<s>
AMD	O
FireStream	O
was	O
AMD	O
's	O
brand	O
name	O
for	O
their	O
Radeon-based	O
product	O
line	O
targeting	O
stream	B-Application
processing	I-Application
and/or	O
GPGPU	B-Architecture
in	O
supercomputers	B-Architecture
.	O
</s>
<s>
Originally	O
developed	O
by	O
ATI	O
Technologies	O
around	O
the	O
Radeon	B-Device
X1900	I-Device
XTX	I-Device
in	O
2006	O
,	O
the	O
product	O
line	O
was	O
previously	O
branded	O
as	O
both	O
ATI	O
FireSTREAM	O
and	O
AMD	O
Stream	O
Processor	O
.	O
</s>
<s>
The	O
AMD	O
FireStream	O
can	O
also	O
be	O
used	O
as	O
a	O
floating-point	B-Algorithm
co-processor	B-General_Concept
for	O
offloading	O
CPU	O
calculations	O
,	O
which	O
is	O
part	O
of	O
the	O
Torrenza	B-General_Concept
initiative	O
.	O
</s>
<s>
The	O
FireStream	O
line	O
has	O
been	O
discontinued	O
since	O
2012	O
,	O
when	O
GPGPU	B-Architecture
workloads	O
were	O
entirely	O
folded	O
into	O
the	O
AMD	B-Application
FirePro	I-Application
line	O
.	O
</s>
<s>
The	O
FireStream	O
line	O
is	O
a	O
series	O
of	O
add-on	O
expansion	B-Device
cards	I-Device
released	O
from	O
2006	O
to	O
2010	O
,	O
based	O
on	O
standard	O
Radeon	B-Device
GPUs	O
but	O
designed	O
to	O
serve	O
as	O
a	O
general-purpose	O
co-processor	B-General_Concept
,	O
rather	O
than	O
rendering	O
and	O
outputting	O
3D	O
graphics	O
.	O
</s>
<s>
Like	O
the	O
FireGL/FirePro	B-Application
line	O
,	O
they	O
were	O
given	O
more	O
memory	O
and	O
memory	O
bandwidth	O
,	O
but	O
the	O
FireStream	O
cards	O
do	O
not	O
necessarily	O
have	O
video	O
output	O
ports	O
.	O
</s>
<s>
All	O
support	O
32-bit	O
single-precision	O
floating	B-Algorithm
point	I-Algorithm
,	O
and	O
all	O
but	O
the	O
first	O
release	O
support	O
64-bit	O
double-precision	O
.	O
</s>
<s>
The	O
line	O
was	O
partnered	O
with	O
new	O
APIs	O
to	O
provide	O
higher	O
performance	O
than	O
existing	O
OpenGL	B-Application
and	O
Direct3D	B-Application
shader	O
APIs	O
could	O
provide	O
,	O
beginning	O
with	O
Close	O
to	O
Metal	O
,	O
followed	O
by	O
OpenCL	B-Application
and	O
the	O
Stream	B-Application
Computing	I-Application
SDK	B-Application
,	O
and	O
eventually	O
integrated	O
into	O
the	O
APP	B-Application
SDK	I-Application
.	O
</s>
<s>
For	O
highly	O
parallel	O
floating	B-Algorithm
point	I-Algorithm
math	I-Algorithm
workloads	O
,	O
the	O
cards	O
can	O
speed	O
up	O
large	O
computations	O
by	O
more	O
than	O
10	O
times	O
;	O
Folding	B-Application
@Home	I-Application
,	O
the	O
earliest	O
and	O
one	O
of	O
the	O
most	O
visible	O
users	O
of	O
the	O
GPGPU	B-Architecture
,	O
obtained	O
20-40	O
times	O
the	O
CPU	O
performance	O
.	O
</s>
<s>
Each	O
pixel	O
and	O
vertex	O
shader	O
,	O
or	O
unified	O
shader	O
in	O
later	O
models	O
,	O
can	O
perform	O
arbitrary	O
floating-point	B-Algorithm
calculations	O
.	O
</s>
<s>
Following	O
the	O
release	O
of	O
the	O
Radeon	B-Device
R520	I-Device
and	O
GeForce	O
G70	O
GPU	O
cores	O
with	O
programmable	O
shaders	O
,	O
the	O
large	O
floating-point	B-Algorithm
throughput	O
drew	O
attention	O
from	O
academic	O
and	O
commercial	O
groups	O
,	O
experimenting	O
with	O
using	O
then	O
for	O
non-graphics	O
work	O
.	O
</s>
<s>
The	O
interest	O
led	O
ATI	O
(	O
and	O
Nvidia	O
)	O
to	O
create	O
GPGPU	B-Architecture
products	O
—	O
able	O
to	O
calculate	O
general	O
purpose	O
mathematical	O
formulas	O
in	O
a	O
massively	O
parallel	O
way	O
—	O
to	O
process	O
heavy	O
calculations	O
traditionally	O
done	O
