<s>
Titan	B-Device
or	O
OLCF-3	B-Device
was	O
a	O
supercomputer	B-Architecture
built	O
by	O
Cray	O
at	O
Oak	O
Ridge	O
National	O
Laboratory	O
for	O
use	O
in	O
a	O
variety	O
of	O
science	O
projects	O
.	O
</s>
<s>
Titan	B-Device
was	O
an	O
upgrade	O
of	O
Jaguar	B-Device
,	O
a	O
previous	O
supercomputer	B-Architecture
at	O
Oak	O
Ridge	O
,	O
that	O
uses	O
graphics	B-Architecture
processing	I-Architecture
units	I-Architecture
(	O
GPUs	B-Architecture
)	O
in	O
addition	O
to	O
conventional	O
central	B-General_Concept
processing	I-General_Concept
units	I-General_Concept
(	O
CPUs	B-Device
)	O
.	O
</s>
<s>
Titan	B-Device
was	O
the	O
first	O
such	O
hybrid	O
to	O
perform	O
over	O
10petaFLOPS	O
.	O
</s>
<s>
Titan	B-Device
was	O
eclipsed	O
at	O
Oak	O
Ridge	O
by	O
Summit	B-Device
in	O
2019	O
,	O
which	O
was	O
built	O
by	O
IBM	O
and	O
features	O
fewer	O
nodes	O
with	O
much	O
greater	O
GPU	B-Architecture
capability	O
per	O
node	O
as	O
well	O
as	O
local	O
per-node	O
non-volatile	B-General_Concept
caching	B-General_Concept
of	O
file	O
data	O
from	O
the	O
system	O
's	O
parallel	B-Application
file	I-Application
system	I-Application
.	O
</s>
<s>
Titan	B-Device
employed	O
AMD	B-General_Concept
Opteron	I-General_Concept
CPUs	B-Device
in	O
conjunction	O
with	O
Nvidia	B-Device
Tesla	I-Device
GPUs	B-Architecture
to	O
improve	O
energy	O
efficiency	O
while	O
providing	O
an	O
order	O
of	O
magnitude	O
increase	O
in	O
computational	O
power	O
over	O
Jaguar	B-Device
.	O
</s>
<s>
It	O
used	O
18,688	O
CPUs	O
paired	O
with	O
an	O
equal	O
number	O
of	O
GPUs	B-Architecture
to	O
perform	O
at	O
a	O
theoretical	O
peak	O
of	O
27petaFLOPS	O
;	O
in	O
the	O
LINPACK	B-Device
benchmark	I-Device
used	O
to	O
rank	O
supercomputers	B-Architecture
 '	O
speed	O
,	O
it	O
performed	O
at	O
17.59petaFLOPS	O
.	O
</s>
<s>
This	O
was	O
enough	O
to	O
take	O
first	O
place	O
in	O
the	O
November	O
2012	O
list	O
by	O
the	O
TOP500	B-Operating_System
organization	O
,	O
but	O
Tianhe-2	B-Device
overtook	O
it	O
on	O
the	O
June	O
2013	O
list	O
.	O
</s>
<s>
Titan	B-Device
was	O
available	O
for	O
any	O
scientific	O
purpose	O
;	O
access	O
depends	O
on	O
the	O
importance	O
of	O
the	O
project	O
and	O
its	O
potential	O
to	O
exploit	O
the	O
hybrid	O
architecture	O
.	O
</s>
<s>
Any	O
selected	O
programs	O
must	O
also	O
be	O
executable	O
on	O
other	O
supercomputers	B-Architecture
to	O
avoid	O
sole	O
dependence	O
on	O
Titan	B-Device
.	O
</s>
<s>
They	O
dealt	O
mostly	O
with	O
molecular	O
scale	O
physics	O
or	O
