<s>
TeraScale	B-Architecture
is	O
the	O
codename	O
for	O
a	O
family	O
of	O
graphics	B-Architecture
processing	I-Architecture
unit	I-Architecture
microarchitectures	B-General_Concept
developed	O
by	O
ATI	O
Technologies/AMD	O
and	O
their	O
second	O
microarchitecture	B-General_Concept
implementing	O
the	O
unified	O
shader	B-Language
model	I-Language
following	O
Xenos	B-Device
.	O
</s>
<s>
TeraScale	B-Architecture
replaced	O
the	O
old	O
fixed-pipeline	B-General_Concept
microarchitectures	B-General_Concept
and	O
competed	O
directly	O
with	O
Nvidia	O
's	O
first	O
unified	O
shader	O
microarchitecture	B-General_Concept
named	O
Tesla	B-Operating_System
.	O
</s>
<s>
TeraScale	B-Architecture
was	O
used	O
in	O
HD	O
2000	O
manufactured	O
in	O
80nm	O
and	O
65	O
nm	O
,	O
HD	O
3000	O
manufactured	O
in	O
65nm	B-Algorithm
and	O
55nm	O
,	O
HD	B-Device
4000	I-Device
manufactured	O
in	O
55nm	O
and	O
40nm	O
,	O
HD	O
5000	O
and	O
HD	O
6000	O
manufactured	O
in	O
40nm	O
.	O
</s>
<s>
TeraScale	B-Architecture
was	O
also	O
used	O
in	O
the	O
AMD	B-Architecture
Accelerated	I-Architecture
Processing	I-Architecture
Units	I-Architecture
code-named	O
"	O
Brazos	O
"	O
,	O
"	O
Llano	O
"	O
,	O
"	O
Trinity	O
"	O
and	O
"	O
Richland	O
"	O
.	O
</s>
<s>
TeraScale	B-Architecture
is	O
even	O
found	O
in	O
some	O
of	O
the	O
succeeding	O
graphics	O
cards	O
brands	O
.	O
</s>
<s>
TeraScale	B-Architecture
is	O
a	O
VLIW	B-General_Concept
SIMD	B-Device
architecture	O
,	O
while	O
Tesla	B-Operating_System
is	O
a	O
RISC	B-Architecture
SIMD	B-Device
architecture	O
,	O
similar	O
to	O
TeraScale	B-Architecture
's	O
successor	O
Graphics	B-Architecture
Core	I-Architecture
Next	I-Architecture
.	O
</s>
<s>
TeraScale	B-Architecture
implements	O
HyperZ	B-Device
.	O
</s>
<s>
a	O
compiler	O
back-end	O
)	O
is	O
available	O
for	O
TeraScale	B-Architecture
,	O
but	O
it	O
seems	O
to	O
be	O
missing	O
in	O
LLVM	O
's	O
matrix	O
.	O
</s>
<s>
Mesa	B-Application
3D	I-Application
makes	O
use	O
of	O
it	O
.	O
</s>
<s>
At	O
SIGGRAPH	O
08	O
in	O
December	O
2008	O
AMD	O
employee	O
Mike	O
Houston	O
described	O
some	O
of	O
the	O
TeraScale	B-Architecture
microarchitecture	B-General_Concept
.	O
</s>
<s>
Previous	O
GPU	B-Architecture
architectures	O
implemented	O
fixed-pipelines	O
,	O
i.e.	O
</s>
<s>
TeraScale	B-Architecture
leverages	O
many	O
flexible	O
shader	O
processors	O
which	O
can	O
be	O
scheduled	O
to	O
process	O
a	O
variety	O
of	O
shader	O
types	O
,	O
thereby	O
significantly	O
increasing	O
GPU	B-Architecture
throughput	O
(	O
dependent	O
on	O
application	O
instruction	O
mix	O
as	O
noted	O
below	O
)	O
.	O
</s>
<s>
The	O
R600	O
core	O
processes	O
vertex	O
,	O
geometry	O
,	O
and	O
pixel	O
shaders	O
as	O
outlined	O
by	O
the	O
Direct3D	B-Application
10.0	O
specification	O
for	O
Shader	B-Language
Model	I-Language
4.0	I-Language
in	O
addition	O
to	O
full	O
OpenGL	B-Application
3.0	O
support	O
.	O
</s>
<s>
The	O
new	O
