<s>
Tensor	B-Device
Processing	I-Device
Unit	I-Device
(	O
TPU	O
)	O
is	O
an	O
AI	B-General_Concept
accelerator	I-General_Concept
application-specific	O
integrated	O
circuit	O
(	O
ASIC	O
)	O
developed	O
by	O
Google	B-Application
for	O
neural	B-Architecture
network	I-Architecture
machine	O
learning	O
,	O
using	O
Google	B-Application
's	I-Application
own	O
TensorFlow	B-Language
software	O
.	O
</s>
<s>
Google	B-Application
began	O
using	O
TPUs	O
internally	O
in	O
2015	O
,	O
and	O
in	O
2018	O
made	O
them	O
available	O
for	O
third	O
party	O
use	O
,	O
both	O
as	O
part	O
of	O
its	O
cloud	O
infrastructure	O
and	O
by	O
offering	O
a	O
smaller	O
version	O
of	O
the	O
chip	O
for	O
sale	O
.	O
</s>
<s>
Compared	O
to	O
a	O
graphics	B-Architecture
processing	I-Architecture
unit	I-Architecture
,	O
TPUs	O
are	O
designed	O
for	O
a	O
high	O
volume	O
of	O
low	O
precision	O
computation	O
(	O
e.g.	O
</s>
<s>
as	O
little	O
as	O
8-bit	O
precision	O
)	O
with	O
more	O
input/output	O
operations	O
per	O
joule	B-Language
,	O
without	O
hardware	O
for	O
rasterisation/texture	O
mapping	O
.	O
</s>
<s>
The	O
TPU	O
ASICs	O
are	O
mounted	O
in	O
a	O
heatsink	O
assembly	O
,	O
which	O
can	O
fit	O
in	O
a	O
hard	O
drive	O
slot	O
within	O
a	O
data	O
center	O
rack	B-Application
,	O
according	O
to	O
Norman	O
Jouppi	O
.	O
</s>
<s>
TPUs	O
are	O
well	O
suited	O
for	O
CNNs	B-Architecture
,	O
while	O
GPUs	B-Architecture
have	O
benefits	O
for	O
some	O
fully-connected	O
neural	B-Architecture
networks	I-Architecture
,	O
and	O
CPUs	B-General_Concept
can	O
have	O
advantages	O
for	O
RNNs	B-Algorithm
.	O
</s>
<s>
The	O
tensor	B-Device
processing	I-Device
unit	I-Device
was	O
announced	O
in	O
May	O
2016	O
at	O
Google	B-Application
I/O	I-Application
,	O
when	O
the	O
company	O
said	O
that	O
the	O
TPU	O
had	O
already	O
been	O
used	O
inside	O
their	B-Architecture
data	I-Architecture
centers	I-Architecture
for	O
over	O
a	O
year	O
.	O
</s>
<s>
The	O
chip	O
has	O
been	O
specifically	O
designed	O
for	O
Google	B-Application
's	I-Application
TensorFlow	B-Language
framework	O
,	O
a	O
symbolic	O
math	O
library	O
which	O
is	O
used	O
for	O
machine	O
learning	O
applications	O
such	O
as	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
However	O
,	O
as	O
of	O
2017	O
Google	B-Application
still	O
used	O
CPUs	B-General_Concept
and	O
GPUs	B-Architecture
for	O
other	O
types	O
of	O
machine	O
learning	O
.	O
</s>
<s>
Other	O
AI	B-General_Concept
accelerator	I-General_Concept
designs	O
are	O
appearing	O
from	O
other	O
vendors	O
also	O
and	O
are	O
aimed	O
at	O
embedded	B-Architecture
and	O
robotics	O
markets	O
.	O
</s>
<s>
Google	B-Application
's	I-Application
TPUs	O
are	O
proprietary	O
.	O
</s>
<s>
Some	O
models	O
are	O
commercially	O
available	O
,	O
and	O
on	O
February	O
12	O
,	O
2018	O
,	O
The	O
New	O
York	O
Times	O
reported	O
that	O
Google	B-Application
"	O
would	O
allow	O
other	O
companies	O
