<s>
General-purpose	B-Architecture
computing	I-Architecture
on	I-Architecture
graphics	I-Architecture
processing	I-Architecture
units	I-Architecture
(	O
GPGPU	B-Architecture
,	O
or	O
less	O
often	O
GPGP	O
)	O
is	O
the	O
use	O
of	O
a	O
graphics	B-Architecture
processing	I-Architecture
unit	I-Architecture
(	O
GPU	B-Architecture
)	O
,	O
which	O
typically	O
handles	O
computation	O
only	O
for	O
computer	O
graphics	O
,	O
to	O
perform	O
computation	O
in	O
applications	O
traditionally	O
handled	O
by	O
the	O
central	B-General_Concept
processing	I-General_Concept
unit	I-General_Concept
(	O
CPU	O
)	O
.	O
</s>
<s>
The	O
use	O
of	O
multiple	O
video	B-Device
cards	I-Device
in	O
one	O
computer	O
,	O
or	O
large	O
numbers	O
of	O
graphics	O
chips	O
,	O
further	O
parallelizes	O
the	O
already	O
parallel	O
nature	O
of	O
graphics	O
processing	O
.	O
</s>
<s>
Essentially	O
,	O
a	O
GPGPU	B-Architecture
pipeline	B-General_Concept
is	O
a	O
kind	O
of	O
parallel	B-Operating_System
processing	I-Operating_System
between	O
one	O
or	O
more	O
GPUs	B-Architecture
and	O
CPUs	O
that	O
analyzes	O
data	O
as	O
if	O
it	O
were	O
in	O
image	O
or	O
other	O
graphic	O
form	O
.	O
</s>
<s>
While	O
GPUs	B-Architecture
operate	O
at	O
lower	O
frequencies	O
,	O
they	O
typically	O
have	O
many	O
times	O
the	O
number	O
of	O
cores	B-Architecture
.	O
</s>
<s>
Thus	O
,	O
GPUs	B-Architecture
can	O
process	O
far	O
more	O
pictures	O
and	O
graphical	O
data	O
per	O
second	O
than	O
a	O
traditional	O
CPU	O
.	O
</s>
<s>
Migrating	O
data	O
into	O
graphical	O
form	O
and	O
then	O
using	O
the	O
GPU	B-Architecture
to	O
scan	O
and	O
analyze	O
it	O
can	O
create	O
a	O
large	O
speedup	B-Operating_System
.	O
</s>
<s>
GPGPU	B-Architecture
pipelines	O
were	O
developed	O
at	O
the	O
beginning	O
of	O
the	O
21st	O
century	O
for	O
graphics	O
processing	O
(	O
e.g.	O
</s>
<s>
In	O
1987	O
,	O
Conway	O
's	O
Game	O
of	O
Life	O
became	O
one	O
of	O
the	O
first	O
examples	O
of	O
general-purpose	O
computing	O
using	O
an	O
early	O
stream	B-Application
processor	I-Application
called	O
a	O
blitter	B-General_Concept
to	O
invoke	O
a	O
special	O
sequence	O
of	O
logical	B-Algorithm
operations	I-Algorithm
on	O
bit	O
vectors	O
.	O
</s>
<s>
General-purpose	O
computing	O
on	O
GPUs	B-Architecture
became	O
more	O
practical	O
and	O
popular	O
after	O
about	O
2001	O
,	O
with	O
the	O
advent	O
of	O
both	O
programmable	O
shaders	O
and	O
floating	B-Algorithm
point	I-Algorithm
support	O
on	O
graphics	B-Architecture
processors	I-Architecture
.	O
</s>
<s>
Notably	O
,	O
problems	O
involving	O
matrices	B-Architecture
and/or	O
vectors	O
especially	O
two-	O
,	O
three-	O
,	O
or	O
four-dimensional	O
vectors	O
were	O
easy	O
to	O
translate	O
to	O
a	O
GPU	B-Architecture
,	O
which	O
acts	O
with	O
native	O
speed	O
and	O
support	O
on	O
those	O
types	O
.	O
</s>
<s>
A	O
significant	O
milestone	O
for	O
GPGPU	B-Architecture
was	O
the	O
year	O
2003	O
when	O
two	O
research	O
groups	O
independently	O
discovered	O
GPU-based	O
approaches	O
for	O
the	O
solution	O
of	O
general	O
linear	O
algebra	O
problems	O
on	O
GPUs	B-Architecture
that	O
ran	O
faster	O
than	O
on	O
CPUs	O
.	O
</s>
<s>
These	O
early	O
efforts	O
to	O
use	O
GPUs	B-Architecture
as	O
general-purpose	O
processors	O
required	O
reformulating	O
computational	O
problems	O
in	O
terms	O
of	O
graphics	O
primitives	O
,	O
as	O
supported	O
by	O
the	O
two	O
major	O
APIs	B-Application
for	O
graphics	B-Architecture
processors	I-Architecture
,	O
OpenGL	B-Application
and	O
DirectX	B-Application
.	O
</s>
<s>
This	O
cumbersome	O
translation	O
was	O
obviated	O
by	O
the	O
advent	O
of	O
general-purpose	O
programming	O
languages	O
and	O
APIs	B-Application
such	O
as	O
Sh/RapidMind	O
,	O
Brook	B-Library
and	O
Accelerator	O
.	O
</s>
<s>
These	O
were	O
followed	O
by	O
Nvidia	O
's	O
CUDA	B-Architecture
,	O
which	O
allowed	O
programmers	O