unified	O
shader	O
functionality	O
is	O
based	O
upon	O
a	O
very	B-General_Concept
long	I-General_Concept
instruction	I-General_Concept
word	I-General_Concept
(	O
VLIW	B-General_Concept
)	O
architecture	O
in	O
which	O
the	O
core	O
executes	O
operations	O
in	O
parallel	O
.	O
</s>
<s>
Notably	O
,	O
the	O
VLIW	B-General_Concept
architecture	O
brings	O
with	O
it	O
some	O
classic	O
challenges	O
inherent	O
to	O
VLIW	B-General_Concept
designs	O
,	O
namely	O
that	O
of	O
maintaining	O
optimal	O
instruction	O
flow	O
.	O
</s>
<s>
Performance	O
of	O
the	O
GPU	B-Architecture
is	O
highly	O
dependent	O
on	O
the	O
mixture	O
of	O
instructions	O
being	O
used	O
by	O
the	O
application	O
and	O
how	O
well	O
the	O
real-time	O
compiler	O
in	O
the	O
driver	O
can	O
organize	O
said	O
instructions	O
.	O
</s>
<s>
TeraScale	B-Architecture
includes	O
multiple	O
units	O
capable	O
of	O
carrying	O
out	O
tessellation	B-Algorithm
.	O
</s>
<s>
Those	O
are	O
similar	O
to	O
the	O
programmable	O
units	O
of	O
the	O
Xenos	B-Device
GPU	I-Device
which	O
is	O
used	O
in	O
the	O
Xbox	O
360	O
.	O
</s>
<s>
Tessellation	B-Algorithm
was	O
officially	O
specified	O
in	O
the	O
major	O
API	O
's	O
starting	O
with	O
DirectX	B-Application
11	O
and	O
OpenGL	B-Application
4	O
.	O
</s>
<s>
TeraScale	B-Architecture
1	O
based	O
GPU	B-Architecture
's	O
(	O
HD	O
2000	O
,	O
3000	O
and	O
4000	O
series	O
)	O
are	O
only	O
conformant	O
to	O
Direct3D	B-Application
10	O
and	O
OpenGL	B-Application
3.3	O
and	O
implements	O
therefore	O
a	O
different	O
tessellation	B-Algorithm
principle	O
which	O
uses	O
vendor	O
specific	O
API	O
extensions	O
.	O
</s>
<s>
The	O
TeraScale	B-Architecture
2	O
based	O
GPU	B-Architecture
's	O
(	O
starting	O
with	O
the	O
Radeon	O
HD	O
5000	O
series	O
)	O
were	O
the	O
first	O
to	O
conform	O
with	O
both	O
Direct3D	B-Application
11	O
and	O
OpenGL	B-Application
4.0	O
tesselation	O
technique	O
.	O
</s>
<s>
Although	O
the	O
TeraScale	B-Architecture
1	O
tessellator	O
is	O
simpler	O
in	O
design	O
,	O
it	O
is	O
described	O
by	O
AMD	O
as	O
a	O
subset	O
of	O
the	O
later	O
tesselation	O
standard	O
.	O
</s>
<s>
The	O
TeraScale	B-Architecture
tessellator	O
units	O
allow	O
the	O
developers	O
to	O
take	O
a	O
simple	O
polygon	O
mesh	O
and	O
subdivide	O
it	O
using	O
a	O
curved	O
surface	O
evaluation	O
function	O
.	O
</s>
<s>
There	O
are	O
different	O
tessellation	B-Algorithm
forms	O
,	O
such	O
as	O
Bézier	O
surfaces	O
with	O
N-patches	O
,	O
B-splines	B-Algorithm
and	O
NURBS	B-Algorithm
,	O
and	O
also	O
some	O
subdivision	O
techniques	O
of	O
the	O
surface	O
,	O
which	O
usually	O
includes	O
displacement	O
map	O
some	O
kind	O
of	O
a	O
texture	O
.	O
</s>
<s>
The	O
TeraScale	B-Architecture
tessellator	O
is	O
reminiscent	O
of	O
ATI	O
TruForm	O
,	O
the	O
brand	O
name	O
of	O
an	O
early	O
hardware	O
tessellation	B-Algorithm
unit	O
used	O
initially	O
in	O
the	O
Radeon	B-Device
8500	I-Device
.	O
</s>
<s>
