<s>
An	O
AI	B-General_Concept
accelerator	I-General_Concept
is	O
a	O
class	O
of	O
specialized	O
hardware	B-General_Concept
accelerator	I-General_Concept
or	O
computer	O
system	O
designed	O
to	O
accelerate	B-General_Concept
artificial	B-Application
intelligence	I-Application
and	O
machine	O
learning	O
applications	O
,	O
including	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
and	O
machine	B-General_Concept
vision	I-General_Concept
.	O
</s>
<s>
Typical	O
applications	O
include	O
algorithms	O
for	O
robotics	O
,	O
Internet	B-Operating_System
of	I-Operating_System
Things	I-Operating_System
,	O
and	O
other	O
data-intensive	O
or	O
sensor-driven	O
tasks	O
.	O
</s>
<s>
They	O
are	O
often	O
manycore	B-General_Concept
designs	O
and	O
generally	O
focus	O
on	O
low-precision	B-General_Concept
arithmetic	O
,	O
novel	O
dataflow	B-General_Concept
architectures	I-General_Concept
or	O
in-memory	B-General_Concept
computing	I-General_Concept
capability	O
.	O
</s>
<s>
,	O
a	O
typical	O
AI	B-Application
integrated	O
circuit	O
chip	O
contains	O
billions	O
of	O
MOSFET	B-Architecture
transistors	O
.	O
</s>
<s>
Computer	O
systems	O
have	O
frequently	O
complemented	O
the	O
CPU	B-General_Concept
with	O
special-purpose	O
accelerators	O
for	O
specialized	O
tasks	O
,	O
known	O
as	O
coprocessors	B-General_Concept
.	O
</s>
<s>
Notable	O
application-specific	O
hardware	B-Device
units	I-Device
include	O
video	B-Device
cards	I-Device
for	O
graphics	O
,	O
sound	B-Device
cards	I-Device
,	O
graphics	B-Architecture
processing	I-Architecture
units	I-Architecture
and	O
digital	B-Architecture
signal	I-Architecture
processors	I-Architecture
.	O
</s>
<s>
As	O
deep	B-Algorithm
learning	I-Algorithm
and	O
artificial	B-Application
intelligence	I-Application
workloads	O
rose	O
in	O
prominence	O
in	O
the	O
2010s	O
,	O
specialized	O
hardware	B-Device
units	I-Device
were	O
developed	O
or	O
adapted	O
from	O
existing	O
products	O
to	O
accelerate	B-General_Concept
these	O
tasks	O
.	O
</s>
<s>
Benchmarks	O
such	O
as	O
MLPerf	B-Algorithm
may	O
be	O
used	O
to	O
evaluate	O
the	O
performance	O
of	O
AI	B-General_Concept
accelerators	I-General_Concept
.	O
</s>
<s>
Later	O
all-digital	O
chips	O
like	O
the	O
Nestor/Intel	O
Ni1000	B-Algorithm
followed	O
.	O
</s>
<s>
As	O
early	O
as	O
1993	O
,	O
digital	B-Architecture
signal	I-Architecture
processors	I-Architecture
were	O
used	O
as	O
neural	B-General_Concept
network	I-General_Concept
accelerators	I-General_Concept
to	O
accelerate	B-General_Concept
optical	B-Application
character	I-Application
recognition	I-Application
software	I-Application
.	O
</s>
<s>
In	O
the	O
1990s	O
,	O
there	O
were	O
also	O
attempts	O
to	O
create	O
parallel	O
high-throughput	O
systems	O
for	O
workstations	O
aimed	O
at	O
various	O
applications	O
,	O
including	O
neural	B-Architecture
network	I-Architecture
simulations	O
.	O
</s>
<s>
FPGA-based	O
accelerators	O
were	O
also	O
first	O