to	O
buy	O
access	O
to	O
those	O
chips	O
through	O
its	O
cloud-computing	O
service.	O
"	O
</s>
<s>
Google	B-Application
has	O
said	O
that	O
they	O
were	O
used	O
in	O
the	O
AlphaGo	B-Application
versus	I-Application
Lee	I-Application
Sedol	I-Application
series	O
of	O
man-machine	O
Go	O
games	O
,	O
as	O
well	O
as	O
in	O
the	O
AlphaZero	B-Application
system	O
,	O
which	O
produced	O
Chess	B-Application
,	O
Shogi	O
and	O
Go	O
playing	O
programs	O
from	O
the	O
game	O
rules	O
alone	O
and	O
went	O
on	O
to	O
beat	O
the	O
leading	O
programs	O
in	O
those	O
games	O
.	O
</s>
<s>
Google	B-Application
has	O
also	O
used	O
TPUs	O
for	O
Google	B-Application
Street	I-Application
View	I-Application
text	O
processing	O
and	O
was	O
able	O
to	O
find	O
all	O
the	O
text	O
in	O
the	O
Street	B-Application
View	I-Application
database	O
in	O
less	O
than	O
five	O
days	O
.	O
</s>
<s>
In	O
Google	B-Algorithm
Photos	I-Algorithm
,	O
an	O
individual	O
TPU	O
can	O
process	O
over	O
100	O
million	O
photos	O
a	O
day	O
.	O
</s>
<s>
It	O
is	O
also	O
used	O
in	O
RankBrain	B-Application
which	O
Google	B-Application
uses	O
to	O
provide	O
search	O
results	O
.	O
</s>
<s>
Google	B-Application
provides	O
third	O
parties	O
access	O
to	O
TPUs	O
through	O
its	O
Cloud	O
TPU	O
service	O
as	O
part	O
of	O
the	B-Application
Google	I-Application
Cloud	O
Platform	O
and	O
through	O
its	O
notebook-based	O
services	O
Kaggle	B-Application
and	O
Colaboratory	O
.	O
</s>
<s>
The	O
clock	O
speed	O
is	O
700MHz	O
and	O
it	O
has	O
a	O
thermal	B-General_Concept
design	I-General_Concept
power	I-General_Concept
of	O
28	O
–	O
40W	O
.	O
</s>
<s>
It	O
has	O
28MiB	O
of	O
on	O
chip	O
memory	O
,	O
and	O
4MiB	O
of	O
32-bit	O
accumulators	B-General_Concept
taking	O
the	O
results	O
of	O
a	O
256×256	O
systolic	B-Architecture
array	I-Architecture
of	O
8-bit	O
multipliers	O
.	O
</s>
<s>
Within	O
the	O
TPU	O
package	O
is	O
8GiB	O
of	O
dual-channel	B-Architecture
2133MHz	O
DDR3	O
SDRAM	O
offering	O
34GB/s	O
of	O
bandwidth	O
.	O
</s>
<s>
Instructions	O
transfer	O
data	O
to	O
or	O
from	O
the	O
host	O
,	O
perform	O
matrix	O
multiplications	O
or	O
convolutions	B-Language
,	O
and	O
apply	O
activation	B-Algorithm
functions	I-Algorithm
.	O
</s>
<s>
Google	B-Application
stated	O
the	O
first-generation	O
TPU	O
design	O
was	O
limited	O
by	O
memory	B-General_Concept
bandwidth	I-General_Concept
and	O
using	O
16	O
GB	O
of	O
High	O
Bandwidth	O
Memory	O
in	O
the	O
second-generation	O
design	O
increased	O
bandwidth	O
to	O
600GB/s	O
and	O
performance	O
to	O
45teraFLOPS	O
.	O
</s>
<s>
Notably	O
,	O
while	O
the	O
first-generation	O
TPUs	O
were	O
limited	O
to	O
integers	O
,	O
the	O
second-generation	O
TPUs	O
can	O
also	O
calculate	O
in	O
floating	B-Algorithm
point	I-Algorithm
.	O
</s>
<s>
Google	B-Application
has	O
stated	O
these	O
second-generation	O
TPUs	O
will	O
be	O
available	O
on	O
the	B-Application
Google	I-Application