climate	B-Application
models	I-Application
,	O
while	O
25	O
others	O
were	O
queued	O
behind	O
them	O
.	O
</s>
<s>
The	O
inclusion	O
of	O
GPUs	B-Architecture
compelled	O
authors	O
to	O
alter	O
their	O
programs	O
.	O
</s>
<s>
The	O
modifications	O
typically	O
increased	O
the	O
degree	O
of	O
parallelism	B-Operating_System
,	O
given	O
that	O
GPUs	B-Architecture
offer	O
many	O
more	O
simultaneous	O
threads	B-Operating_System
than	O
CPUs	B-Device
.	O
</s>
<s>
The	O
changes	O
often	O
yield	O
greater	O
performance	O
even	O
on	O
CPU-only	O
machines	O
.	O
</s>
<s>
Plans	O
to	O
create	O
a	O
supercomputer	B-Architecture
capable	O
of	O
20	O
petaFLOPS	O
at	O
the	O
Oak	B-Device
Ridge	I-Device
Leadership	I-Device
Computing	I-Device
Facility	I-Device
(	O
OLCF	B-Device
)	O
at	O
Oak	O
Ridge	O
National	O
Laboratory	O
(	O
ORNL	O
)	O
originated	O
as	O
far	O
back	O
as	O
2005	O
,	O
when	O
Jaguar	B-Device
was	O
built	O
.	O
</s>
<s>
Titan	B-Device
will	O
itself	O
be	O
replaced	O
by	O
an	O
approximately	O
200petaFLOPS	O
system	O
in	O
2016	O
as	O
part	O
of	O
ORNL	O
's	O
plan	O
to	O
operate	O
an	O
exascale	O
(	O
1000petaFLOPS	O
to	O
1exaFLOPS	O
)	O
machine	O
by	O
2020	O
.	O
</s>
<s>
The	O
initial	O
plan	O
to	O
build	O
a	O
new	O
15,000	O
squaremeter	O
(	O
160,000	O
ft2	O
)	O
building	O
for	O
Titan	B-Device
,	O
was	O
discarded	O
in	O
favor	O
of	O
using	O
Jaguar	B-Device
's	O
existing	O
infrastructure	O
.	O
</s>
<s>
The	O
precise	O
system	O
architecture	O
was	O
not	O
finalized	O
until	O
2010	O
,	O
although	O
a	O
deal	O
with	O
Nvidia	O
to	O
supply	O
the	O
GPUs	B-Architecture
was	O
signed	O
in	O
2009	O
.	O
</s>
<s>
Titan	B-Device
was	O
first	O
announced	O
at	O
the	O
private	O
ACM/IEEE	O
Supercomputing	B-Architecture
Conference	O
(	O
SC10	O
)	O
on	O
November	O
16	O
,	O
2010	O
,	O
and	O
was	O
publicly	O
announced	O
on	O
October	O
11	O
,	O
2011	O
,	O
as	O
the	O
first	O
phase	O
of	O
the	O
Titan	B-Device
upgrade	O
began	O
.	O
</s>
<s>
Jaguar	B-Device
had	O
received	O
various	O
upgrades	O
since	O
its	O
creation	O
.	O
</s>
<s>
It	O
began	O
with	O
the	O
Cray	B-Device
XT3	I-Device
platform	O
that	O
yielded	O
25	O
teraFLOPS	O
.	O
</s>
<s>
By	O
2008	O
,	O
Jaguar	B-Device
had	O
been	O
expanded	O
with	O
more	O
cabinets	O
and	O
upgraded	O
to	O
the	O
XT4	B-Device
platform	O
,	O
reaching	O
263	O
teraFLOPS	O
.	O
</s>
<s>
In	O
2009	O
,	O
it	O
was	O
upgraded	O
to	O
the	O
XT5	B-Device
platform	O
,	O
hitting	O
1.4	O
petaFLOPS	O
.	O
</s>
<s>
Its	O
final	O
upgrades	O
brought	O
Jaguar	B-Device