to	O
ignore	O
the	O
underlying	O
graphical	O
concepts	O
in	O
favor	O
of	O
more	O
common	O
high-performance	B-Architecture
computing	I-Architecture
concepts	O
.	O
</s>
<s>
Newer	O
,	O
hardware-vendor-independent	O
offerings	O
include	O
Microsoft	O
's	O
DirectCompute	B-Library
and	O
Apple/Khronos	O
Group	O
's	O
OpenCL	B-Application
.	O
</s>
<s>
This	O
means	O
that	O
modern	O
GPGPU	B-Architecture
pipelines	O
can	O
leverage	O
the	O
speed	O
of	O
a	O
GPU	B-Architecture
without	O
requiring	O
full	O
and	O
explicit	O
conversion	O
of	O
the	O
data	O
to	O
a	O
graphical	O
form	O
.	O
</s>
<s>
Mark	O
Harris	O
,	O
the	O
founder	O
of	O
GPGPU.org,	O
coined	O
the	O
term	O
GPGPU	B-Architecture
.	O
</s>
<s>
Any	O
language	O
that	O
allows	O
the	O
code	O
running	O
on	O
the	O
CPU	O
to	O
poll	O
a	O
GPU	B-Architecture
shader	O
for	O
return	O
values	O
,	O
can	O
create	O
a	O
GPGPU	B-Architecture
framework	O
.	O
</s>
<s>
Programming	O
standards	O
for	O
parallel	B-Operating_System
computing	I-Operating_System
include	O
OpenCL	B-Application
(	O
vendor-independent	O
)	O
,	O
OpenACC	B-Operating_System
,	O
OpenMP	B-Application
and	O
OpenHMPP	B-Application
.	O
</s>
<s>
,	O
OpenCL	B-Application
is	O
the	O
dominant	O
open	O
general-purpose	O
GPU	B-Architecture
computing	I-Architecture
language	O
,	O
and	O
is	O
an	O
open	O
standard	O
defined	O
by	O
the	B-Library
Khronos	I-Library
Group	I-Library
.	O
</s>
<s>
OpenCL	B-Application
provides	O
a	O
cross-platform	B-Operating_System
GPGPU	B-Architecture
platform	O
that	O
additionally	O
supports	O
data	O
parallel	O
compute	O
on	O
CPUs	O
.	O
</s>
<s>
OpenCL	B-Application
is	O
actively	O
supported	O
on	O
Intel	O
,	O
AMD	O
,	O
Nvidia	O
,	O
and	O
ARM	O
platforms	O
.	O
</s>
<s>
The	B-Library
Khronos	I-Library
Group	I-Library
has	O
also	O
standardised	O
and	O
implemented	O
SYCL	B-Language
,	O
a	O
higher-level	O
programming	O
model	O
for	O
OpenCL	B-Application
as	O
a	O
single-source	O
domain	O
specific	O
embedded	O
language	O
based	O
on	O
pure	O
C++11	O
.	O
</s>
<s>
The	O
dominant	O
proprietary	O
framework	O
is	O
Nvidia	B-Architecture
CUDA	I-Architecture
.	O
</s>
<s>
Nvidia	O
launched	O
CUDA	B-Architecture
in	O
2006	O
,	O
a	O
software	B-Application
development	I-Application
kit	I-Application
(	O
SDK	O
)	O
and	O
application	B-Application
programming	I-Application
interface	I-Application
(	O
API	B-Application
)	O
that	O
allows	O
using	O
the	O
programming	B-Language
language	I-Language
C	I-Language
to	O
code	O
algorithms	O
for	O
execution	O
on	O
GeForce	O
8	O
series	O
and	O
later	O
GPUs	B-Architecture
.	O
</s>
<s>
ROCm	B-Operating_System
,	O
launched	O
in	O
2016	O
,	O
is	O
AMD	O
's	O
open-source	O
response	O
to	O
CUDA	B-Architecture
.	O
</s>
<s>
It	O
is	O
,	O
as	O
of	O
2022	O
,	O
on	O
par	O
with	O
CUDA	B-Architecture
with	O
regards	O
to	O
features	O
,	O
and	O
still	O
lacking	O
in	O
consumer	O
support	O
.	O
</s>
<s>
Altimesh	O
Hybridizer	O
created	O
by	O
Altimesh	O
compiles	O
Common	O
Intermediate	O
Language	O
to	O
CUDA	B-Architecture
binaries	O
.	O
</s>
<s>
Debugging	O
and	O
profiling	O
is	O
integrated	O
with	O
Visual	B-Application
Studio	I-Application
and	O
Nsight	O
.	O
</s>
<s>
It	O
is	O
available	O
as	O
a	O
Visual	B-Application
Studio	I-Application
extension	O
on	O
Visual	B-Application
Studio	I-Application
Marketplace	O
.	O
</s>
<s>
Microsoft	O
introduced	O
the	O
DirectCompute	B-Library
GPU	B-Architecture
computing	I-Architecture
API	B-Application
,	O
released	O
with	O
the	O
DirectX	B-Application
11	I-Application
API	B-Application
.	O
</s>
<s>
,	O
created	O
by	O
QuantAlea	O
,	O
introduces	O
native	O
GPU	B-Architecture
computing	I-Architecture
capabilities	O
for	O
the	O
Microsoft	O
.NET	O
languages	O
F#	B-Operating_System
and	O
C#	B-Application
.	O
</s>
<s>
Alea	O
GPU	B-Architecture
also	O
provides	O
a	O
simplified	O
GPU	B-Architecture
