
using	O
traditional	O
supersampling	B-Algorithm
to	O
64	O
samples	O
per	O
pixel	O
,	O
as	O
well	O
as	O
the	O
motion	O
vectors	O
for	O
each	O
frame	O
.	O
</s>
<s>
This	O
data	O
is	O
also	O
augmented	B-General_Concept
using	O
common	O
augmentations	O
such	O
as	O
rotations	O
,	O
colour	O
changes	O
,	O
and	O
random	O
noise	O
to	O
help	O
generalize	O
the	O
test	O
data	O
.	O
</s>
<s>
This	O
first	O
iteration	O
received	O
a	O
mixed	O
response	O
,	O
with	O
many	O
criticizing	O
the	O
often	O
soft	O
appearance	O
and	O
artifacting	O
in	O
certain	O
situations	O
;	O
likely	O
a	O
side	O
effect	O
of	O
the	O
limited	O
data	O
from	O
only	O
using	O
a	O
single	O
frame	O
input	O
to	O
the	O
neural	B-Architecture
networks	I-Architecture
which	O
could	O
not	O
be	O
trained	O
to	O
perform	O
optimally	O
in	O
all	O
scenarios	O
and	O
edge-cases	O
.	O
</s>
<s>
Nvidia	O
also	O
demonstrated	O
the	O
ability	O
for	O
the	O
auto-encoder	B-Algorithm
networks	O
to	O
learn	O
the	O
ability	O
to	O
recreate	O
depth-of-field	O
and	O
motion	O
blur	O
,	O
although	O
this	O
functionality	O
has	O
never	O
been	O
included	O
in	O
a	O
publicly	O
released	O
product	O
.	O
</s>
<s>
DLSS	B-Device
2.0	O
is	O
a	O
temporal	B-Algorithm
anti-aliasing	I-Algorithm
upsampling	B-Algorithm
(	O
TAAU	O
)	O
implementation	O
,	O
using	O
data	O
from	O
previous	O
frames	O
extensively	O
through	O
sub-pixel	O
jittering	O
to	O
resolve	O
fine	O
detail	O
and	O
reduce	O
aliasing	O
.	O
</s>
<s>
The	O
data	O
DLSS	B-Device
2.0	O
collects	O
includes	O
:	O
the	O
raw	O
low-resolution	O
input	O
,	O
motion	O
vectors	O
,	O
depth	O
buffers	O
,	O
and	O
exposure	O
/	O
brightness	O
information	O
.	O
</s>
<s>
It	O
can	O
also	O
be	O
used	O
as	O
a	O
simpler	O
TAA	O
implementation	O
where	O
the	O
image	O
is	O
rendered	O
at	O
100%	O
resolution	B-General_Concept
,	O
rather	O
than	O
being	O
upsampled	B-Algorithm
by	O
DLSS	B-Device
,	O
Nvidia	O
brands	O
this	O
as	O
DLAA	O
(	O
Deep	B-Algorithm
Learning	I-Algorithm
Anti-Aliasing	I-Algorithm
)	O
.	O
</s>
<s>
TAA(U )	O
is	O
used	O
in	O
many	O
modern	O
video	O
games	O
and	O
game	O
engines	O
,	O
however	O
all	O
previous	O
implementations	O
have	O
used	O
some	O
form	O
of	O
manually	O
written	O
heuristics	B-Algorithm
to	O
prevent	O
temporal	O
artifacts	O
such	O
as	O
ghosting	O
and	O
flickering	O
.	O
</s>
<s>
This	O
helps	O
to	O
identify	O
and	O
fix	O
many	O
temporal	O
artifacts	O
,	O
but	O
deliberately	O
removing	O
fine	O
details	O
in	O
this	O
way	O
is	O
analogous	O
to	O
applying	O
a	O
blur	B-Algorithm
filter	I-Algorithm
,	O
and	O
thus	O
the	O
final	O
image	O
can	O
appear	O
blurry	O
when	O
using	O
this	O
method	O
.	O
</s>
<s>
DLSS	B-Device
2.0	O
uses	O
a	O
convolutional	B-Architecture
auto-encoder	B-Algorithm
neural	B-Architecture
network	I-Architecture
trained	O
to	O
identify	O
and	O
fix	O
temporal	O
artifacts	O
,	O
instead	O
of	O
manually	O
programmed	O
heuristics	B-Algorithm
as	O
mentioned	O
above	O
.	O
</s>
<s>
Because	O
of	O
this	O
,	O
DLSS	B-Device
2.0	O
can	O
generally	O
resolve	O
detail	O
better	O