to	O
1.76	O
petaFLOPS	O
.	O
</s>
<s>
Titan	B-Device
was	O
funded	O
primarily	O
by	O
the	O
US	O
Department	O
of	O
Energy	O
through	O
ORNL	O
.	O
</s>
<s>
Funding	O
was	O
sufficient	O
to	O
purchase	O
the	O
CPUs	B-Device
but	O
not	O
all	O
of	O
the	O
GPUs	B-Architecture
so	O
the	O
National	O
Oceanic	O
and	O
Atmospheric	O
Administration	O
agreed	O
to	O
fund	O
the	O
remaining	O
nodes	O
in	O
return	O
for	O
computing	O
time	O
.	O
</s>
<s>
ORNL	O
scientific	O
computing	O
chief	O
Jeff	O
Nichols	O
noted	O
that	O
Titan	B-Device
cost	O
approximately	O
$60million	O
upfront	O
,	O
of	O
which	O
the	O
NOAA	O
contribution	O
was	O
less	O
than	O
$10million	O
,	O
but	O
precise	O
figures	O
were	O
covered	O
by	O
non-disclosure	O
agreements	O
.	O
</s>
<s>
Between	O
October	O
and	O
December	O
,	O
96	O
of	O
Jaguar	B-Device
's	O
200	O
cabinets	O
,	O
each	O
containing	O
24XT5	O
blades	B-Architecture
(	O
two	O
6-core	O
CPUs	B-Device
per	O
node	O
,	O
four	O
nodes	O
per	O
blade	O
)	O
,	O
were	O
upgraded	O
to	O
XK7	B-Device
blade	O
(	O
one	O
16-core	B-Architecture
CPU	B-Device
per	O
node	O
,	O
four	O
nodes	O
per	O
blade	O
)	O
while	O
the	O
remainder	O
of	O
the	O
machine	O
remained	O
in	O
use	O
.	O
</s>
<s>
In	O
December	O
,	O
computation	O
was	O
moved	O
to	O
the	O
96	O
XK7	B-Device
cabinets	O
while	O
the	O
remaining	O
104	O
cabinets	O
were	O
upgraded	O
to	O
XK7	B-Device
blades	B-Architecture
.	O
</s>
<s>
ORNL	O
's	O
external	O
ESnet	O
connection	O
was	O
upgraded	O
from	O
10Gbit/s	O
to	O
100Gbit/s	O
and	O
the	O
system	O
interconnect	O
(	O
the	O
network	O
over	O
which	O
CPUs	B-Device
communicate	O
with	O
each	O
other	O
)	O
was	O
updated	O
.	O
</s>
<s>
The	O
Seastar	O
design	O
used	O
in	O
Jaguar	B-Device
was	O
upgraded	O
to	O
the	O
Gemini	O
interconnect	O
used	O
in	O
Titan	B-Device
which	O
connects	O
the	O
nodes	O
into	O
a	O
direct	O
3D	O
torus	B-Operating_System
interconnect	I-Operating_System
network	O
.	O
</s>
<s>
Gemini	O
uses	O
wormhole	B-Protocol
flow	I-Protocol
control	I-Protocol
internally	O
.	O
</s>
<s>
The	O
system	O
memory	B-Architecture
was	O
doubled	O
to	O
584TiB	O
.	O
</s>
<s>
960	O
of	O
the	O
XK7	B-Device
nodes	O
(	O
10cabinets	O
)	O
were	O
fitted	O
with	O
a	O
Fermi	B-General_Concept
based	O
GPU	B-Architecture
as	O
Kepler	B-General_Concept
GPUs	B-Architecture
were	O
not	O
then	O
available	O
;	O
these	O
960nodes	O
were	O
referred	O
to	O
as	O
TitanDev	O
and	O
used	O
to	O
test	O
code	O
.	O
</s>
<s>
This	O
first	O
phase	O
of	O
the	O
upgrade	O
increased	O