programming	O
model	O
based	O
on	O
GPU	B-Architecture
parallel-for	O
and	O
parallel	O
aggregate	O
using	O
delegates	O
and	O
automatic	O
memory	O
management	O
.	O
</s>
<s>
MATLAB	B-Language
supports	O
GPGPU	B-Architecture
acceleration	O
using	O
the	O
Parallel	B-Operating_System
Computing	I-Operating_System
Toolbox	O
and	O
MATLAB	B-Language
Distributed	B-Architecture
Computing	I-Architecture
Server	O
,	O
and	O
third-party	O
packages	O
like	O
Jacket	O
.	O
</s>
<s>
GPGPU	B-Architecture
processing	O
is	O
also	O
used	O
to	O
simulate	O
Newtonian	O
physics	O
by	O
physics	O
engines	O
,	O
and	O
commercial	O
implementations	O
include	O
Havok	B-General_Concept
Physics	I-General_Concept
,	I-General_Concept
FX	I-General_Concept
and	O
PhysX	B-Operating_System
,	O
both	O
of	O
which	O
are	O
typically	O
used	O
for	O
computer	O
and	O
video	O
games	O
.	O
</s>
<s>
C++	B-Language
Accelerated	O
Massive	O
Parallelism	B-Operating_System
(	O
C++	B-Language
AMP	I-Language
)	O
is	O
a	O
library	O
that	O
accelerates	O
execution	O
of	O
C++	B-Language
code	I-Language
by	O
exploiting	O
the	O
data-parallel	O
hardware	O
on	O
GPUs	B-Architecture
.	O
</s>
<s>
Due	O
to	O
a	O
trend	O
of	O
increasing	O
power	O
of	O
mobile	O
GPUs	B-Architecture
,	O
general-purpose	O
programming	O
became	O
available	O
also	O
on	O
the	O
mobile	O
devices	O
running	O
major	O
mobile	B-Operating_System
operating	I-Operating_System
systems	I-Operating_System
.	O
</s>
<s>
Google	B-Application
Android	I-Application
4.2	O
enabled	O
running	O
RenderScript	B-Application
code	O
on	O
the	O
mobile	O
device	O
GPU	B-Architecture
.	O
</s>
<s>
Apple	O
introduced	O
the	O
proprietary	O
Metal	B-Library
API	I-Library
for	O
iOS	B-Application
applications	O
,	O
able	O
to	O
execute	O
arbitrary	O
code	O
through	O
Apple	O
's	O
GPU	B-Architecture
compute	B-Operating_System
shaders	I-Operating_System
.	O
</s>
<s>
Computer	O
video	B-Device
cards	I-Device
are	O
produced	O
by	O
various	O
vendors	O
,	O
such	O
as	O
Nvidia	O
,	O
AMD	O
.	O
</s>
<s>
Cards	O
from	O
such	O
vendors	O
differ	O
on	O
implementing	O
data-format	O
support	O
,	O
such	O
as	O
integer	O
and	O
floating-point	B-Algorithm
formats	I-Algorithm
(	O
32-bit	O
and	O
64-bit	O
)	O
.	O
</s>
<s>
Microsoft	O
introduced	O
a	O
Shader	O
Model	O
standard	O
,	O
to	O
help	O
rank	O
the	O
various	O
features	O
of	O
graphic	B-Device
cards	I-Device
into	O
a	O
simple	O
Shader	O
Model	O
version	O
number	O
(	O
1.0	O
,	O
2.0	O
,	O
3.0	O
,	O
etc	O
.	O
</s>
<s>
Pre-DirectX	O
9	O
video	B-Device
cards	I-Device
only	O
supported	O
paletted	B-Data_Structure
or	O
integer	O
color	O
types	O
.	O
</s>
<s>
Sometimes	O
another	O
alpha	B-Algorithm
value	O
is	O
added	O
,	O
to	O
be	O
used	O
for	O
transparency	O
.	O
</s>
<s>
8	O
bits	O
per	O
pixel	B-Algorithm
–	O
Sometimes	O
palette	B-Data_Structure
mode	O
,	O
where	O
each	O
value	O
is	O
an	O
index	O
in	O
a	O
table	O
with	O
the	O
real	O
color	O
value	O
specified	O
in	O
one	O
of	O
the	O
other	O
formats	O
.	O
</s>
<s>
16	O
bits	O
per	O
pixel	B-Algorithm
–	O
Usually	O
the	O
bits	O
are	O
allocated	O
as	O
five	O
bits	O
for	O
red	O
,	O
six	O
bits	O
for	O
green	O
,	O
and	O
five	O
bits	O
for	O
blue	O
.	O
</s>
<s>
24	O
bits	O
per	O
pixel	B-Algorithm
–	O
There	O
are	O
eight	O
bits	O
for	O
each	O
of	O
red	O
,	O
green	O
,	O
and	O
blue	O
.	O
</s>
<s>
32	O
bits	O
per	O
pixel	B-Algorithm
–	O
There	O
are	O
eight	O
bits	O
for	O
each	O
of	O
red	O
,	O
green	O
,	O
blue	O
,	O
and	O
alpha	B-Algorithm
.	O
</s>
<s>
For	O
early	O
fixed-function	O
or	O
limited	O
programmability	O
graphics	O
(	O
i.e.	O
,	O
up	O
to	O
and	O
including	O
DirectX	O
8.1-compliant	O
GPUs	B-Architecture
)	O
this	O
was	O
sufficient	O
because	O
this	O
is	O
also	O
the	O
representation	O
used	O
in	O
displays	O
.	O
</s>
<s>
Given	O
sufficient	O
graphics	O
processing	O
power	O
even	O
graphics	O
programmers	O