than	O
other	O
TAA	O
and	O
TAAU	O
implementations	O
,	O
while	O
also	O
removing	O
most	O
temporal	O
artifacts	O
.	O
</s>
<s>
This	O
is	O
why	O
DLSS	B-Device
2.0	O
can	O
sometimes	O
produce	O
a	O
sharper	O
image	O
than	O
rendering	O
at	O
higher	O
,	O
or	O
even	O
native	O
resolutions	O
using	O
traditional	O
TAA	O
.	O
</s>
<s>
However	O
,	O
no	O
temporal	O
solution	O
is	O
perfect	O
,	O
and	O
artifacts	O
(	O
ghosting	O
in	O
particular	O
)	O
are	O
still	O
visible	O
in	O
some	O
scenarios	O
when	O
using	O
DLSS	B-Device
2.0	O
.	O
</s>
<s>
Because	O
temporal	O
artifacts	O
occur	O
in	O
most	O
art	O
styles	O
and	O
environments	O
in	O
broadly	O
the	O
same	O
way	O
,	O
the	O
neural	B-Architecture
network	I-Architecture
that	O
powers	O
DLSS	B-Device
2.0	O
does	O
not	O
need	O
to	O
be	O
retrained	O
when	O
being	O
used	O
in	O
different	O
games	O
.	O
</s>
<s>
Despite	O
this	O
,	O
Nvidia	O
does	O
frequently	O
ship	O
new	O
minor	O
revisions	O
of	O
DLSS	B-Device
2.0	O
with	O
new	O
titles	O
,	O
so	O
this	O
could	O
suggest	O
some	O
minor	O
training	O
optimizations	O
may	O
be	O
performed	O
as	O
games	O
are	O
released	O
,	O
although	O
Nvidia	O
does	O
not	O
provide	O
changelogs	O
for	O
these	O
minor	O
revisions	O
to	O
confirm	O
this	O
.	O
</s>
<s>
The	O
main	O
advancements	O
compared	O
to	O
DLSS	B-Device
1.0	O
include	O
:	O
Significantly	O
improved	O
detail	O
retention	O
,	O
a	O
generalized	O
neural	B-Architecture
network	I-Architecture
that	O
does	O
not	O
need	O
to	O
be	O
re-trained	O
per-game	O
,	O
and	O
~	O
2x	O
less	O
overhead	O
(	O
~	O
1-2ms	O
vs	O
~	O
2-4ms	O
)	O
.	O
</s>
<s>
It	O
should	O
also	O
be	O
noted	O
that	O
forms	O
of	O
TAAU	O
such	O
as	O
DLSS	B-Device
2.0	O
are	O
not	O
upscalers	O
in	O
the	O
same	O
sense	O
as	O
techniques	O
such	O
as	O
ESRGAN	O
or	O
DLSS	B-Device
1.0	O
,	O
which	O
attempt	O
to	O
create	O
new	O
information	O
from	O
a	O
low-resolution	O
source	O
;	O
instead	O
TAAU	O
works	O
to	O
recover	O
data	O
from	O
previous	O
frames	O
,	O
rather	O
than	O
creating	O
new	O
data	O
.	O
</s>
<s>
In	O
practice	O
,	O
this	O
means	O
low	O
resolution	B-General_Concept
textures	O
in	O
games	O
will	O
still	O
appear	O
low-resolution	O
when	O
using	O
current	O
TAAU	O
techniques	O
.	O
</s>
<s>
This	O
is	O
why	O
Nvidia	O
recommends	O
game	O
developers	O
use	O
higher	O
resolution	B-General_Concept
textures	O
than	O
they	O
would	O
normally	O
for	O
a	O
given	O
rendering	O
resolution	B-General_Concept
by	O
applying	O
a	O
mip-map	O
bias	O
when	O
DLSS	B-Device
2.0	O
is	O
enabled	O
.	O
</s>
<s>
Augments	O
DLSS	B-Device
2.0	O
by	O
making	O
use	O
of	O
an	O
optical-flow	O
frame	O
generation	O
technique	O
.	O
</s>
<s>
The	O
DLSS	B-Device
frame	O
generation	O
algorithm	O
takes	O
two	O
rendered	O
frames	O
from	O
the	O
rendering	B-General_Concept
pipeline	I-General_Concept
,	O
and	O
generates	O
a	O
new	O
frame	O
that	O
smoothly	O
transitions	O
between	O
them	O
.	O
</s>
<s>
DLSS	B-Device
3.0	O
makes	O
use	O
of	O
a	O
new	O
generation	O
Optical	O
Flow	O
Accelerator	O
(	O
OFA	O
)	O
included	O
