<s>
Selene	B-Device
is	O
a	O
supercomputer	B-Architecture
developed	O
by	O
Nvidia	O
,	O
capable	O
of	O
achieving	O
63.460	O
petaflops	O
,	O
ranking	O
as	O
the	O
fifth	O
fastest	B-Operating_System
supercomputer	I-Operating_System
in	O
the	O
world	O
,	O
when	O
it	O
entered	O
the	O
list	O
.	O
</s>
<s>
Selene	B-Device
is	O
based	O
on	O
the	O
Nvidia	B-Operating_System
DGX	I-Operating_System
system	O
consisting	O
of	O
AMD	O
CPUs	O
,	O
Nvidia	O
A100	O
GPUs	O
,	O
and	O
Mellanox	O
HDDR	O
networking	B-Architecture
.	O
</s>
<s>
Selene	B-Device
is	O
based	O
on	O
the	O
Nvidia	B-Operating_System
DGX	I-Operating_System
Superpod	O
,	O
which	O
is	O
a	O
high	O
performance	O
turnkey	O
supercomputer	B-Architecture
solution	O
provided	O
by	O
Nvidia	O
using	O
DGX	O
hardware	O
.	O
</s>
<s>
DGX	O
Superpod	O
is	O
a	O
tightly	O
integrated	O
system	O
that	O
combines	O
high	O
performance	O
DGX	O
compute	O
nodes	O
with	O
fast	O
storage	O
and	O
high	O
bandwidth	O
networking	B-Architecture
.	O
</s>
<s>
Selene	B-Device
was	O
built	O
in	O
three	O
months	O
and	O
is	O
the	O
fastest	O
industrial	O
system	O
in	O
the	O
US	O
while	O
being	O
the	O
second-most	O
energy-efficient	O
supercomputing	B-Architecture
system	O
ever	O
.	O
</s>
<s>
Selene	B-Device
utilizing	O
1080	O
AMD	O
Epyc	O
CPUs	O
and	O
4320	O
A100	O
GPUs	O
is	O
used	O
to	O
train	O
BERT	B-General_Concept
,	O
the	O
natural	B-Language
language	I-Language
processor	I-Language
,	O
in	O
less	O
than	O
16	O
seconds	O
,	O
which	O
usually	O
takes	O
most	O
smaller	O
systems	O
about	O
20	O
minutes	O
to	O
execute	O
.	O
</s>
<s>
IEEE	O
Spectrum	O
reported	O
that	O
as	O
per	O
December	O
2021	O
among	O
all	O
the	O
commercially	O
available	O
supercomputing	B-Architecture
systems	O
Selene	B-Device
topped	O
all	O
the	O
results	O
of	O
MLPerf	B-Algorithm
benchmark	O
,	O
which	O
is	O
the	O
benchmark	O
developed	O
by	O
the	O
consortium	O
of	O
artificial	B-Application
intelligence	I-Application
developers	O
from	O
academia	O
,	O
research	O
labs	O
,	O
and	O
industry	O
aiming	O
to	O
unbiasedly	O
evaluate	O
the	O
training	O
and	O
inference	O
performance	O
for	O
hardware	O
,	O
software	O
,	O
and	O
services	O
used	O
for	O
AI	B-Application
.	O
</s>
<s>
Selene	B-Device
is	O
deployed	O
by	O
the	O
Argonne	O
National	O
Laboratory	O
to	O
research	O
different	O
ways	O
to	O
end	O
the	O
coronavirus	O
.	O
</s>
<s>
Nvidia	O
used	O
Selene	B-Device
to	O
train	O
its	O
GauGAN2	O
AI	B-Application
model	O
,	O
which	O
is	O
used	O
in	O
Nvidia	O
Canvas	O
software	O
to	O
create	O
art	O
using	O
artificial	B-Application
intelligence	I-Application
,	O
using	O
10	O
million	O
landscape	O
images	O
for	O
training	O
.	O
</s>
<s>
GauGAN2	O
AI	B-Application
model	O
uses	O
segmentation	O
mapping	O
,	O
inpainting	O
,	O
and	O
text-to-image	B-General_Concept
generation	I-General_Concept
in	O
a	O
single	O
model	O
to	O
create	O
art	O
.	O
</s>
