<s>
SXM	B-Application
is	O
a	O
high	O
bandwidth	O
socket	B-General_Concept
solution	O
for	O
connecting	O
Nvidia	O
Compute	O
Accelerators	O
to	O
a	O
system	O
.	O
</s>
<s>
Each	O
generation	O
of	O
Nvidia	B-Device
Tesla	I-Device
since	O
P100	O
models	O
,	O
the	O
DGX	B-Operating_System
computer	I-Operating_System
series	I-Operating_System
and	O
the	O
HGX	O
boards	O
come	O
with	O
an	O
SXM	B-Application
socket	B-General_Concept
type	O
that	O
realizes	O
high	O
bandwidth	O
,	O
power	O
delivery	O
and	O
more	O
for	O
the	O
matching	O
GPU	O
daughter	O
cards	O
.	O
</s>
<s>
in	O
their	O
models	O
of	O
the	O
DGX	B-Operating_System
system	I-Operating_System
series	O
.	O
</s>
<s>
Current	O
socket	B-General_Concept
generations	O
are	O
SXM	B-Application
for	O
Pascal	B-General_Concept
based	O
GPUs	O
,	O
SXM2	B-Application
and	O
SXM3	O
for	O
Volta	B-General_Concept
based	O
GPUs	O
,	O
SXM4	B-Application
for	O
Ampere	B-General_Concept
based	O
GPUs	O
,	O
and	O
SXM5	B-Application
for	O
Hopper	B-General_Concept
based	O
GPUs	O
.	O
</s>
<s>
The	O
DGX-1	O
system	O
was	O
the	O
first	O
to	O
be	O
equipped	O
with	O
SXM-2	O
sockets	O
and	O
thus	O
was	O
the	O
first	O
to	O
carry	O
the	O
form	O
factor	O
compatible	O
SXM	B-Application
modules	O
with	O
P100	O
GPUs	O
and	O
later	O
was	O
unveiled	O
to	O
be	O
capable	O
of	O
allowing	O
upgrading	O
to	O
(	O
or	O
being	O
pre-equipped	O
with	O
)	O
SXM2	B-Application
modules	O
with	O
V100	O
GPUs	O
.	O
</s>
<s>
SXM	B-Application
boards	O
are	O
typically	O
built	O
with	O
four	O
or	O
eight	O
GPU	O
slots	O
,	O
although	O
some	O
solutions	O
such	O
as	O
the	O
Nvidia	B-Operating_System
DGX-2	I-Operating_System
connect	O
multiple	O
boards	O
to	O
deliver	O
high	O
performance	O
.	O
</s>
<s>
While	O
third	O
party	O
solutions	O
for	O
SXM	B-Application
boards	O
exist	O
,	O
most	O
System	O
Integrators	O
such	O
as	O
Supermicro	O
use	O
prebuilt	O
Nvidia	O
HGX	O
boards	O
,	O
which	O
come	O
in	O
four	O
or	O
eight	O
socket	B-General_Concept
configurations	O
.	O
</s>
<s>
This	O
solution	O
greatly	O
lowers	O
the	O
cost	O
and	O
difficulty	O
of	O
SXM	B-Application
based	O
GPU	O
servers	O
,	O
and	O
enables	O
compatibility	O
and	O
reliability	O
across	O
all	O
boards	O
of	O
the	O
same	O
generation	O
.	O
</s>
<s>
SXM	B-Application
modules	O
on	O
e.g.	O
</s>
<s>
This	O
as	O
well	O
reduces	O
bottlenecks	O
which	O
would	O
normally	O
be	O
located	O
within	O
CPU	B-General_Concept
and	O
PCIe	O
.	O
</s>
<s>
For	O
example	O
a	O
Hopper-based	B-General_Concept
H100	O
SXM5	B-Application
based	O
GPU	O
can	O
use	O
up	O
to	O
900GB/s	O
of	O
bandwidth	O
across	O
18	O
NVLink	O
4	O
channels	O
,	O
with	O
each	O
contributing	O
a	O
50GB/s	O
of	O
bandwidth	O
;	O
This	O
compared	O
to	O
PCIe	O
5.0	O
,	O
which	O
can	O
handle	O
up	O
to	O
64GB/s	O
of	O
bandwidth	O
within	O
a	O
x16	O
slot	O
.	O
</s>
<s>
Power	O
delivery	O
is	O
also	O
handled	O
by	O
the	O
SXM	B-Application
socket	B-General_Concept
,	O
negating	O
the	O
need	O
for	O
external	O
power	O
cables	O
such	O
as	O
those	O
needed	O
in	O
PCIe	O
equivalent	O
cards	O
.	O
</s>
<s>
This	O
,	O
combined	O
with	O
the	O
horizontal	O
mounting	O
allows	O
cooling	O
options	O
of	O
higher	O
efficiency	O
which	O
in	O
turn	O
allows	O
the	O
SXM	B-Application
based	O
GPUs	O
to	O
operate	O
at	O
a	O
much	O
higher	O
TDP	B-General_Concept
.	O
</s>
<s>
The	O
Hopper-based	B-General_Concept
H100	O
,	O
for	O
example	O
,	O
can	O
draw	O
up	O
to	O
700W	O
solely	O
from	O
the	O
SXM	B-Application
socket	B-General_Concept
.	O
</s>
<s>
The	O
early	O
Nvidia	B-Operating_System
Tegra	I-Operating_System
automotive	O
targeted	O
evaluation	O
board	O
,	O
'	O
Drive	O
PX2	O
 '	O
,	O
had	O
two	O
MXM	B-Architecture
(	O
Mobile	B-Architecture
PCI	I-Architecture
Express	I-Architecture
Module	I-Architecture
)	O
sockets	O
on	O
both	O
sides	O
of	O
the	O
card	O
,	O
this	O
dual	O
MXM	B-Architecture
design	O
can	O
be	O
considered	O
a	O
predecessor	O
to	O
the	O
Nvidia	B-Device
Tesla	I-Device
implementation	O
of	O
the	O
SXM	B-Application
socket	B-General_Concept
.	O
</s>
