<s>
GPU	B-General_Concept
virtualization	I-General_Concept
refers	O
to	O
technologies	O
that	O
allow	O
the	O
use	O
of	O
a	O
GPU	B-Architecture
to	O
accelerate	B-General_Concept
graphics	O
or	O
GPGPU	B-Architecture
applications	O
running	O
on	O
a	O
virtual	B-Architecture
machine	I-Architecture
.	O
</s>
<s>
GPU	B-General_Concept
virtualization	I-General_Concept
is	O
used	O
in	O
various	O
applications	O
such	O
as	O
desktop	B-Protocol
virtualization	I-Protocol
,	O
cloud	B-Application
gaming	I-Application
and	O
computational	O
science	O
(	O
e.g.	O
</s>
<s>
GPU	B-General_Concept
virtualization	I-General_Concept
implementations	O
generally	O
involve	O
one	O
or	O
more	O
of	O
the	O
following	O
techniques	O
:	O
device	O
emulation	O
,	O
API	B-General_Concept
remoting	B-Operating_System
,	O
fixed	O
pass-through	O
and	O
mediated	O
pass-through	O
.	O
</s>
<s>
Each	O
technique	O
presents	O
different	O
trade-offs	O
regarding	O
virtual	B-Architecture
machine	I-Architecture
to	O
GPU	B-Architecture
consolidation	B-Algorithm
ratio	I-Algorithm
,	O
graphics	B-Architecture
acceleration	I-Architecture
,	O
rendering	O
fidelity	B-Algorithm
and	O
feature	B-Architecture
support	O
,	O
portability	B-Architecture
to	O
different	O
hardware	O
,	O
isolation	O
between	O
virtual	B-Architecture
machines	I-Architecture
,	O
and	O
support	O
for	O
suspending/resuming	O
and	O
live	O
migration	O
.	O
</s>
<s>
In	O
API	B-General_Concept
remoting	B-Operating_System
or	O
API	B-General_Concept
forwarding	O
,	O
calls	O
to	O
graphical	O
APIs	B-General_Concept
from	O
guest	O
applications	O
are	O
forwarded	O
to	O
the	O
host	O
by	O
remote	B-Operating_System
procedure	I-Operating_System
call	I-Operating_System
,	O
and	O
the	O
host	O
then	O
executes	O
graphical	O
commands	O
from	O
multiple	O
guests	O
using	O
the	O
host	O
's	O
GPU	B-Architecture
as	O
a	O
single	O
user	O
.	O
</s>
<s>
This	O
technique	O
allows	O
sharing	O
GPU	B-Architecture
resources	O
between	O
multiple	O
guests	O
and	O
the	O
host	O
when	O
the	O
GPU	B-Architecture
does	O
not	O
support	O
hardware-assisted	B-General_Concept
virtualization	I-General_Concept
.	O
</s>
<s>
Hypervisors	O
usually	O
use	O
shared	B-Operating_System
memory	I-Operating_System
between	O
guest	O
and	O
host	O
to	O
maximize	O
performance	O
and	O
minimize	O
latency	O
.	O
</s>
<s>
Using	O
a	O
network	B-General_Concept
interface	I-General_Concept
instead	O
(	O
a	O
common	O
approach	O
in	O
distributed	B-Operating_System
rendering	I-Operating_System
)	O
,	O
third-party	O
software	O
can	O
add	O
support	O
for	O
specific	O
APIs	B-General_Concept
(	O
e.g.	O
</s>
<s>
rCUDA	B-General_Concept
for	O
CUDA	B-Architecture
)	O
or	O
add	O
support	O
for	O
typical	O
APIs	B-General_Concept
(	O
e.g.	O
</s>
<s>
VMGL	O
for	O
OpenGL	B-Application
)	O
when	O
it	O
is	O
not	O
supported	O
by	O
the	O
hypervisor	O
's	O
software	O
package	O
,	O
although	O
network	B-Protocol
delay	I-Protocol
and	O
serialization	B-Application
overhead	O
may	O
outweigh	O
the	O
benefits	O
.	O
</s>
<s>
+Application	O
support	O
from	O
API	B-General_Concept
remoting	B-Operating_System
virtualization	O
technologiesTechnologyDirect3DOpenGLVulkanOpenCLVMware	O