explored	O
in	O
the	O
1990s	O
for	O
both	O
inference	O
and	O
training	O
.	O
</s>
<s>
Smartphones	B-Application
began	O
incorporating	O
AI	B-General_Concept
accelerators	I-General_Concept
starting	O
with	O
the	O
Qualcomm	O
Snapdragon	O
820	O
in	O
2015	O
.	O
</s>
<s>
Architectures	O
such	O
as	O
the	O
Cell	B-General_Concept
microprocessor	I-General_Concept
have	O
features	O
significantly	O
overlapping	O
with	O
AI	B-General_Concept
accelerators	I-General_Concept
including	O
:	O
support	O
for	O
packed	O
low	O
precision	B-General_Concept
arithmetic	O
,	O
dataflow	B-General_Concept
architecture	I-General_Concept
,	O
and	O
prioritizing	O
'	O
throughput	O
 '	O
over	O
latency	O
.	O
</s>
<s>
The	O
Cell	B-General_Concept
microprocessor	I-General_Concept
was	O
subsequently	O
applied	O
to	O
a	O
number	O
of	O
tasks	O
including	O
AI	B-Application
.	O
</s>
<s>
In	O
the	O
2000s	O
,	O
CPUs	O
also	O
gained	O
increasingly	O
wide	O
SIMD	B-Device
units	O
,	O
driven	O
by	O
video	O
and	O
gaming	O
workloads	O
;	O
as	O
well	O
as	O
support	O
for	O
packed	O
low-precision	B-General_Concept
data	B-General_Concept
types	O
.	O
</s>
<s>
Due	O
to	O
increasing	O
performance	O
of	O
CPUs	O
,	O
they	O
are	O
also	O
being	O
used	O
for	O
running	O
AI	B-Application
workloads	O
.	O
</s>
<s>
Graphics	B-Architecture
processing	I-Architecture
units	I-Architecture
or	O
GPUs	B-Architecture
are	O
specialized	O
hardware	O
for	O
the	O
manipulation	O
of	O
images	O
and	O
calculation	O
of	O
local	O
image	O
properties	O
.	O
</s>
<s>
The	O
mathematical	O
basis	O
of	O
neural	B-Architecture
networks	I-Architecture
and	O
image	B-General_Concept
manipulation	I-General_Concept
are	O
similar	O
,	O
embarrassingly	B-Operating_System
parallel	I-Operating_System
tasks	O
involving	O
matrices	O
,	O
leading	O
GPUs	B-Architecture
to	O
become	O
increasingly	O
used	O
for	O
machine	O
learning	O
tasks	O
.	O
</s>
<s>
,	O
GPUs	B-Architecture
are	O
popular	O
for	O
AI	B-Application
work	O
,	O
and	O
they	O
continue	O
to	O
evolve	O
in	O
a	O
direction	O
to	O
facilitate	O
deep	B-Algorithm
learning	I-Algorithm
,	O
both	O
for	O
training	O
and	O
inference	O
in	O
devices	O
such	O
as	O
self-driving	O
cars	O
.	O
</s>
<s>
GPU	B-Architecture
developers	O
such	O
as	O
Nvidia	O
NVLink	O
are	O
developing	O
additional	O
connective	O
capability	O
for	O
the	O
kind	O
of	O
dataflow	O
workloads	O
AI	B-Application
benefits	O
from	O
.	O
</s>
<s>
As	O
GPUs	B-Architecture
have	O
been	O
increasingly	O
applied	O
to	O
AI	B-Application
acceleration	O
,	O
GPU	B-Architecture
manufacturers	O
have	O
incorporated	O
neural	O
network-specific	O
hardware	O
to	O
further	O
accelerate	B-General_Concept
these	O
tasks	O
.	O
</s>
<s>
Tensor	O
cores	O
are	O
intended	O
to	O
speed	O
up	O
the	O
training	O
of	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
frameworks	O
are	O
still	O
evolving	O
,	O
making	O
it	O
hard	O
to	O
design	O
custom	O
hardware	O
.	O
</s>
<s>