on	O
CPUs	B-General_Concept
and	O
specialized	O
floating-point	B-Algorithm
math	I-Algorithm
co-processors	B-General_Concept
.	O
</s>
<s>
GPGPUs	B-Architecture
were	O
projected	O
to	O
have	O
immediate	O
performance	O
gains	O
of	O
a	O
factor	O
of	O
10	O
or	O
more	O
,	O
over	O
compared	O
to	O
contemporary	O
multi-socket	O
CPU-only	O
calculation	O
.	O
</s>
<s>
With	O
the	O
development	O
of	O
the	O
high-performance	O
X1900	B-Device
XFX	O
nearly	O
finished	O
,	O
ATI	O
based	O
its	O
first	O
Stream	O
Processor	O
design	O
on	O
it	O
,	O
announcing	O
it	O
as	O
the	O
upcoming	O
ATI	O
FireSTREAM	O
together	O
with	O
the	O
new	O
Close	O
to	O
Metal	O
API	O
at	O
SIGGRAPH	O
2006	O
.	O
</s>
<s>
The	O
core	O
itself	O
was	O
mostly	O
unchanged	O
,	O
except	O
for	O
doubling	O
the	O
onboard	O
memory	O
and	O
bandwidth	O
,	O
similar	O
to	O
the	O
FireGL	B-Application
V7350	I-Application
;	O
new	O
driver	O
and	O
software	O
support	O
made	O
up	O
most	O
of	O
the	O
difference	O
.	O
</s>
<s>
Folding	B-Application
@home	I-Application
began	O
using	O
the	O
X1900	B-Device
for	O
general	O
computation	O
,	O
using	O
a	O
pre-release	O
of	O
version	O
6.5	O
of	O
the	O
ATI	O
Catalyst	O
driver	O
,	O
and	O
reported	O
20-40x	O
improvement	O
in	O
GPU	O
over	O
CPU	O
.	O
</s>
<s>
Asynchronous	O
DMA	B-General_Concept
also	O
improved	O
performance	O
by	O
allowing	O
a	O
larger	O
memory	O
pool	O
without	O
the	O
CPU	O
's	O
help	O
.	O
</s>
<s>
Plans	O
included	O
the	O
development	O
of	O
a	O
stream	O
processor	O
on	O
an	O
MXM	B-Architecture
module	I-Architecture
by	O
2008	O
,	O
for	O
laptop	O
computing	O
,	O
but	O
was	O
never	O
released	O
.	O
</s>
<s>
This	O
generation	O
again	O
doubled	O
the	O
performance	O
relative	O
to	O
the	O
previous	O
,	O
to	O
2	O
teraflops	O
in	O
the	O
9350	O
and	O
2.6	O
teraflops	O
in	O
the	O
9370	O
,	O
and	O
was	O
the	O
first	O
built	O
from	O
the	O
ground	O
up	O
for	O
OpenCL	B-Application
.	O
</s>
<s>
The	O
Northern	O
and	O
Southern	O
Islands	O
generations	O
were	O
skipped	O
,	O
and	O
in	O
2012	O
,	O
AMD	O
announced	O
that	O
the	O
new	O
FirePro	O
W	O
(	O
workstation	O
)	O
and	O
S	O
(	O
server	O
)	O
series	O
based	O
on	O
the	O
new	O
Graphics	B-Architecture
Core	I-Architecture
Next	I-Architecture
architecture	O
would	O
take	O
the	O
place	O
of	O
FireStream	O
cards	O
.	O
</s>
<s>
One	O
of	O
the	O
supporting	O
firms	O
was	O
PeakStream	O
(	O
acquired	O
by	O
Google	B-Application
in	O
June	O
2007	O
)	O
,	O
who	O
was	O
first	O
to	O
provide	O
an	O
open	O
beta	O
version	O
of	O
software	O
to	O
support	O
CTM	O
and	O
AMD	O
FireStream	O
as	O
well	O
as	O
x86	B-Operating_System
and	O
Cell	O
(	O
Cell	O
Broadband	O
Engine	O
)	O
processors	O
.	O
</s>
<s>
The	O
FireStream	O
was	O
claimed	O
to	O
be	O
20	O
times	O
faster	O
in	O
typical	O
applications	O
than	O
regular	O
CPUs	B-General_Concept
after	O
running	O
PeakStream	O
's	O
software	O
.	O
</s>
<s>
RapidMind	O
also	O
provided	O
stream	B-Application
processing	I-Application
software	O
that	O
worked	O
with	O
ATI	O
and	O
NVIDIA	O
,	O
as	O
well	O
as	O
Cell	O
processors	O
.	O
</s>
<s>
After	O
abandoning	O
their	O
short-lived	O
Close	O
to	O
Metal	O
API	O
,	O
AMD	O
focused	O
on	O
OpenCL	B-Application
.	O
</s>
<s>
AMD	O
first	O
released	O
its	O
Stream	B-Application
Computing	I-Application
SDK	B-Application
(	O
v1.0	O
)	O
,	O
in	O
December	O
2007	O
under	O
the	O
AMD	O
EULA	O
,	O
to	O
be	O
run	O