in	O
Ada	B-General_Concept
Lovelace	I-General_Concept
generation	O
RTX	B-Algorithm
GPUs	B-Architecture
.	O
</s>
<s>
The	O
new	O
OFA	O
is	O
faster	O
and	O
more	O
accurate	O
than	O
the	O
OFA	O
already	O
available	O
in	O
previous	O
Turing	O
and	O
Ampere	O
RTX	B-Algorithm
GPUs	B-Architecture
.	O
</s>
<s>
This	O
results	O
in	O
DLSS	B-Device
3.0	O
being	O
exclusive	O
for	O
the	O
RTX	B-Algorithm
4000	O
Series	O
.	O
</s>
<s>
At	O
release	O
,	O
DLSS	B-Device
3.0	O
does	O
not	O
work	O
for	O
VR	O
displays	O
.	O
</s>
<s>
DLSS	B-Device
requires	O
and	O
applies	O
its	O
own	O
anti-aliasing	O
method	O
.	O
</s>
<s>
Unlike	O
TAA	O
,	O
DLSS	B-Device
does	O
not	O
sample	O
every	O
pixel	O
in	O
every	O
frame	O
.	O
</s>
<s>
DLSS	B-Device
uses	O
machine	O
learning	O
to	O
combine	O
samples	O
in	O
the	O
current	O
frame	O
and	O
past	O
frames	O
,	O
and	O
it	O
can	O
be	O
thought	O
of	O
as	O
an	O
advanced	O
and	O
superior	O
TAA	O
implementation	O
made	O
possible	O
by	O
the	O
available	O
tensor	O
cores	O
.	O
</s>
<s>
Nvidia	O
offers	O
deep	B-Algorithm
learning	I-Algorithm
anti-aliasing	I-Algorithm
(	O
DLAA	O
)	O
.	O
</s>
<s>
DLAA	O
provides	O
the	O
same	O
AI-driven	O
anti-aliasing	O
DLSS	B-Device
uses	O
,	O
but	O
without	O
any	O
upscaling	B-Algorithm
or	O
downscaling	O
functionality	O
.	O
</s>
<s>
With	O
the	O
exception	O
of	O
the	O
shader-core	O
version	O
implemented	O
in	O
Control	B-Application
,	O
DLSS	B-Device
is	O
only	O
available	O
on	O
GeForce	B-Device
RTX	I-Device
20	I-Device
,	O
GeForce	B-Device
RTX	I-Device
30	I-Device
,	O
GeForce	B-Device
RTX	I-Device
40	I-Device
,	O
and	O
Quadro	O
RTX	B-Algorithm
series	O
of	O
video	B-Device
cards	I-Device
,	O
using	O
dedicated	O
AI	B-General_Concept
accelerators	I-General_Concept
called	O
Tensor	O
Cores	O
.	O
</s>
<s>
Tensor	O
Cores	O
are	O
available	O
since	O
the	O
Nvidia	B-General_Concept
Volta	I-General_Concept
GPU	B-Architecture
microarchitecture	B-General_Concept
,	O
which	O
was	O
first	O
used	O
on	O
the	O
Tesla	B-Device
V100	I-Device
line	O
of	O
products	O
.	O
</s>
<s>
They	O
are	O
used	O
for	O
doing	O
fused	B-Algorithm
multiply-add	I-Algorithm
(	O
FMA	O
)	O
operations	O
that	O
are	O
used	O
extensively	O
in	O
neural	B-Architecture
network	I-Architecture
calculations	O
for	O
applying	O
a	O
large	O
series	O
of	O
multiplications	O
on	O
weights	O
,	O
followed	O
by	O
the	O
addition	O
of	O
a	O
bias	O
.	O
</s>
<s>
Each	O
core	O
can	O
do	O
1024	O
bits	O
of	O
FMA	O
operations	O
per	O
clock	O
,	O
so	O
1024	O
INT1	O
,	O
256	O
INT4	O
,	O
128	O
INT8	O
,	O
and	O
64	O
FP16	O
operations	O
per	O
clock	O
per	O
tensor	O
core	O
,	O
and	O
most	O
Turing	O
GPUs	B-Architecture
have	O
a	O
few	O
hundred	O
tensor	O
cores	O
.	O
</s>
<s>
The	O
Tensor	O
Cores	O
use	O
CUDA	O
Warp-Level	O
Primitives	O
on	O
32	O
parallel	O
threads	B-Operating_System
to	O
take	O
advantage	O
of	O
their	O
parallel	O
architecture	O
.	O
</s>
<s>
A	O
Warp	O
is	O
a	O
set	O
of	O
32	O
threads	B-Operating_System
which	O
are	O
configured	O
to	O
execute	O
the	O
same	O
instruction	O
.	O
</s>