would	O
like	O
to	O
use	O
better	O
formats	O
,	O
such	O
as	O
floating	B-Algorithm
point	I-Algorithm
data	O
formats	O
,	O
to	O
obtain	O
effects	O
such	O
as	O
high-dynamic-range	O
imaging	O
.	O
</s>
<s>
Many	O
GPGPU	B-Architecture
applications	O
require	O
floating	B-Algorithm
point	I-Algorithm
accuracy	O
,	O
which	O
came	O
with	O
video	B-Device
cards	I-Device
conforming	O
to	O
the	O
DirectX	B-Application
9	I-Application
specification	O
.	O
</s>
<s>
DirectX	B-Application
9	I-Application
Shader	O
Model	O
2.x	O
suggested	O
the	O
support	O
of	O
two	O
precision	O
types	O
:	O
full	O
and	O
partial	O
precision	O
.	O
</s>
<s>
Full	O
precision	O
support	O
could	O
either	O
be	O
FP32	O
or	O
FP24	O
(	O
floating	B-Algorithm
point	I-Algorithm
32	O
-	O
or	O
24-bit	O
per	O
component	O
)	O
or	O
greater	O
,	O
while	O
partial	O
precision	O
was	O
FP16	O
.	O
</s>
<s>
ATI	O
's	O
Radeon	B-Device
R300	I-Device
series	I-Device
of	O
GPUs	B-Architecture
supported	O
FP24	O
precision	O
only	O
in	O
the	O
programmable	O
fragment	O
pipeline	B-General_Concept
(	O
although	O
FP32	O
was	O
supported	O
in	O
the	O
vertex	O
processors	O
)	O
while	O
Nvidia	O
's	O
NV30	O
series	O
supported	O
both	O
FP16	O
and	O
FP32	O
;	O
other	O
vendors	O
such	O
as	O
S3	O
Graphics	O
and	O
XGI	O
supported	O
a	O
mixture	O
of	O
formats	O
up	O
to	O
FP24	O
.	O
</s>
<s>
The	O
implementations	O
of	O
floating	B-Algorithm
point	I-Algorithm
on	O
Nvidia	O
GPUs	B-Architecture
are	O
mostly	O
IEEE	O
compliant	O
;	O
however	O
,	O
this	O
is	O
not	O
true	O
across	O
all	O
vendors	O
.	O
</s>
<s>
While	O
64-bit	O
floating	B-Algorithm
point	I-Algorithm
values	I-Algorithm
(	O
double	O
precision	O
float	B-Algorithm
)	O
are	O
commonly	O
available	O
on	O
CPUs	O
,	O
these	O
are	O
not	O
universally	O
supported	O
on	O
GPUs	B-Architecture
.	O
</s>
<s>
Some	O
GPU	B-Architecture
architectures	O
sacrifice	O
IEEE	O
compliance	O
,	O
while	O
others	O
lack	O
double-precision	O
.	O
</s>
<s>
Efforts	O
have	O
occurred	O
to	O
emulate	O
double-precision	O
floating	B-Algorithm
point	I-Algorithm
values	I-Algorithm
on	O
GPUs	B-Architecture
;	O
however	O
,	O
the	O
speed	O
tradeoff	O
negates	O
any	O
benefit	O
to	O
offloading	O
the	O
computing	O
onto	O
the	O
GPU	B-Architecture
in	O
the	O
first	O
place	O
.	O
</s>
<s>
Most	O
operations	O
on	O
the	O
GPU	B-Architecture
operate	O
in	O
a	O
vectorized	O
fashion	O
:	O
one	O
operation	O
can	O
be	O
performed	O
on	O
up	O
to	O
four	O
values	O
at	O
once	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
one	O
color	O
is	O
to	O
be	O
modulated	O
by	O
another	O
color	O
,	O
the	O
GPU	B-Architecture
can	O
produce	O
the	O
resulting	O
color	O
in	O
one	O
operation	O
.	O
</s>
<s>
Many	O
other	O
applications	O
can	O
put	O
this	O
to	O
good	O
use	O
,	O
and	O
because	O
of	O
their	O
higher	O
performance	O
,	O
vector	O
instructions	O
,	O
termed	O
single	B-Device
instruction	I-Device
,	I-Device
multiple	I-Device
data	I-Device
(	O
SIMD	B-Device
)	O
,	O
have	O
long	O
been	O
available	O
on	O
CPUs	O
.	O
</s>
<s>
Originally	O
,	O
data	O
was	O
simply	O
passed	O
one-way	O
from	O
a	O
central	B-General_Concept
processing	I-General_Concept
unit	I-General_Concept
(	O
CPU	O
)	O
to	O
a	O
graphics	B-Architecture
processing	I-Architecture
unit	I-Architecture
(	O
GPU	B-Architecture
)	O
,	O
then	O
to	O
a	O
display	B-Device
device	I-Device
.	O
</s>
<s>
As	O
time	O
progressed	O
,	O
however	O
,	O
it	O
became	O
valuable	O
for	O
GPUs	B-Architecture
to	O
store	O
at	O
first	O
simple	O
,	O
then	O
complex	O
structures	O
of	O
data	O
to	O
be	O
passed	O
back	O
to	O
the	O
CPU	O
that	O
analyzed	O
an	O
image	O
,	O
or	O
a	O
set	O
of	O
scientific-data	O
represented	O
as	O
a	O
2D	O
or	O
3D	O
format	O
that	O
a	O
video	B-Device
card	I-Device
can	O
understand	O
.	O
</s>
<s>
Because	O
the	O
GPU	B-Architecture
has	O
access	O
to	O
every	O
draw	O