A	O
few	O
games	O
(	O
such	O
as	O
Madden	B-Application
NFL	I-Application
2004	I-Application
,	O
Serious	B-Device
Sam	I-Device
,	O
Unreal	B-Application
Tournament	I-Application
2003	I-Application
and	O
2004	B-Application
,	O
and	O
unofficially	O
Morrowind	B-Application
)	O
,	O
had	O
the	O
support	O
for	O
the	O
ATI	O
's	O
tesselation	O
technology	O
included	O
.	O
</s>
<s>
Such	O
a	O
slow	O
adaptation	O
has	O
to	O
do	O
with	O
the	O
fact	O
that	O
it	O
was	O
not	O
a	O
feature	O
shared	O
with	O
NVIDIA	O
GPUs	B-Architecture
,	O
since	O
those	O
had	O
implemented	O
a	O
competing	O
tessellation	B-Algorithm
solution	O
using	O
Quintic-RT	O
patches	O
which	O
had	O
achieved	O
even	O
less	O
support	O
from	O
the	O
major	O
game	O
developers	O
.	O
</s>
<s>
Since	O
the	O
Xbox	O
360	O
's	O
GPU	B-Architecture
is	O
based	O
on	O
the	O
ATI	O
's	O
architecture	O
,	O
Microsoft	O
saw	O
the	O
hardware-accelerated	O
surface	O
tessellation	B-Algorithm
as	O
a	O
major	O
GPU	B-Architecture
feature	O
.	O
</s>
<s>
A	O
couple	O
of	O
years	O
later	O
the	O
tesselation	O
feature	O
became	O
mandatory	O
with	O
the	O
release	O
of	O
the	O
DirectX	B-Application
11	O
in	O
2009	O
.	O
</s>
<s>
While	O
the	O
tessellation	B-Algorithm
principle	O
introduced	O
with	O
TeraScale	B-Architecture
was	O
not	O
part	O
of	O
the	O
OpenGL	B-Application
3.3	O
or	O
Direct3D	B-Application
10.0	O
requirements	O
,	O
and	O
competitors	O
such	O
as	O
the	O
GeForce	O
8	O
series	O
lacked	O
similar	O
hardware	O
,	O
Microsoft	O
has	O
added	O
the	O
tessellation	B-Algorithm
feature	O
as	O
part	O
of	O
their	O
DirectX	B-Application
10.1	I-Application
future	O
plans	O
.	O
</s>
<s>
Finally	O
,	O
Microsoft	O
introduced	O
tessellation	B-Algorithm
as	O
a	O
required	O
capability	O
not	O
with	O
DirectX	B-Application
10.1	I-Application
but	O
DirectX	B-Application
11	O
.	O
</s>
<s>
GCN	O
geometric	O
processor	O
is	O
AMD	O
's	O
(	O
which	O
acquired	O
the	O
ATI	O
's	O
GPU	B-Architecture
business	O
)	O
most	O
current	O
solution	O
for	O
carrying	O
out	O
tessellation	B-Algorithm
using	O
the	O
GPU	B-Architecture
.	O
</s>
<s>
Although	O
the	O
R600	O
is	O
a	O
significant	O
departure	O
from	O
previous	O
designs	O
,	O
it	O
still	O
shares	O
many	O
features	O
with	O
its	O
predecessor	O
,	O
the	O
Radeon	B-Device
R520	I-Device
.	O
</s>
<s>
The	O
Ultra-Threaded	O
Dispatch	O
Processor	O
is	O
a	O
major	O
architectural	O
component	O
of	O
the	O
R600	O
core	O
,	O
just	O
as	O
it	O
was	O
with	O
the	O
Radeon	B-Device
X1000	I-Device
GPUs	B-Architecture
.	O
</s>
<s>
There	O
are	O
lower	O
levels	O
of	O
management	O
as	O
well	O
;	O
each	O
SIMD	B-Device
array	O
of	O
80	O
stream	O
processors	O
has	O
its	O
own	O
sequencer	O
and	O
arbiter	O
.	O
</s>
<s>
Texturing	O
and	O
final	O
output	O
aboard	O
the	O
R600	O
core	O
is	O
similar	O
but	O
also	O
distinct	O
from	O
R580	B-Device
.	O
</s>
<s>
R600	O
is	O
equipped	O
with	O
4	O
texture	O
units	O
that	O