Reconfigurable	B-Architecture
devices	O
such	O
as	O
field-programmable	B-Architecture
gate	I-Architecture
arrays	I-Architecture
(	O
FPGA	B-Architecture
)	O
make	O
it	O
easier	O
to	O
evolve	O
hardware	O
,	O
frameworks	O
,	O
and	O
software	O
alongside	O
each	O
other	O
.	O
</s>
<s>
Microsoft	O
has	O
used	O
FPGA	B-Architecture
chips	O
to	O
accelerate	B-General_Concept
inference	O
.	O
</s>
<s>
While	O
GPUs	B-Architecture
and	O
FPGAs	B-Architecture
perform	O
far	O
better	O
than	O
CPUs	O
for	O
AI-related	O
tasks	O
,	O
a	O
factor	O
of	O
up	O
to	O
10	O
in	O
efficiency	O
may	O
be	O
gained	O
with	O
a	O
more	O
specific	O
design	O
,	O
via	O
an	O
application-specific	O
integrated	O
circuit	O
(	O
ASIC	O
)	O
.	O
</s>
<s>
These	O
accelerators	O
employ	O
strategies	O
such	O
as	O
optimized	O
memory	O
use	O
and	O
the	O
use	O
of	O
lower	B-Algorithm
precision	I-Algorithm
arithmetic	I-Algorithm
to	O
accelerate	B-General_Concept
calculation	O
and	O
increase	O
throughput	O
of	O
computation	O
.	O
</s>
<s>
Some	O
adopted	O
low-precision	B-General_Concept
floating-point	B-Algorithm
formats	I-Algorithm
used	O
AI	B-Application
acceleration	O
are	O
half-precision	O
and	O
the	O
bfloat16	O
floating-point	B-Algorithm
format	I-Algorithm
.	O
</s>
<s>
Companies	O
such	O
as	O
Google	O
,	O
Qualcomm	O
,	O
Amazon	O
,	O
Apple	O
,	O
Facebook	O
,	O
AMD	O
and	O
Samsung	O
are	O
all	O
designing	O
their	O
own	O
AI	B-Application
ASICs	O
.	O
</s>
<s>
Cerebras	O
Systems	O
has	O
also	O
built	O
a	O
dedicated	O
AI	B-General_Concept
accelerator	I-General_Concept
based	O
on	O
the	O
largest	O
processor	O
in	O
the	O
industry	O
,	O
the	O
second-generation	O
Wafer	O
Scale	O
Engine	O
(	O
WSE-2	O
)	O
,	O
to	O
support	O
deep	B-Algorithm
learning	I-Algorithm
workloads	O
.	O
</s>
<s>
In	O
June	O
2017	O
,	O
IBM	O
researchers	O
announced	O
an	O
architecture	O
in	O
contrast	O
to	O
the	O
Von	B-Architecture
Neumann	I-Architecture
architecture	I-Architecture
based	O
on	O
in-memory	B-General_Concept
computing	I-General_Concept
and	O
phase-change	B-Device
memory	I-Device
arrays	O
applied	O
to	O
temporal	O
correlation	O
detection	O
,	O
intending	O
to	O
generalize	O
the	O
approach	O
to	O
heterogeneous	O
computing	O
and	O
massively	B-Operating_System
parallel	I-Operating_System
systems	O
.	O
</s>
<s>
In	O
October	O
2018	O
,	O
IBM	O
researchers	O
announced	O
an	O
architecture	O
based	O
on	O
in-memory	B-General_Concept
processing	I-General_Concept
and	O
modeled	O
on	O
the	O
human	O
brain	O
's	O
synaptic	O
network	O
to	O
accelerate	B-General_Concept
deep	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
The	O
system	O
is	O
based	O
on	O
phase-change	B-Device
memory	I-Device
arrays	O
.	O
</s>
<s>
Their	O
algorithm	O
is	O
based	O
on	O
in-memory	B-General_Concept
computing	I-General_Concept
with	O
analog	O
resistive	O
memories	O
which	O
performs	O
with	O
high	O
efficiencies	O
of	O
time	O
and	O
energy	O
,	O
via	O
conducting	O
matrix	O