operation	O
,	O
it	O
can	O
analyze	O
data	O
in	O
these	O
forms	O
quickly	O
,	O
whereas	O
a	O
CPU	O
must	O
poll	O
every	O
pixel	B-Algorithm
or	O
data	O
element	O
much	O
more	O
slowly	O
,	O
as	O
the	O
speed	O
of	O
access	O
between	O
a	O
CPU	O
and	O
its	O
larger	O
pool	O
of	O
random-access	B-Architecture
memory	I-Architecture
(	O
or	O
in	O
an	O
even	O
worse	O
case	O
,	O
a	O
hard	B-Device
drive	I-Device
)	O
is	O
slower	O
than	O
GPUs	B-Architecture
and	O
video	B-Device
cards	I-Device
,	O
which	O
typically	O
contain	O
smaller	O
amounts	O
of	O
more	O
expensive	O
memory	O
that	O
is	O
much	O
faster	O
to	O
access	O
.	O
</s>
<s>
Transferring	O
the	O
portion	O
of	O
the	O
data	O
set	O
to	O
be	O
actively	O
analyzed	O
to	O
that	O
GPU	B-Architecture
memory	O
in	O
the	O
form	O
of	O
textures	O
or	O
other	O
easily	O
readable	O
GPU	B-Architecture
forms	O
results	O
in	O
speed	O
increase	O
.	O
</s>
<s>
The	O
distinguishing	O
feature	O
of	O
a	O
GPGPU	B-Architecture
design	O
is	O
the	O
ability	O
to	O
transfer	O
information	O
bidirectionally	O
back	O
from	O
the	O
GPU	B-Architecture
to	O
the	O
CPU	O
;	O
generally	O
the	O
data	O
throughput	O
in	O
both	O
directions	O
is	O
ideally	O
high	O
,	O
resulting	O
in	O
a	O
multiplier	O
effect	O
on	O
the	O
speed	O
of	O
a	O
specific	O
high-use	O
algorithm	O
.	O
</s>
<s>
GPGPU	B-Architecture
pipelines	O
may	O
improve	O
efficiency	O
on	O
especially	O
large	O
data	O
sets	O
and/or	O
data	O
containing	O
2D	O
or	O
3D	O
imagery	O
.	O
</s>
<s>
It	O
is	O
used	O
in	O
complex	O
graphics	B-General_Concept
pipelines	I-General_Concept
as	O
well	O
as	O
scientific	O
computing	O
;	O
more	O
so	O
in	O
fields	O
with	O
large	O
data	O
sets	O
like	O
genome	O
mapping	O
,	O
or	O
where	O
two	O
-	O
or	O
three-dimensional	O
analysis	O
is	O
useful	O
especially	O
at	O
present	O
biomolecule	O
analysis	O
,	O
protein	O
study	O
,	O
and	O
other	O
complex	O
organic	O
chemistry	O
.	O
</s>
<s>
Such	O
pipelines	O
can	O
also	O
vastly	O
improve	O
efficiency	O
in	O
image	B-Algorithm
processing	I-Algorithm
and	O
computer	B-Application
vision	I-Application
,	O
among	O
other	O
fields	O
;	O
as	O
well	O
as	O
parallel	B-Operating_System
processing	I-Operating_System
generally	O
.	O
</s>
<s>
Some	O
very	O
heavily	O
optimized	O
pipelines	O
have	O
yielded	O
speed	O
increases	O
of	O
several	O
hundred	O
times	O
the	O
original	O
CPU-based	O
pipeline	B-General_Concept
on	O
one	O
high-use	O
task	O
.	O
</s>
<s>
A	O
simple	O
example	O
would	O
be	O
a	O
GPU	B-Architecture
program	O
that	O
collects	O
data	O
about	O
average	O
lighting	O
values	O
as	O
it	O
renders	O
some	O
view	O
from	O
either	O
a	O
camera	O
or	O
a	O
computer	O
graphics	O
program	O
back	O
to	O
the	O
main	O
program	O
on	O
the	O
CPU	O
,	O
so	O
that	O
the	O
CPU	O
can	O
then	O
make	O
adjustments	O
to	O
the	O
overall	O
screen	O
view	O
.	O
</s>
<s>
A	O
more	O
advanced	O
example	O
might	O
use	O
edge	B-Algorithm
detection	I-Algorithm
to	O
return	O
both	O
numerical	O
information	O
and	O
a	O
processed	O
image	O
representing	O
outlines	O
to	O
a	O
computer	B-Application
vision	I-Application
program	O
controlling	O
,	O
say	O
,	O
a	O
mobile	O
robot	O
.	O
</s>
<s>
Because	O
the	O
GPU	B-Architecture
has	O
fast	O
and	O
local	O
hardware	O
access	O
to	O
every	O
pixel	B-Algorithm
or	O
other	O
picture	B-Algorithm
element	I-Algorithm
in	O
an	O
image	O
,	O
it	O
can	O
analyze	O
and	O
average	O
it	O
(	O
for	O
the	O
first	O
example	O
)	O
or	O
apply	O
a	O
Sobel	B-Algorithm
edge	I-Algorithm
filter	I-Algorithm
or	O
other	O
convolution	B-Language
filter	O
(	O
for	O
the	O
second	O
)	O
with	O
much	O
greater	O
speed	O
than	O
a	O
CPU	O
,	O
which	O
typically	O
must	O
access	O
slower	O
random-access	B-Architecture
memory	I-Architecture
copies	O
of	O
the	O
graphic	O
in	O
question	O
.	O
</s>
<s>
GPGPU	B-Architecture
is	O
fundamentally	O
a	O
software	O
concept	O