on	O
Windows	B-Application
XP	I-Application
.	O
</s>
<s>
The	O
SDK	B-Application
includes	O
"	O
Brook+	O
"	O
,	O
an	O
AMD	O
hardware	O
optimized	O
version	O
of	O
the	O
Brook	B-Library
language	O
developed	O
by	O
Stanford	O
University	O
,	O
itself	O
a	O
variant	O
of	O
the	O
ANSI	O
C	B-Language
(	O
C	B-Language
language	I-Language
)	O
,	O
open-sourced	B-Application
and	O
optimized	O
for	O
stream	B-Application
computing	I-Application
.	O
</s>
<s>
The	O
AMD	B-Library
Core	I-Library
Math	I-Library
Library	I-Library
(	O
ACML	B-Library
)	O
and	O
AMD	B-Library
Performance	I-Library
Library	I-Library
(	O
APL	O
)	O
with	O
optimizations	O
for	O
the	O
AMD	O
FireStream	O
and	O
the	O
COBRA	O
video	O
library	O
(	O
further	O
renamed	O
as	O
"	O
Accelerated	O
Video	B-Device
Transcoding	I-Device
"	O
or	O
AVT	O
)	O
for	O
video	B-Device
transcoding	I-Device
acceleration	O
will	O
also	O
be	O
included	O
.	O
</s>
<s>
Another	O
important	O
part	O
of	O
the	O
SDK	B-Application
,	O
the	O
Compute	O
Abstraction	O
Layer	O
(	O
CAL	O
)	O
,	O
is	O
a	O
software	O
development	O
layer	O
aimed	O
for	O
low-level	O
access	O
,	O
through	O
the	O
CTM	O
hardware	O
interface	O
,	O
to	O
the	O
GPU	O
architecture	O
for	O
performance	O
tuning	O
software	O
written	O
in	O
various	O
high-level	O
programming	O
languages	O
.	O
</s>
<s>
In	O
August	O
2011	O
,	O
AMD	O
released	O
version	O
2.5	O
of	O
the	O
ATI	O
APP	O
Software	B-Application
Development	I-Application
Kit	I-Application
,	O
which	O
includes	O
support	O
for	O
OpenCL	B-Application
1.1	I-Application
,	O
a	O
parallel	B-Operating_System
computing	I-Operating_System
language	O
developed	O
by	O
the	B-Library
Khronos	I-Library
Group	I-Library
.	O
</s>
<s>
The	O
concept	O
of	O
compute	B-Operating_System
shaders	I-Operating_System
,	O
officially	O
called	O
DirectCompute	O
,	O
in	O
Microsoft	O
's	O
next	O
generation	O
API	O
called	O
DirectX	O
11	O
is	O
already	O
included	O
in	O
graphics	O
drivers	O
with	O
DirectX	O
11	O
support	O
.	O
</s>
<s>
According	O
to	O
an	O
AMD-demonstrated	O
system	O
with	O
two	O
dual-core	O
AMD	B-General_Concept
Opteron	I-General_Concept
processors	O
and	O
two	O
Radeon	B-Device
R600	O
GPU	O
cores	O
running	O
on	O
Microsoft	B-Application
Windows	I-Application
XP	I-Application
Professional	O
,	O
1teraflop	O
(	O
TFLOP	O
)	O
can	O
be	O
achieved	O
by	O
a	O
universal	O
multiply-add	O
(	O
MADD	O
)	O
calculation	O
.	O
</s>
<s>
In	O
a	O
demonstration	O
of	O
Kaspersky	O
SafeStream	O
anti-virus	O
scanning	O
that	O
had	O
been	O
optimized	O
for	O
AMD	O
stream	O
processors	O
,	O
was	O
able	O
to	O
scan	O
21	O
times	O
faster	O
with	O
the	O
R670	O
based	O
acceleration	O
than	O
with	O
search	O
running	O
entirely	O
on	O
an	O
Opteron	B-General_Concept
,	O
in	O
2007	O
.	O
</s>
<s>
Recursive	O
functions	O
are	O
not	O
supported	O
in	O
Brook+	O
because	O
all	O
function	O
calls	O
are	O
inlined	O
at	O
compile	O
time	O
.	O
</s>
<s>
Conversion	O
of	O
floating-point	B-Algorithm
numbers	I-Algorithm
to	O
integers	O
on	O
GPUs	O
is	O
done	O
differently	O
than	O
on	O
x86	B-Operating_System
CPUs	B-General_Concept
;	O
it	O
is	O
not	O
fully	O
IEEE-754	O
compliant	O
.	O
</s>
<s>
Doing	O
"	O
global	O
synchronization	O
"	O
on	O
the	O
GPU	O
is	O
not	O
very	O
efficient	O
,	O
which	O
forces	O
the	O
GPU	O
to	O
divide	O
the	O
kernel	B-Operating_System
and	O
do	O
synchronization	O
on	O
the	O
CPU	O
.	O
</s>