are	O
decoupled	O
(	O
independent	O
)	O
from	O
the	O
shader	O
core	O
,	O
like	O
in	O
the	O
R520	B-Device
and	O
R580	B-Device
GPUs	B-Architecture
.	O
</s>
<s>
The	O
render	B-General_Concept
output	I-General_Concept
units	I-General_Concept
(	O
ROPs	B-General_Concept
)	O
of	O
Radeon	O
HD	O
2000	O
series	O
now	O
performs	O
the	O
task	O
of	O
Multisample	B-Algorithm
anti-aliasing	I-Algorithm
(	O
MSAA	O
)	O
with	O
programmable	O
sample	O
grids	O
and	O
maximum	O
of	O
8	O
sample	O
points	O
,	O
instead	O
of	O
using	O
pixel	O
shaders	O
as	O
in	O
the	O
Radeon	B-Device
X1000	I-Device
series	I-Device
.	O
</s>
<s>
Also	O
new	O
is	O
the	O
capability	O
to	O
filter	O
FP16	O
textures	O
,	O
popular	O
with	O
HDR	B-Device
lighting	I-Device
,	O
at	O
full-speed	O
.	O
</s>
<s>
Anti-aliasing	O
capabilities	O
are	O
more	O
robust	O
on	O
R600	O
than	O
on	O
the	O
R520	B-Device
series	O
.	O
</s>
<s>
In	O
addition	O
to	O
the	O
ability	O
to	O
perform	O
8×	O
MSAA	O
,	O
up	O
from	O
6×	O
MSAA	O
on	O
the	O
R300	O
through	O
R580	B-Device
,	O
R600	O
has	O
a	O
new	O
custom	O
filter	O
anti-aliasing	O
(	O
CFAA	O
)	O
mode	O
.	O
</s>
<s>
CFAA	O
is	O
performed	O
by	O
shader	O
,	O
instead	O
of	O
in	O
the	O
ROPs	B-General_Concept
.	O
</s>
<s>
Memory	B-General_Concept
controllers	I-General_Concept
are	O
connected	O
via	O
internal	O
bi-directional	O
ring	O
bus	O
wrapped	O
around	O
the	O
processor	O
.	O
</s>
<s>
All	O
variants	O
support	O
PCI	O
Express	O
2.0	O
,	O
DirectX	B-Application
10.1	I-Application
with	O
Shader	O
Model	O
4.1	O
features	O
,	O
dedicated	O
ATI	O
Unified	O
Video	O
Decoder	O
(	O
UVD	O
)	O
for	O
all	O
models	O
and	O
PowerPlay	O
technology	O
for	O
desktop	O
video	O
cards	O
.	O
</s>
<s>
TeraScale	B-Architecture
2	O
(	O
VLIW5	O
)	O
was	O
introduced	O
with	O
Radeon	O
HD	O
5000	O
Series	O
GPUs	B-Architecture
in	O
"	O
Evergreen	O
"	O
generation	O
.	O
</s>
<s>
5870	O
(	O
Cypress	O
)	O
,	O
5770	O
(	O
Juniper	O
)	O
and	O
5670	O
(	O
Redwood	O
)	O
support	O
max	O
resolution	O
of	O
the	O
6	O
times	O
2560×1600	O
pixels	O
,	O
while	O
the	O
5470	O
(	O
Cedar	O
)	O
supports	O
4	O
times	O
2560×1600	O
pixels	O
,	O
important	O
for	O
AMD	O
Eyefinity	O
multi-monitor	B-General_Concept
support	O
.	O
</s>
<s>
With	O
the	O
release	O
of	O
Cypress	O
,	O
the	O
Terascale	B-Architecture
graphics	O
engine	O
architecture	O
has	O
been	O
upgraded	O
with	O
twice	O
the	O
number	O
of	O
stream	O
cores	O
,	O
texture	O
units	O
and	O
ROP	O
units	O
compared	O
to	O
the	O
RV770	O
.	O
</s>
<s>
The	O
architecture	O
of	O
stream	O
cores	O
is	O
largely	O
unchanged	O
,	O
but	O
adds	O
support	O
for	O
DirectX	B-Application
11/DirectCompute	O
11	O
capabilities	O
with	O
new	O
instructions	O
.	O
</s>
<s>
This	O
combination	O
of	O
is	O
referred	O
to	O
as	O
a	O
SIMD	B-Device
core	O
.	O
</s>
<s>
Unlike	O
the	O
predecessor	O
Radeon	B-Device
R700	I-Device
,	O
as	O
DirectX	B-Application
11	O
mandates	O
full	O
developer	O