,	O
not	O
a	O
hardware	O
concept	O
;	O
it	O
is	O
a	O
type	O
of	O
algorithm	O
,	O
not	O
a	O
piece	O
of	O
equipment	O
.	O
</s>
<s>
Specialized	O
equipment	O
designs	O
may	O
,	O
however	O
,	O
even	O
further	O
enhance	O
the	O
efficiency	O
of	O
GPGPU	B-Architecture
pipelines	O
,	O
which	O
traditionally	O
perform	O
relatively	O
few	O
algorithms	O
on	O
very	O
large	O
amounts	O
of	O
data	O
.	O
</s>
<s>
Massively	O
parallelized	B-Operating_System
,	O
gigantic-data-level	O
tasks	O
thus	O
may	O
be	O
parallelized	B-Operating_System
even	O
further	O
via	O
specialized	O
setups	O
such	O
as	O
rack	O
computing	O
(	O
many	O
similar	O
,	O
highly	O
tailored	O
machines	O
built	O
into	O
a	O
rack	O
)	O
,	O
which	O
adds	O
a	O
third	O
layer	O
many	O
computing	O
units	O
each	O
using	O
many	O
CPUs	O
to	O
correspond	O
to	O
many	O
GPUs	B-Architecture
.	O
</s>
<s>
Some	O
Bitcoin	B-Protocol
"	O
miners	O
"	O
used	O
such	O
setups	O
for	O
high-quantity	O
processing	O
.	O
</s>
<s>
Historically	O
,	O
CPUs	O
have	O
used	O
hardware-managed	O
caches	B-General_Concept
,	O
but	O
the	O
earlier	O
GPUs	B-Architecture
only	O
provided	O
software-managed	O
local	O
memories	O
.	O
</s>
<s>
However	O
,	O
as	O
GPUs	B-Architecture
are	O
being	O
increasingly	O
used	O
for	O
general-purpose	O
applications	O
,	O
state-of-the-art	O
GPUs	B-Architecture
are	O
being	O
designed	O
with	O
hardware-managed	O
multi-level	O
caches	B-General_Concept
which	O
have	O
helped	O
the	O
GPUs	B-Architecture
to	O
move	O
towards	O
mainstream	O
computing	O
.	O
</s>
<s>
For	O
example	O
,	O
GeForce	O
200	O
series	O
GT200	O
architecture	O
GPUs	B-Architecture
did	O
not	O
feature	O
an	O
L2	O
cache	O
,	O
the	O
Fermi	B-General_Concept
GPU	B-Architecture
has	O
768KiB	O
last-level	O
cache	O
,	O
the	O
Kepler	B-General_Concept
GPU	B-Architecture
has	O
1.5MiB	O
last-level	O
cache	O
,	O
the	O
Maxwell	B-General_Concept
GPU	B-Architecture
has	O
2MiB	O
last-level	O
cache	O
,	O
and	O
the	O
Pascal	B-General_Concept
GPU	B-Architecture
has	O
4MiB	O
last-level	O
cache	O
.	O
</s>
<s>
GPUs	B-Architecture
have	O
very	O
large	O
register	B-General_Concept
files	I-General_Concept
,	O
which	O
allow	O
them	O
to	O
reduce	O
context-switching	O
latency	O
.	O
</s>
<s>
Register	B-General_Concept
file	I-General_Concept
size	O
is	O
also	O
increasing	O
over	O
different	O
GPU	B-Architecture
generations	O
,	O
e.g.	O
,	O
the	O
total	O
register	B-General_Concept
file	I-General_Concept
size	O
on	O
Maxwell	B-General_Concept
(	O
GM200	O
)	O
,	O
Pascal	B-General_Concept
and	O
Volta	O
GPUs	B-Architecture
are	O
6MiB	O
,	O
14MiB	O
and	O
20MiB	O
,	O
respectively	O
.	O
</s>
<s>
By	O
comparison	O
,	O
the	O
size	O
of	O
a	O
register	B-General_Concept
file	I-General_Concept
on	I-General_Concept
CPUs	I-General_Concept
is	O
small	O
,	O
typically	O
tens	O
or	O
hundreds	O
of	O
kilobytes	O
.	O
</s>
<s>
The	O
high	O
performance	O
of	O
GPUs	B-Architecture
comes	O
at	O
the	O
cost	O
of	O
high	O
power	O
consumption	O
,	O
which	O
under	O
full	O
load	O
is	O
in	O
fact	O
as	O
much	O
power	O
as	O
the	O
rest	O
of	O
the	O
PC	O
system	O
combined	O
.	O
</s>
<s>
The	O
maximum	O
power	O
consumption	O
of	O
the	O
Pascal	B-General_Concept
series	O
GPU	B-Architecture
(	O
Tesla	B-Device
P100	I-Device
)	O
was	O
specified	O
to	O
be	O
250W	O
.	O
</s>
<s>
GPUs	B-Architecture
are	O
designed	O
specifically	O
for	O
graphics	O
and	O
thus	O
are	O
very	O
restrictive	O
in	O
operations	O
and	O
programming	O
.	O
</s>
<s>
Due	O
to	O
their	O
design	O
,	O
GPUs	B-Architecture
are	O
only	O
effective	O
for	O
problems	O
that	O
can	O
be	O
solved	O
using	O
stream	B-Application
processing	I-Application
and	O
the	O
hardware	O
can	O
only	O
be	O
used	O
in	O
certain	O
ways	O
.	O
</s>
<s>
The	O
following	O
discussion	O
referring	O
to	O
vertices	O
,	O
fragments	O
and	O
textures	O
concerns	O
mainly	O
the	O
legacy	O