Compute	B-General_Concept
Engine	I-General_Concept
for	O
use	O
in	O
TensorFlow	B-Language
applications	O
.	O
</s>
<s>
Google	B-Application
announced	O
that	O
processors	O
themselves	O
are	O
twice	O
as	O
powerful	O
as	O
the	O
second-generation	O
TPUs	O
,	O
and	O
would	O
be	O
deployed	O
in	O
pods	O
with	O
four	O
times	O
as	O
many	O
chips	O
as	O
the	O
preceding	O
generation	O
.	O
</s>
<s>
On	O
May	O
18	O
,	O
2021	O
,	O
Google	B-Application
CEO	O
Sundar	O
Pichai	O
spoke	O
about	O
TPU	O
v4	O
Tensor	B-Device
Processing	I-Device
Units	I-Device
during	O
his	O
keynote	O
at	O
the	B-Application
Google	I-Application
I/O	O
virtual	O
conference	O
.	O
</s>
<s>
In	O
July	O
2018	O
,	O
Google	B-Application
announced	O
the	O
Edge	B-Device
TPU	O
.	O
</s>
<s>
The	O
Edge	B-Device
TPU	O
is	O
Google	B-Application
's	I-Application
purpose-built	O
ASIC	O
chip	O
designed	O
to	O
run	O
machine	O
learning	O
(	O
ML	O
)	O
models	O
for	O
edge	B-Device
computing	I-Device
,	O
meaning	O
it	O
is	O
much	O
smaller	O
and	O
consumes	O
far	O
less	O
power	O
compared	O
to	O
the	O
TPUs	O
hosted	O
in	O
Google	B-Application
datacenters	O
(	O
also	O
known	O
as	O
Cloud	O
TPUs	O
)	O
.	O
</s>
<s>
In	O
January	O
2019	O
,	O
Google	B-Application
made	O
the	O
Edge	B-Device
TPU	O
available	O
to	O
developers	O
with	O
a	O
line	O
of	O
products	O
under	O
the	O
Coral	O
brand	O
.	O
</s>
<s>
The	O
Edge	B-Device
TPU	O
is	O
capable	O
of	O
4	O
trillion	O
operations	O
per	O
second	O
with	O
2	O
W	O
of	O
electrical	O
power	O
.	O
</s>
<s>
The	O
product	O
offerings	O
include	O
a	O
single-board	B-Device
computer	I-Device
(	O
SBC	B-Device
)	O
,	O
a	O
system	B-Architecture
on	I-Architecture
module	I-Architecture
(	O
SoM	O
)	O
,	O
a	O
USB	B-Protocol
accessory	O
,	O
a	O
mini	O
PCI-e	O
card	O
,	O
and	O
an	O
M.2	B-Protocol
card	O
.	O
</s>
<s>
The	O
SBC	B-Device
Coral	O
Dev	O
Board	O
and	O
Coral	O
SoM	O
both	O
run	O
Mendel	O
Linux	O
OS	O
–	O
a	O
derivative	O
of	O
Debian	O
.	O
</s>
<s>
The	O
USB	B-Protocol
,	O
PCI-e	O
,	O
and	O
M.2	B-Protocol
products	O
function	O
as	O
add-ons	O
to	O
existing	O
computer	O
systems	O
,	O
and	O
support	O
Debian-based	O
Linux	O
systems	O
on	O
x86-64	O
and	O
ARM64	O
hosts	O
(	O
including	O
Raspberry	B-Operating_System
Pi	I-Operating_System
)	O
.	O
</s>
<s>
The	O
machine	O
learning	O
runtime	O
used	O
to	O
execute	O
models	O
on	O
the	O
Edge	B-Device
TPU	O
is	O
based	O
on	O
TensorFlow	B-Language
Lite	I-Language
.	O
</s>
<s>
The	O
Edge	B-Device
TPU	O
is	O
only	O
capable	O
of	O
accelerating	O
forward-pass	O
operations	O
,	O
which	O
means	O
it	O
's	O
primarily	O
useful	O
for	O
performing	O
inferences	O
(	O
although	O
it	O
is	O
possible	O
to	O
perform	O
lightweight	O
transfer	O
learning	O
on	O
the	O
Edge	B-Device
TPU	O
)	O
.	O
</s>
<s>
The	O
Edge	B-Device
TPU	O
also	O
only	O
supports	O
8-bit	O
math	O
,	O
meaning	O
that	O
for	O
a	O
network	O
to	O
be	O
compatible	O
with	O