the	O
peak	O
performance	O
of	O
Jaguar	B-Device
to	O
3.3petaFLOPS	O
.	O
</s>
<s>
Beginning	O
on	O
September	O
13	O
,	O
2012	O
,	O
Nvidia	O
K20X	O
GPUs	B-Architecture
were	O
fitted	O
to	O
all	O
of	O
Jaguar	B-Device
's	O
XK7	B-Device
compute	O
blades	B-Architecture
,	O
including	O
the	O
960	O
TitanDev	O
nodes	O
.	O
</s>
<s>
In	O
October	O
,	O
the	O
task	O
was	O
completed	O
and	O
the	O
computer	O
was	O
finally	O
renamed	O
Titan	B-Device
.	O
</s>
<s>
In	O
March	O
2013	O
,	O
Nvidia	O
launched	O
the	O
GTX	O
Titan	B-Device
,	O
a	O
consumer	O
graphics	B-Device
card	I-Device
that	O
uses	O
the	O
same	O
GPU	B-Architecture
die	O
as	O
the	O
K20X	O
GPUs	B-Architecture
in	O
Titan	B-Device
.	O
</s>
<s>
Titan	B-Device
underwent	O
acceptance	O
testing	O
in	O
early	O
2013	O
but	O
only	O
completed	O
92%	O
of	O
the	O
tests	O
,	O
short	O
of	O
the	O
required	O
95%	O
.	O
</s>
<s>
The	O
problem	O
was	O
discovered	O
to	O
be	O
excess	O
gold	O
in	O
the	O
female	O
edge	O
connectors	O
of	O
the	O
motherboards	B-Device
 '	O
PCIe	O
slots	O
causing	O
cracks	O
in	O
the	O
motherboards	B-Device
 '	O
solder	O
.	O
</s>
<s>
Throughout	O
the	O
repairs	O
users	O
were	O
given	O
access	O
to	O
the	O
available	O
CPUs	B-Device
.	O
</s>
<s>
Titan	B-Device
's	O
hardware	O
has	O
a	O
theoretical	O
peak	O
performance	O
of	O
27petaFLOPS	O
with	O
"	O
perfect	O
"	O
software	O
.	O
</s>
<s>
On	O
November	O
12	O
,	O
2012	O
,	O
the	O
TOP500	B-Operating_System
organization	O
that	O
ranks	O
the	O
world	O
's	O
supercomputers	B-Architecture
by	O
LINPACK	B-Application
performance	O
,	O
ranked	O
Titan	B-Device
first	O
at	O
17.59petaFLOPS	O
,	O
displacing	O
IBM	B-Device
Sequoia	I-Device
.	O
</s>
<s>
Titan	B-Device
also	O
ranked	O
third	O
on	O
the	O
Green500	B-Application
,	O
the	O
same	O
500	O
supercomputers	B-Architecture
ranked	O
in	O
terms	O
of	O
energy	O
efficiency	O
.	O
</s>
<s>
In	O
the	O
June	O
2013	O
TOP500	B-Operating_System
ranking	O
,	O
Titan	B-Device
fell	O
to	O
second	O
place	O
behind	O
Tianhe-2	B-Device
and	O
to	O
twenty-ninth	O
on	O
the	O
Green500	B-Application
list	O
.	O
</s>
<s>
Titan	B-Device
did	O
not	O
re-test	O
for	O
the	O
June	O
2013	O
ranking	O
,	O
because	O
it	O
would	O
still	O
have	O
ranked	O
second	O
,	O
at	O
27petaFLOPS	O
.	O
</s>
<s>
Titan	B-Device
uses	O
Jaguar	B-Device
's	O
200	O
cabinets	O
,	O
covering	O
404squaremeters	O
(	O
4,352	O
ft2	O
)	O
,	O
with	O
replaced	O
internals	O
and	O
upgraded	O
networking	O
.	O
</s>
<s>
Reusing	O
Jaguar	B-Device
's	O
power	O
and	O