model	O
of	O
GPGPU	B-Architecture
programming	O
,	O
where	O
graphics	O
APIs	B-Application
(	O
OpenGL	B-Application
or	O
DirectX	B-Application
)	O
were	O
used	O
to	O
perform	O
general-purpose	O
computation	O
.	O
</s>
<s>
With	O
the	O
introduction	O
of	O
the	O
CUDA	B-Architecture
(	O
Nvidia	O
,	O
2007	O
)	O
and	O
OpenCL	B-Application
(	O
vendor-independent	O
,	O
2008	O
)	O
general-purpose	O
computing	O
APIs	B-Application
,	O
in	O
new	O
GPGPU	B-Architecture
codes	O
it	O
is	O
no	O
longer	O
necessary	O
to	O
map	O
the	O
computation	O
to	O
graphics	O
primitives	O
.	O
</s>
<s>
The	O
stream	B-Application
processing	I-Application
nature	O
of	O
GPUs	B-Architecture
remains	O
valid	O
regardless	O
of	O
the	O
APIs	B-Application
used	O
.	O
</s>
<s>
GPUs	B-Architecture
can	O
only	O
process	O
independent	O
vertices	O
and	O
fragments	O
,	O
but	O
can	O
process	O
many	O
of	O
them	O
in	O
parallel	O
.	O
</s>
<s>
In	O
this	O
sense	O
,	O
GPUs	B-Architecture
are	O
stream	O
processors	O
processors	O
that	O
can	O
operate	O
in	O
parallel	O
by	O
running	O
one	O
kernel	B-Algorithm
on	O
many	O
records	O
in	O
a	O
stream	O
at	O
once	O
.	O
</s>
<s>
Streams	O
provide	O
data	O
parallelism	B-Operating_System
.	O
</s>
<s>
Kernels	B-Algorithm
are	O
the	O
functions	O
that	O
are	O
applied	O
to	O
each	O
element	O
in	O
the	O
stream	O
.	O
</s>
<s>
In	O
the	O
GPUs	B-Architecture
,	O
vertices	O
and	O
fragments	O
are	O
the	O
elements	O
in	O
streams	O
and	O
vertex	O
and	O
fragment	O
shaders	O
are	O
the	O
kernels	B-Algorithm
to	O
be	O
run	O
on	O
them	O
.	O
</s>
<s>
It	O
is	O
important	O
for	O
GPGPU	B-Architecture
applications	O
to	O
have	O
high	O
arithmetic	O
intensity	O
else	O
the	O
memory	O
access	O
latency	O
will	O
limit	O
computational	O
speedup	B-Operating_System
.	O
</s>
<s>
Ideal	O
GPGPU	B-Architecture
applications	O
have	O
large	O
data	O
sets	O
,	O
high	O
parallelism	B-Operating_System
,	O
and	O
minimal	O
dependency	O
between	O
data	O
elements	O
.	O
</s>
<s>
There	O
are	O
a	O
variety	O
of	O
computational	O
resources	O
available	O
on	O
the	O
GPU	B-Architecture
:	O
</s>
<s>
The	O
most	O
common	O
form	O
for	O
a	O
stream	O
to	O
take	O
in	O
GPGPU	B-Architecture
is	O
a	O
2D	O
grid	O
because	O
this	O
fits	O
naturally	O
with	O
the	O
rendering	O
model	O
built	O
into	O
GPUs	B-Architecture
.	O
</s>
<s>
Many	O
computations	O
naturally	O
map	O
into	O
grids	O
:	O
matrix	O
algebra	O
,	O
image	B-Algorithm
processing	I-Algorithm
,	O
physically	O
based	O
simulation	O
,	O
and	O
so	O
on	O
.	O
</s>
<s>
Certain	O
operations	O
can	O
be	O
done	O
automatically	O
by	O
the	O
GPU	B-Architecture
because	O
of	O
this	O
.	O
</s>
<s>
Compute	B-Operating_System
kernels	I-Operating_System
can	O
be	O
thought	O
of	O
as	O
the	O
body	O
of	O
loops	O
.	O
</s>
<s>
On	O
the	O
GPU	B-Architecture
,	O
the	O
programmer	O
only	O
specifies	O
the	O
body	O
of	O
the	O
loop	O
as	O
the	O
kernel	B-Algorithm
and	O
what	O
data	O
to	O
loop	O
over	O
by	O
invoking	O
geometry	O
processing	O
.	O
</s>
<s>
Such	O
flow	O
control	O
structures	O
have	O
only	O
recently	O
been	O
added	O
to	O
GPUs	B-Architecture
.	O
</s>
<s>
Recent	O
GPUs	B-Architecture
allow	O
branching	O
,	O
but	O
usually	O
with	O
a	O
performance	O
penalty	O
.	O
</s>
<s>
Branching	O
should	O
generally	O
be	O
avoided	O
in	O
inner	O
loops	O
,	O
whether	O
in	O
CPU	O
or	O
GPU	B-Architecture
code	O
,	O
and	O
various	O
methods	O
,	O
such	O
as	O
static	O
branch	O
resolution	O
,	O
pre-computation	O
,	O
predication	O
,	O
loop	O
splitting	O
,	O
and	O
Z-cull	O
can	O
be	O
used	O
to	O
achieve	O
branching	O
when	O
hardware	O
support	O
does	O
not	O
exist	O
.	O
</s>
<s>
The	O
map	O
operation	O
simply	O
applies	O
the	O
given	O
function	O
(	O
the	O
kernel	B-Algorithm
)	O
to	O
every	O
element	O
in	O
the	O
stream	O
.	O
</s>
<s>
The	O
map	O
operation	O
is	O
simple	O
to	O
implement	O
on	O
the	O