Virtual	O
Shared	O
Graphics	B-Architecture
Acceleration	I-Architecture
(	O
vSGA	O
)	O
Parallels	B-Language
Desktop	I-Language
for	I-Language
Mac	I-Language
3D	O
accelerationHyper-V	O
RemoteFX	O
vGPUVirtualBox	O
Guest	B-Operating_System
Additions	I-Operating_System
3D	O
driverThincast	O
Workstation	O
-	O
Virtual	O
3DHi	O
!	O
</s>
<s>
In	O
fixed	O
pass-through	O
or	O
GPU	B-Architecture
pass-through	O
(	O
a	O
special	O
case	O
of	O
PCI	B-Protocol
pass-through	O
)	O
,	O
a	O
GPU	B-Architecture
is	O
accessed	O
directly	O
by	O
a	O
single	O
virtual	B-Architecture
machine	I-Architecture
exclusively	O
and	O
permanently	O
.	O
</s>
<s>
This	O
technique	O
achieves	O
96100%	O
of	O
native	O
performance	O
and	O
high	O
fidelity	B-Algorithm
,	O
but	O
the	O
acceleration	B-General_Concept
provided	O
by	O
the	O
GPU	B-Architecture
cannot	O
be	O
shared	O
between	O
multiple	O
virtual	B-Architecture
machines	I-Architecture
.	O
</s>
<s>
As	O
such	O
,	O
it	O
has	O
the	O
lowest	O
consolidation	B-Algorithm
ratio	I-Algorithm
and	O
the	O
highest	O
cost	O
,	O
as	O
each	O
graphics-accelerated	O
virtual	B-Architecture
machine	I-Architecture
requires	O
an	O
additional	O
physical	O
GPU	B-Architecture
.	O
</s>
<s>
VirtualBox	B-Operating_System
removed	O
support	O
for	O
PCI	B-Protocol
pass-through	O
in	O
version	O
6.1.0	O
.	O
</s>
<s>
For	O
certain	O
GPU	B-Architecture
models	O
,	O
Nvidia	O
and	O
AMD	O
video	O
card	O
drivers	O
attempt	O
to	O
detect	O
the	O
GPU	B-Architecture
is	O
being	O
accessed	O
by	O
a	O
virtual	B-Architecture
machine	I-Architecture
and	O
disable	O
some	O
or	O
all	O
GPU	B-Architecture
features	O
.	O
</s>
<s>
NVIDIA	O
has	O
recently	O
changed	O
virtualization	O
rules	O
for	O
consumer	O
GPUs	B-Architecture
by	O
disabling	O
the	O
check	O
in	O
GeForce	O
Game	O
Ready	O
driver	O
465.xx	O
and	O
later	O
.	O
</s>
<s>
For	O
NVIDIA	O
,	O
various	O
architectures	O
of	O
desktop	O
and	O
laptop	O
consumer	O
GPUs	B-Architecture
can	O
be	O
passed	O
through	O
in	O
various	O
ways	O
.	O
</s>
<s>
For	O
desktop	O
graphics	O
cards	O
,	O
passthrough	O
can	O
be	O
done	O
via	O
the	O
KVM	B-Application
using	O
either	O
the	O
legacy	O
or	O
UEFI	O
BIOS	O
configuration	O
via	O
SeaBIOS	O
and	O
OVMF	O
,	O
respectively	O
.	O
</s>
<s>
For	O
desktops	O
,	O
most	O
graphics	O
cards	O
can	O
be	O
passed	O
through	O
,	O
although	O
for	O
graphics	O
cards	O
with	O
the	O
Pascal	O
architecture	O
or	O
older	O
,	O
the	O
VBIOS	O
of	O
the	O
graphics	O
card	O
must	O
be	O
passed	O
through	O
in	O
the	O
virtual	B-Architecture
machine	I-Architecture
if	O
the	O
GPU	B-Architecture
is	O
used	O
to	O
boot	O
the	O
host	O
.	O
</s>
<s>
For	O
laptops	O
that	O
have	O
NVIDIA	O
Optimus	O
on	O
as	O
well	O
as	O
rendering	O
through	O
the	O
CPU	O
's	O
integrated	O
graphics	O
framebuffer	O
as	O
opposed	O
to	O
its	O
own	O
,	O
the	O
passthrough	O
is	O
more	O
complicated	O
,	O
requiring	O
a	O
remote	O
rendering	O
display	O
or	O
service	O
,	O
the	O
use	O
of	O
Intel	O
GVT-g	O
,	O
as	O
well	O
as	O
integrating	O
the	O
VBIOS	O
into	O
the	O
boot	O
configuration	O
due	O
to	O
the	O
VBIOS	O
being	O
present	O
in	O
the	O
laptop	O
's	O
system	O
BIOS	O
as	O
opposed	O
to	O