cooling	O
systems	O
saved	O
approximately	O
$20	O
million	O
.	O
</s>
<s>
At	O
its	O
peak	O
,	O
Titan	B-Device
draws	O
8.2MW	O
,	O
1.2MW	O
more	O
than	O
Jaguar	B-Device
,	O
but	O
runs	O
almost	O
ten	O
times	O
as	O
fast	O
in	O
terms	O
of	O
floating	B-Algorithm
point	I-Algorithm
calculations	O
.	O
</s>
<s>
Titan	B-Device
has	O
18,688	O
nodes	O
(	O
4	O
nodes	O
per	O
blade	O
,	O
24	O
blades	B-Architecture
per	O
cabinet	O
)	O
,	O
each	O
containing	O
a	O
16-core	B-Architecture
AMD	B-General_Concept
Opteron	I-General_Concept
6274	O
CPU	B-Device
with	O
32GB	O
of	O
DDR3	O
ECC	B-General_Concept
memory	I-General_Concept
and	O
an	O
Nvidia	B-Device
Tesla	I-Device
K20X	O
GPU	B-Architecture
with	O
6GB	O
GDDR5	O
ECC	B-General_Concept
memory	I-General_Concept
.	O
</s>
<s>
There	O
are	O
a	O
total	O
of	O
299,008	O
processor	O
cores	O
,	O
and	O
a	O
total	O
of	O
693.6TiB	O
of	O
CPU	B-Device
and	O
GPU	B-Architecture
RAM	B-Architecture
.	O
</s>
<s>
Initially	O
,	O
Titan	B-Device
used	O
Jaguar	B-Device
's	O
10PB	O
of	O
Lustre	B-Application
storage	O
with	O
a	O
transfer	O
speed	O
of	O
240GB/s	O
,	O
but	O
in	O
April	O
2013	O
,	O
the	O
storage	O
was	O
upgraded	O
to	O
40PB	O
with	O
a	O
transfer	O
rate	O
of	O
1.4TB/s	O
.	O
</s>
<s>
GPUs	B-Architecture
were	O
selected	O
for	O
their	O
vastly	O
higher	O
parallel	B-Operating_System
processing	I-Operating_System
efficiency	O
over	O
CPUs	B-Device
.	O
</s>
<s>
Although	O
the	O
GPUs	B-Architecture
have	O
a	O
slower	O
clock	O
speed	O
than	O
the	O
CPUs	B-Device
,	O
each	O
GPU	B-Architecture
contains	O
2,688	O
CUDA	B-Architecture
cores	O
at	O
732MHz	O
,	O
resulting	O
in	O
a	O
faster	O
overall	O
system	O
.	O
</s>
<s>
Consequently	O
,	O
the	O
CPUs	B-Device
 '	O
cores	O
are	O
used	O
to	O
allocate	O
tasks	O
to	O
the	O
GPUs	B-Architecture
rather	O
than	O
directly	O
processing	O
the	O
data	O
as	O
in	O
conventional	O
supercomputers	B-Architecture
.	O
</s>
<s>
Titan	B-Device
runs	O
the	O
Cray	O
Linux	B-Application
Environment	O
,	O
a	O
full	O
version	O
of	O
Linux	B-Application
on	O
the	O
login	O
nodes	O
that	O
users	O
directly	O
access	O
,	O
but	O
a	O
smaller	O
,	O
more	O
efficient	O
version	O
on	O
the	O
compute	O
nodes	O
.	O
</s>
<s>
Titan	B-Device
's	O
components	O
are	O
air-cooled	O
by	O
heat	O
sinks	O
,	O
but	O
the	O
air	O
is	O
chilled	O
before	O
being	O
pumped	O
through	O
the	O
cabinets	O
.	O
</s>
<s>
The	O
system	O
has	O
a	O
cooling	O
capacity	O
of	O
23.2	O
MW	O
(	O
6600	O
tons	O
)	O
and	O
works	O
by	O
chilling	O
water	O
to	O