GPU	B-Architecture
.	O
</s>
<s>
The	O
programmer	O
generates	O
a	O
fragment	O
for	O
each	O
pixel	B-Algorithm
on	O
screen	O
and	O
applies	O
a	O
fragment	O
program	O
to	O
each	O
one	O
.	O
</s>
<s>
While	O
at	O
first	O
glance	O
the	O
operation	O
may	O
seem	O
inherently	O
serial	O
,	O
efficient	O
parallel	O
scan	O
algorithms	O
are	O
possible	O
and	O
have	O
been	O
implemented	O
on	O
graphics	B-Architecture
processing	I-Architecture
units	I-Architecture
.	O
</s>
<s>
In	O
dedicated	O
compute	B-Operating_System
kernels	I-Operating_System
,	O
scatter	O
can	O
be	O
performed	O
by	O
indexed	O
writes	O
.	O
</s>
<s>
In	O
dedicated	O
compute	B-Operating_System
kernels	I-Operating_System
,	O
gather	O
may	O
be	O
performed	O
by	O
indexed	O
reads	O
.	O
</s>
<s>
The	O
most	O
common	O
implementation	O
on	O
GPUs	B-Architecture
is	O
using	O
radix	B-Algorithm
sort	I-Algorithm
for	O
integer	O
and	O
floating	B-Algorithm
point	I-Algorithm
data	O
and	O
coarse-grained	O
merge	B-Algorithm
sort	I-Algorithm
and	O
fine-grained	O
sorting	B-Algorithm
networks	I-Algorithm
for	O
general	O
comparable	O
data	O
.	O
</s>
<s>
The	O
GPU	B-Architecture
is	O
not	O
used	O
to	O
speed	B-Operating_System
up	I-Operating_System
the	O
search	O
for	O
an	O
individual	O
element	O
,	O
but	O
instead	O
is	O
used	O
to	O
run	O
multiple	O
searches	O
in	O
parallel	O
.	O
</s>
<s>
A	O
variety	O
of	O
data	O
structures	O
can	O
be	O
represented	O
on	O
the	O
GPU	B-Architecture
:	O
</s>
<s>
The	O
following	O
are	O
some	O
of	O
the	O
areas	O
where	O
GPUs	B-Architecture
have	O
been	O
used	O
for	O
general	O
purpose	O
computing	O
:	O
</s>
<s>
Implementations	O
of	O
:	O
the	O
GPU	B-Architecture
Tabu	O
Search	O
algorithm	O
solving	O
the	O
Resource	O
Constrained	O
Project	O
Scheduling	O
problem	O
is	O
freely	O
available	O
on	O
GitHub	O
;	O
the	O
GPU	B-Architecture
algorithm	O
solving	O
the	O
Nurse	B-Application
scheduling	I-Application
problem	I-Application
problem	O
is	O
freely	O
available	O
on	O
GitHub	O
.	O
</s>
<s>
GPGPU	B-Architecture
usage	O
in	O
Bioinformatics	O
:	O
</s>
<s>
Application	O
Description	O
Supported	O
features	O
Expected	O
speed-up	O
†	O
GPU‡	O
Multi-GPU	O
support	O
Release	O
status	O
AbaloneModels	O
molecular	O
dynamics	O
of	O
biopolymers	O
for	O
simulations	O
of	O
proteins	O
,	O
DNA	O
and	O
ligandsExplicit	O
and	O
implicit	O
solvent	O
,	O
hybrid	O
Monte	O
Carlo4	O
–	O
120xT	O
2075	O
,	O
2090	O
,	O
K10	O
,	O
K20	O
,	O
K20XAvailable	O
now	O
,	O
version	O
1.8.88	O
ACEMDGPU	O
simulation	O
of	O
molecular	O
mechanics	O
force	O
fields	O
,	O
implicit	O
and	O
explicit	O
solvent	O
Written	O
for	O
use	O
on	O
GPUs160	O
ns/day	O
GPU	B-Architecture
version	O
onlyT	O
2075	O
,	O
2090	O
,	O
K10	O
,	O
K20	O
,	O
K20XAvailable	O
now	O
AMBERSuite	O
of	O
programs	O
to	O
simulate	O
molecular	O
dynamics	O
on	O
biomoleculePMEMD	O
:	O
explicit	O
and	O
implicit	O
solvent	O
89.44	O
ns/day	O
JAC	O
NVE	O
T	O
2075	O
,	O
2090	O
,	O
K10	O
,	O
K20	O
,	O
K20XAvailable	O
now	O
,	O
version	O
12	O
+	O
bugfix9	O
DL-POLYSimulate	O
macromolecules	O
,	O
polymers	O
,	O
ionic	O
systems	O
,	O
etc	O
.	O
</s>
<s>
†	O
Expected	O
speedups	B-Operating_System
are	O
highly	O
dependent	O
on	O
system	O
configuration	O
.	O
</s>
<s>
GPU	B-Architecture
performance	O
compared	O
against	O
multi-core	B-Architecture
x86	B-Operating_System
CPU	O
socket	O
.	O
</s>
<s>
GPU	B-Architecture
performance	O
benchmarked	O
on	O
GPU	B-Architecture
supported	O
features	O
and	O
may	O
be	O
a	O
kernel	B-Algorithm
to	O
kernel	B-Algorithm
performance	O
comparison	O
.	O
</s>
<s>
Speedups	B-Operating_System
as	O
per	O
Nvidia	O
in-house	O
testing	O
or	O
ISV	O
's	O
documentation	O
.	O
</s>
<s>
‡	O
Q	O
=	O
Quadro	B-Application
GPU	I-Application
,	O
T	O
=	O
Tesla	B-Device
GPU	I-Device
.	O
</s>
<s>
Nvidia	O
recommended	O
GPUs	B-Architecture
for	O
this	O
application	O
.	O
</s>