control	O
over	O
interpolation	O
,	O
dedicated	O
interpolators	O
were	O
removed	O
,	O
relying	O
instead	O
on	O
the	O
SIMD	B-Device
cores	O
.	O
</s>
<s>
The	O
stream	O
cores	O
can	O
handle	O
the	O
higher	O
rounding	O
precision	O
fused	O
multiply	B-Algorithm
–	I-Algorithm
add	I-Algorithm
(	O
FMA	O
)	O
instruction	O
in	O
both	O
single	O
and	O
double	O
precision	O
which	O
increases	O
precision	O
over	O
multiply	B-Algorithm
–	I-Algorithm
add	I-Algorithm
(	O
MAD	O
)	O
and	O
is	O
compliant	O
to	O
IEEE	O
754-2008	O
standard	O
.	O
</s>
<s>
The	O
instruction	O
sum	B-Algorithm
of	I-Algorithm
absolute	I-Algorithm
differences	I-Algorithm
(	O
SAD	O
)	O
has	O
been	O
natively	O
added	O
to	O
the	O
processors	O
.	O
</s>
<s>
Each	O
SIMD	B-Device
core	O
is	O
equipped	O
with	O
32KiB	O
local	O
data	O
share	O
and	O
8kiB	O
of	O
L1	O
cache	O
,	O
while	O
all	O
SIMD	B-Device
cores	O
share	O
64KiB	O
global	O
data	O
share	O
.	O
</s>
<s>
Each	O
memory	B-General_Concept
controller	I-General_Concept
ties	O
to	O
two	O
quad	O
ROPs	B-General_Concept
,	O
one	O
per	O
64-bit	O
channel	O
,	O
and	O
dedicated	O
512KiB	O
L2	O
cache	O
.	O
</s>
<s>
APU	O
that	O
include	O
a	O
TeraScale	B-Architecture
2	O
IGP	O
:	O
</s>
<s>
TeraScale	B-Architecture
3	O
(	O
VLIW4	O
)	O
replaces	O
the	O
previous	O
5-way	O
VLIW	B-General_Concept
designs	O
with	O
a	O
4-way	O
VLIW	B-General_Concept
design	O
.	O
</s>
<s>
The	O
new	O
design	O
also	O
incorporates	O
an	O
additional	O
tessellation	B-Algorithm
unit	O
to	O
improve	O
Direct3D	B-Application
11	O
performance	O
.	O
</s>
<s>
TeraScale	B-Architecture
3	O
is	O
introduced	O
in	O
the	O
Radeon	O
HD	O
6900-branded	O
graphics	O
cards	O
and	O
also	O
implemented	O
in	O
the	O
Trinity	O
and	O
Richland	O
APUs	O
.	O
</s>
<s>
AMD	O
PowerTune	O
,	O
dynamic	B-General_Concept
frequency	I-General_Concept
scaling	I-General_Concept
for	O
GPUs	B-Architecture
,	O
was	O
introduced	O
with	O
the	O
Radeon	O
HD	O
6900	O
series	O
on	O
December	O
15	O
,	O
2010	O
and	O
has	O
seen	O
continued	O
development	O
,	O
as	O
documented	O
in	O
some	O
reviews	O
by	O
AnandTech	O
.	O
</s>
<s>
At	O
HPG11	O
in	O
August	O
2011	O
AMD	O
employees	O
Michael	O
Mantor	O
(	O
Senior	O
Fellow	O
Architect	O
)	O
and	O
Mike	O
Houston	O
(	O
Fellow	O
Architect	O
)	O
presented	O
Graphics	B-Architecture
Core	I-Architecture
Next	I-Architecture
,	O
the	O
microarchitecture	B-General_Concept
succeeding	O
TeraScale	B-Architecture
.	O
</s>
<s>
Microarchitecture	B-General_Concept
TeraScale	B-Architecture
1	O
TeraScale	B-Architecture
2	O
TeraScale	B-Architecture
3	O
Chip1	O
R600	O
RV610	O
RV620	O
RV630	O
RV635	O
RV670	O
RV710	O
RV711	O
RV730	O
RV740	O
RV770	O
RV790	O
Cedar(RV810 )	O
Redwood(RV830 )	O
Juniper(RV840 )	O
Cypress(RV870 )	O
Caicos(RV910 )	O
Turks(RV930 )	O
Barts(RV940 )	O
Cayman(RV970 )	O
Code	O
name	O
Pele	O
Laka	O
Koopa	O
Shaka	O
Wario	O
Boom	O
Luigi	O
Mario	O
Walden	O
Wekiva	O
Spartan	O
?	O
</s>