the	O
Edge	B-Device
TPU	O
,	O
it	O
needs	O
to	O
either	O
be	O
trained	O
using	O
the	O
TensorFlow	B-Language
quantization-aware	O
training	O
technique	O
,	O
or	O
since	O
late	O
2019	O
it	O
's	O
also	O
possible	O
to	O
use	O
post-training	O
quantization	O
.	O
</s>
<s>
On	O
November	O
12	O
,	O
2019	O
,	O
Asus	O
announced	O
a	O
pair	O
of	O
single-board	B-Device
computer	I-Device
(	O
SBCs	O
)	O
featuring	O
the	O
Edge	B-Device
TPU	O
.	O
</s>
<s>
The	O
Asus	B-Device
Tinker	I-Device
Edge	I-Device
T	I-Device
and	I-Device
Tinker	I-Device
Edge	I-Device
R	I-Device
Board	I-Device
designed	O
for	O
IoT	B-Operating_System
and	O
edge	B-Device
AI	B-Application
.	O
</s>
<s>
The	O
SBCs	O
officially	O
support	O
Android	B-Application
and	O
Debian	O
operating	B-General_Concept
systems	I-General_Concept
.	O
</s>
<s>
ASUS	O
has	O
also	O
demonstrated	O
a	O
mini	O
PC	O
called	O
Asus	O
PN60T	O
featuring	O
the	O
Edge	B-Device
TPU	O
.	O
</s>
<s>
On	O
January	O
2	O
,	O
2020	O
,	O
Google	B-Application
announced	O
the	O
Coral	O
Accelerator	O
Module	O
and	O
Coral	O
Dev	O
Board	O
Mini	O
,	O
to	O
be	O
demonstrated	O
at	O
CES	O
2020	O
later	O
the	O
same	O
month	O
.	O
</s>
<s>
The	O
Coral	O
Accelerator	O
Module	O
is	O
a	O
multi-chip	B-Algorithm
module	I-Algorithm
featuring	O
the	O
Edge	B-Device
TPU	O
,	O
PCIe	O
and	O
USB	B-Protocol
interfaces	O
for	O
easier	O
integration	O
.	O
</s>
<s>
The	O
Coral	O
Dev	O
Board	O
Mini	O
is	O
a	O
smaller	O
SBC	B-Device
featuring	O
the	O
Coral	O
Accelerator	O
Module	O
and	O
MediaTek	B-Architecture
8167s	I-Architecture
SoC	I-Architecture
.	O
</s>
<s>
On	O
October	O
15	O
,	O
2019	O
,	O
Google	B-Application
announced	O
the	O
Pixel	B-Device
4	I-Device
smartphone	O
,	O
which	O
contains	O
an	O
Edge	B-Device
TPU	O
called	O
the	O
Pixel	O
Neural	O
Core	O
.	O
</s>
<s>
Google	B-Application
describe	O
it	O
as	O
"	O
customized	O
to	O
meet	O
the	O
requirements	O
of	O
key	O
camera	O
features	O
in	O
Pixel	B-Device
4	I-Device
"	O
,	O
using	O
a	O
neural	B-Architecture
network	I-Architecture
search	O
that	O
sacrifices	O
some	O
accuracy	O
in	O
favor	O
of	O
minimizing	O
latency	O
and	O
power	O
use	O
.	O
</s>
<s>
Google	B-Application
followed	O
the	O
Pixel	O
Neural	O
Core	O
by	O
integrating	O
an	O
Edge	B-Device
TPU	O
into	O
a	O
custom	O
system-on-chip	O
named	O
Google	B-Application
Tensor	B-Device
,	O
which	O
was	O
released	O
in	O
2021	O
with	O
the	O
Pixel	B-Device
6	I-Device
line	O
of	O
smartphones	O
.	O
</s>
<s>
The	B-Application
Google	I-Application
Tensor	B-Device
SoC	O
demonstrated	O
"	O
extremely	O
large	O
performance	O
advantages	O
over	O
the	O
competition	O
"	O
in	O
machine	O
learning-focused	O
benchmarks	O
;	O
although	O
instantaneous	O
power	O
consumption	O
also	O
was	O
relatively	O
high	O
,	O
the	O
improved	O
performance	O
meant	O
less	O
energy	O
was	O
consumed	O
due	O
to	O
shorter	O
periods	O
requiring	O
peak	O
performance	O
.	O
</s>