the	O
GPU	B-Architecture
itself	O
.	O
</s>
<s>
For	O
laptops	O
that	O
have	O
a	O
GPU	B-Architecture
with	O
NVIDIA	O
Optimus	O
and	O
have	O
a	O
dedicated	O
framebuffer	O
,	O
the	O
configurations	O
may	O
vary	O
.	O
</s>
<s>
In	O
mediated	O
device	O
pass-through	O
or	O
full	O
GPU	B-General_Concept
virtualization	I-General_Concept
,	O
the	O
GPU	B-Architecture
hardware	O
provides	O
contexts	B-Operating_System
with	O
virtual	B-General_Concept
memory	I-General_Concept
ranges	I-General_Concept
for	O
each	O
guest	O
through	O
IOMMU	B-General_Concept
and	O
the	O
hypervisor	O
sends	O
graphical	O
commands	O
from	O
guests	O
directly	O
to	O
the	O
GPU	B-Architecture
.	O
</s>
<s>
This	O
technique	O
is	O
a	O
form	O
of	O
hardware-assisted	B-General_Concept
virtualization	I-General_Concept
and	O
achieves	O
near-native	O
performance	O
and	O
high	O
fidelity	B-Algorithm
.	O
</s>
<s>
If	O
the	O
hardware	O
exposes	O
contexts	B-Operating_System
as	O
full	O
logical	O
devices	O
,	O
then	O
guests	O
can	O
use	O
any	O
API	B-General_Concept
.	O
</s>
<s>
Otherwise	O
,	O
APIs	B-General_Concept
and	O
drivers	O
must	O
manage	O
the	O
additional	O
complexity	O
of	O
GPU	B-Architecture
contexts	B-Operating_System
.	O
</s>
<s>
As	O
a	O
disadvantage	O
,	O
there	O
may	O
be	O
little	O
isolation	O
between	O
virtual	B-Architecture
machines	I-Architecture
when	O
accessing	O
GPU	B-Architecture
resources	O
.	O
</s>
<s>
While	O
API	B-General_Concept
remoting	B-Operating_System
is	O
generally	O
available	O
for	O
current	O
and	O
older	O
GPUs	B-Architecture
,	O
mediated	O
pass-through	O
requires	O
hardware	O
support	O
available	O
only	O
on	O
specific	O
devices	O
.	O
</s>
<s>
GPU	B-Architecture
architectures	O
are	O
very	O
complex	O
and	O
change	O
quickly	O
,	O
and	O
their	O
internal	O
details	O
are	O
often	O
kept	O
secret	O
.	O
</s>
<s>
It	O
is	O
generally	O
not	O
feasible	O
to	O
fully	O
virtualize	O
new	O
generations	O
of	O
GPUs	B-Architecture
,	O
only	O
older	O
and	O
simpler	O
generations	O
.	O
</s>
<s>
For	O
example	O
,	O
PCem	B-Language
,	O
a	O
specialized	O
emulator	O
of	O
the	O
IBM	O
PC	O
architecture	O
,	O
can	O
emulate	O
a	O
S3	O
ViRGE/DX	O
graphics	O
device	O
,	O
which	O
supports	O
Direct3D	B-Application
3	O
,	O
and	O
a	O
3dfx	O
Voodoo2	O
,	O
which	O
supports	O
Glide	B-Operating_System
,	O
among	O
others	O
.	O
</s>
<s>
When	O
using	O
a	O
VGA	B-Protocol
or	O
an	O
SVGA	O
virtual	O
display	O
adapter	O
,	O
the	O
guest	O
may	O
not	O
have	O
3D	O
graphics	B-Architecture
acceleration	I-Architecture
,	O
providing	O
only	O
minimal	O
functionality	O
to	O
allow	O
access	O
to	O
the	O
machine	O
via	O
a	O
graphics	O
terminal	O
.	O
</s>
<s>
The	O
virtual	B-Architecture
machine	I-Architecture
manager	O
may	O
also	O
provide	O
common	O
API	B-General_Concept
implementations	O
using	O
software	O
rendering	O
to	O
enable	O
3D	O
graphics	O
applications	O
on	O
the	O
guest	O
,	O
albeit	O
at	O
speeds	O
that	O
may	O
be	O
low	O
as	O
3%	O
of	O
hardware-accelerated	B-General_Concept
native	O
performance	O
.	O
</s>
<s>
The	O
following	O
software	O
technologies	O
implement	O
graphics	O
APIs	B-General_Concept
using	O
software	O
rendering	O
:	O
</s>