5.5	O
°C	O
(	O
42°F	O
)	O
,	O
which	O
in	O
turn	O
cools	O
recirculated	O
air	O
.	O
</s>
<s>
Researchers	O
also	O
have	O
access	O
to	O
EVEREST	O
(	O
Exploratory	O
Visualization	B-Application
Environment	O
for	O
Research	O
and	O
Technology	O
)	O
to	O
better	O
understand	O
the	O
data	O
that	O
Titan	B-Device
outputs	O
.	O
</s>
<s>
EVEREST	O
is	O
a	O
visualization	B-Application
room	O
with	O
a	O
10	O
by	O
3meter	O
(	O
33	O
by	O
10ft	O
)	O
screen	O
and	O
a	O
smaller	O
,	O
secondary	O
screen	O
.	O
</s>
<s>
The	O
screens	O
are	O
37	O
and	O
33megapixels	O
respectively	O
with	O
stereoscopic	B-Algorithm
3D	I-Algorithm
capability	O
.	O
</s>
<s>
In	O
2009	O
,	O
the	O
Oak	B-Device
Ridge	I-Device
Leadership	I-Device
Computing	I-Device
Facility	I-Device
that	O
manages	O
Titan	B-Device
narrowed	O
the	O
fifty	O
applications	O
for	O
first	O
use	O
of	O
the	O
supercomputer	B-Architecture
down	O
to	O
six	O
"	O
vanguard	O
"	O
codes	O
chosen	O
for	O
the	O
importance	O
of	O
the	O
research	O
and	O
for	O
their	O
ability	O
to	O
fully	O
utilize	O
the	O
system	O
.	O
</s>
<s>
The	O
six	O
vanguard	O
projects	O
to	O
use	O
Titan	B-Device
were	O
:	O
</s>
<s>
In	O
2009	O
,	O
using	O
Jaguar	B-Device
,	O
it	O
produced	O
the	O
first	O
fully	O
resolved	O
simulation	O
of	O
autoigniting	O
hydrocarbon	O
flames	O
relevant	O
to	O
the	O
efficiency	O
of	O
direct	O
injection	O
diesel	O
engines	O
.	O
</s>
<s>
WL-LSMS	O
simulates	O
the	O
interactions	O
between	O
electrons	O
and	O
atoms	O
in	O
magnetic	O
materials	O
at	O
temperatures	O
other	O
than	O
absolute	B-Application
zero	I-Application
.	O
</s>
<s>
An	O
earlier	O
version	O
of	O
the	O
code	O
was	O
the	O
first	O
to	O
perform	O
at	O
greater	O
than	O
one	O
petaFLOPS	O
on	O
Jaguar	B-Device
.	O
</s>
<s>
The	O
performance	O
of	O
Denovo	O
on	O
conventional	O
CPU-based	O
machines	O
doubled	O
after	O
the	O
tweaks	O
for	O
Titan	B-Device
and	O
it	O
performs	O
3.5	O
times	O
faster	O
on	O
Titan	B-Device
than	O
it	O
did	O
on	O
Jaguar	B-Device
.	O
</s>
<s>
Large-scale	B-Application
Atomic/Molecular	I-Application
Massively	I-Application
Parallel	I-Application
Simulator	I-Application
(	O
LAMMPS	B-Application
)	O
is	O
a	O
molecular	O
dynamics	O
code	O
that	O
simulates	O
particles	O
across	O
a	O
range	O
of	O
scales	O
,	O
from	O
quantum	O
to	O
relativistic	O
,	O
to	O
improve	O
materials	O
science	O
with	O
potential	O
applications	O
in	O
semiconductor	O
,	O
biomolecule	O
and	O
polymer	B-Language
development	O
.	O
</s>
<s>
CAM-SE	O
will	O
allow	O