–	O
vector	O
multiplication	O
in	O
one	O
step	O
using	O
Ohm	O
's	O
law	O
and	O
Kirchhoff	O
's	O
law	O
.	O
</s>
<s>
published	O
experiments	O
with	O
a	O
large-area	O
active	O
channel	O
material	O
for	O
developing	O
logic-in-memory	O
devices	O
and	O
circuits	O
based	O
on	O
floating-gate	B-Algorithm
field-effect	O
transistors	O
(	O
FGFETs	O
)	O
.	O
</s>
<s>
Such	O
atomically	O
thin	O
semiconductors	O
are	O
considered	O
promising	O
for	O
energy-efficient	O
machine	O
learning	O
applications	O
,	O
where	O
the	O
same	O
basic	O
device	O
structure	O
is	O
used	O
for	O
both	O
logic	O
operations	O
and	O
data	B-General_Concept
storage	O
.	O
</s>
<s>
proposed	O
an	O
integrated	O
photonic	O
hardware	B-General_Concept
accelerator	I-General_Concept
for	O
parallel	O
convolutional	O
processing	O
.	O
</s>
<s>
The	O
authors	O
identify	O
two	O
key	O
advantages	O
of	O
integrated	O
photonics	O
over	O
its	O
electronic	O
counterparts	O
:	O
(	O
1	O
)	O
massively	B-Operating_System
parallel	I-Operating_System
data	B-General_Concept
transfer	O
through	O
wavelength	O
division	O
multiplexing	B-Architecture
in	O
conjunction	O
with	O
frequency	O
combs	O
,	O
and	O
(	O
2	O
)	O
extremely	O
high	O
data	B-General_Concept
modulation	O
speeds	O
.	O
</s>
<s>
Their	O
system	O
can	O
execute	O
trillions	O
of	O
multiply-accumulate	O
operations	O
per	O
second	O
,	O
indicating	O
the	O
potential	O
of	O
integrated	O
photonics	O
in	O
data-heavy	O
AI	B-Application
applications	O
.	O
</s>
<s>
As	O
of	O
2016	O
,	O
the	O
field	O
is	O
still	O
in	O
flux	O
and	O
vendors	O
are	O
pushing	O
their	O
own	O
marketing	O
term	O
for	O
what	O
amounts	O
to	O
an	O
"	O
AI	B-General_Concept
accelerator	I-General_Concept
"	O
,	O
in	O
the	O
hope	O
that	O
their	O
designs	O
and	O
APIs	B-Application
will	O
become	O
the	O
dominant	O
design	O
.	O
</s>
<s>
In	O
the	O
past	O
when	O
consumer	O
graphics	B-Architecture
accelerators	I-Architecture
emerged	O
,	O
the	O
industry	O
eventually	O
adopted	O
Nvidia	O
's	O
self-assigned	O
term	O
,	O
"	O
the	O
GPU	B-Architecture
"	O
,	O
as	O
the	O
collective	O
noun	O
for	O
"	O
graphics	B-Architecture
accelerators	I-Architecture
"	O
,	O
which	O
had	O
taken	O
many	O
forms	O
before	O
settling	O
on	O
an	O
overall	O
pipeline	B-General_Concept
implementing	O
a	O
model	O
presented	O
by	O
Direct3D	B-Application
.	O
</s>
<s>
Search	B-Application
engines	I-Application
,	O
increasing	O
the	O
energy	O
efficiency	O
of	O
data	B-Operating_System
centers	I-Operating_System
and	O
ability	O
to	O
use	O
increasingly	O
advanced	O
queries	B-Library
.	O
</s>
<s>
the	O
Movidius	B-General_Concept
Myriad	I-General_Concept
2	I-General_Concept
has	O
been	O
demonstrated	O
successfully	O
guiding	O
autonomous	O
drones	O
.	O
</s>
<s>
Voice	B-Application
user	I-Application
interface	I-Application
,	O
e.g.	O
</s>
<s>
in	O
mobile	O
phones	O
,	O
a	O
target	O
for	O
Qualcomm	B-General_Concept
Zeroth	I-General_Concept
.	O
</s>