greater	O
accuracy	O
in	O
climate	B-Application
simulations	I-Application
.	O
</s>
<s>
Non-Equilibrium	O
Radiation	O
Diffusion	O
(	O
NRDF	O
)	O
plots	O
non-charged	O
particles	O
through	O
supernovae	O
with	O
potential	O
applications	O
in	O
laser	O
fusion	O
,	O
fluid	O
dynamics	O
,	O
medical	B-Application
imaging	I-Application
,	O
nuclear	O
reactors	O
,	O
energy	O
storage	O
and	O
combustion	O
.	O
</s>
<s>
Its	O
Chimera	O
code	O
uses	O
hundreds	O
of	O
partial	O
differential	O
equations	O
to	O
track	O
the	O
energy	O
,	O
angle	O
,	O
angle	O
of	O
scatter	O
and	O
type	O
of	O
each	O
neutrino	O
modeled	O
in	O
a	O
star	O
going	O
supernova	B-Application
,	O
resulting	O
in	O
millions	O
of	O
individual	O
equations	O
.	O
</s>
<s>
Bonsai	O
is	O
a	O
gravitational	O
tree	O
code	O
for	O
n-body	B-Algorithm
simulation	I-Algorithm
.	O
</s>
<s>
VERA	O
is	O
a	O
light-water	O
reactor	O
simulation	O
written	O
at	O
the	O
Consortium	O
for	O
Advanced	O
Simulation	O
of	O
Light	O
Water	O
Reactors	O
(	O
CASL	O
)	O
on	O
Jaguar	B-Device
.	O
</s>
<s>
Although	O
not	O
one	O
of	O
the	O
first	O
six	O
projects	O
,	O
VERA	O
was	O
planned	O
to	O
run	O
on	O
Titan	B-Device
after	O
optimization	O
with	O
assistance	O
from	O
CAAR	O
and	O
testing	O
on	O
TitanDev	O
.	O
</s>
<s>
Computer	O
scientist	O
Tom	O
Evans	O
found	O
that	O
the	O
adaption	O
to	O
Titan	B-Device
's	O
hybrid	O
architecture	O
was	O
more	O
difficult	O
than	O
to	O
previous	O
CPU-based	O
supercomputers	B-Architecture
.	O
</s>
<s>
He	O
aimed	O
to	O
simulate	O
an	O
entire	O
reactor	O
fuel	O
cycle	O
,	O
an	O
eighteen	O
to	O
thirty-six	O
month-long	O
process	O
,	O
in	O
one	O
week	O
on	O
Titan	B-Device
.	O
</s>
<s>
In	O
2013	O
thirty-one	O
codes	O
were	O
planned	O
to	O
run	O
on	O
Titan	B-Device
,	O
typically	O
four	O
or	O
five	O
at	O
any	O
one	O
time	O
.	O
</s>
<s>
The	O
code	O
of	O
many	O
projects	O
has	O
to	O
be	O
modified	O
to	O
suit	O
the	O
GPU	B-Architecture
processing	O
of	O
Titan	B-Device
,	O
but	O
each	O
code	O
is	O
required	O
to	O
be	O
executable	O
on	O
CPU-based	O
systems	O
so	O
that	O
projects	O
do	O
not	O
become	O
solely	O
dependent	O
on	O
Titan	B-Device
.	O
</s>
<s>
OLCF	B-Device
formed	O
the	O
Center	O
for	O
Accelerated	O
Application	O
Readiness	O
(	O
CAAR	O
)	O
to	O
aid	O
with	O
the	O
adaptation	O
process	O
.	O
</s>
<s>
It	O
holds	O
developer	O
workshops	O
at	O
Nvidia	O
headquarters	O
to	O
educate	O
users	O
about	O
the	O
architecture	O
,	O
compilers	B-Language
and	O
applications	O
on	O
Titan	B-Device
.	O
</s>
<s>
CAAR	O
has	O
been	O
working	O
on	O
compilers	B-Language
with	O
Nvidia	O
and	O
code	O
vendors	O
to	O
integrate	O
directives	O
for	O
GPUs	B-Architecture
into	O
their	O
programming	O
languages	O
.	O
</s>
<s>
Researchers	O
can	O
thus	O
express	O
parallelism	B-Operating_System
in	O
their	O
code	O
with	O
their	O
existing	O
programming	O
language	O
,	O
typically	O
Fortran	B-Application
,	O
C	B-Language
or	O
C++	B-Language
,	O
and	O
the	O
compiler	B-Language
can	O
express	O
it	O
to	O
the	O
GPUs	B-Architecture
.	O
</s>
<s>
Dr.	O
Bronson	O
Messer	O
,	O
a	O
computational	O
astrophysicist	O
,	O
said	O
of	O
the	O
task	O
:	O
"	O
an	O
application	O
using	O
Titan	B-Device
to	O
the	O
utmost	O
must	O
also	O
find	O
a	O
way	O
to	O
keep	O
the	O
GPU	B-Architecture
busy	O
,	O
remembering	O
all	O
the	O
while	O
that	O
the	O
GPU	B-Architecture
is	O
fast	O
,	O
but	O
less	O
flexible	O
than	O
the	O
CPU.	O
"	O
</s>
<s>
Moab	B-General_Concept
Cluster	I-General_Concept
Suite	I-General_Concept
is	O
used	O
to	O
prioritize	O
jobs	O
to	O
nodes	O
to	O
keep	O
utilization	O
high	O
;	O
it	O
improved	O
efficiency	O
from	O
70%	O
to	O
approximately	O
95%	O
in	O
the	O
tested	O
software	O
.	O
</s>
<s>
Some	O
projects	O
found	O
that	O
the	O
changes	O
increased	O
efficiency	O
of	O
their	O
code	O
on	O
non-GPU	O
machines	O
;	O
the	O
performance	O
of	O
Denovo	O
doubled	O
on	O
CPU-based	O
machines	O
.	O
</s>
<s>
The	O
amount	O
of	O
code	O
alteration	O
required	O
to	O
run	O
on	O
the	O
GPUs	B-Architecture
varies	O
by	O
project	O
.	O
</s>
<s>
According	O
to	O
Dr.	O
Messer	O
of	O
NRDF	O
,	O
only	O
a	O
small	O
percentage	O
of	O
his	O
code	O
runs	O
on	O
GPUs	B-Architecture
because	O
the	O
calculations	O
are	O
relatively	O
simple	O
but	O
processed	O
repeatedly	O
and	O
in	O
parallel	O
.	O
</s>
<s>
NRDF	O
is	O
written	O
in	O
CUDA	B-Architecture
Fortran	I-Architecture
,	O
a	O
version	O
of	O
Fortran	B-Application
with	O
CUDA	B-Architecture
extensions	O
for	O
the	O
GPUs	B-Architecture
.	O
</s>
<s>
Chimera	O
's	O
third	O
"	O
head	O
"	O
was	O
the	O
first	O
to	O
run	O
on	O
the	O
GPUs	B-Architecture
as	O
the	O
nuclear	O
burning	O
could	O
most	O
easily	O
be	O
simulated	O
by	O
GPU	B-Architecture
architecture	O
.	O
</s>
<s>
On	O
Jaguar	B-Device
,	O
the	O
project	O
modeled	O
14	O
or	O
15	O
nuclear	O
species	O
but	O
Messer	O
anticipated	O
simulating	O
up	O
to	O
200	O
species	O
,	O
allowing	O
far	O
greater	O
precision	O
when	O
comparing	O
the	O
simulation	O
to	O
empirical	O
observation	O
.	O
</s>
