►
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
Lightning Talk: Cloud TPU? Don’t Mind If I Do - Nathan Murthy, Tesla
Application-Specific Integrated Circuits (ASICs) have unleashed a wave technological of breakthroughs that have accelerated the pace of development and performance of products up and down the stack. Machine learning libraries like Tensorflow can benefit from these new tightly-integrated hardware architectures. Once exclusive to Google’s data centers, Tensor Processing Units are now available as cloud compute resources to everyone. At Tesla these breakthroughs offer potentially new avenues for our AI/ML projects to grow.
To learn more: https://sched.co/GrVP
A
So,
hey
everyone
I'm
here
to
assure
you,
if
you
stick
around
for
Kendall's
talk,
that
we
will
still
be
employing
human
beings
for
the
next
five
to
ten
years.
So
trust
me
trust
me
and
if
you
guys
want,
we
can
have
some
sidebars
afterwards.
I
can
tell
you
all
the
great
stories
of
human
being
scrubbing.
You
know,
tape
from
the
factory
and
everything
there's
definitely
no
robots
for
those
that
we
can
program
yet
I'm,
hey
everyone.
My
name
is
Nathan
Nathan
Murthy.
A
The
tile
of
my
talk
is
called
cloud
CP,
you
don't
mind
if
I
do
feel
free
to
take
a
picture
of
the
first
slide.
You
wanna
stay
in
touch
after
the
talk
through
Twitter.
So
those
of
you
who
don't
know
Tesla
not
only
builds
cars
electric
vehicles,
but
we
also
have
an
energy
component
to
our
business,
which
is
another
core
foundation
for
what
we
do.
A
One
of
the
big
pieces
of
energy
technology
I
got
the
opportunity
to
work
on
was
the
horns:
devil
power
reserve
in
Australia,
which
today
is
the
largest
lithium-ion
battery
pack
in
the
world.
In
the
last
year,
we
were
able
to
save
Australian
electricity
customers,
40
million
dollars
by
integrating
this
battery
system
with
wind
power
sighted
at
this
power
reserve,
and
what
the
opportunity
for
us
was
the
ability
to
provide
this
battery,
it's
kind
of
like
a
shock
absorber
for
the
grid
to
kind
of
hedge
electricity
customers
from
these
large
spikes.
A
In
order
to
do
this
efficiently
with
our
battery
pack,
we
need
to
be
able
to
predict
these
sort
of
high-priced
events
and
to
do
that,
we
sort
of
build
a
built
out
a
component
that
allowed
us
to
be
able
to
make
intelligent
decisions
about
the
state
of
the
system
and
act
on
that
state,
using
what
we
call
Tesla
auto
bidder
for
the
energy
component.
Now
for
those
of
you
who
are
familiar
Tesla,
how
many
people
know
about
autopilot?
A
Okay,
Google,
so
roughly
a
50%,
so
auto
bidder
is
roughly
the
analog
of
autopilot
for
the
electricity
grid.
I
won't
go
to
too
much
detail,
but
this
auto
bidding
component
if
I
were
to
break
down
into
those
spikes
of
energy
that
you
saw
essentially
sends
electricity
bids
across
12
or
9
to
12
different
products
in
the
electricity
market.
Every
five
minutes
you'll
see
that
there
is
a
pricing
forecast
step,
an
optimization
step
before
roughly
two
minutes
before
you
even
send
a
bid.
A
A
Tensorflow
is
a
very
popular
tool
that
we've
been
using
and
what's
great
about
tensorflow
as
an
open-source
tool.
Is
that
if
you
run
it
in
traditional
hardware,
you
can
still
take
advantage
of
CPUs
and
GPUs,
but
if
you
run
it
on
Google
cloud,
you
get
to
take
advantage
of
their
TPU
beta
product.
So
I've
got
a
quick
demo
about.
What's
so
great
about
CPUs
I.
A
A
Still
dialing
still
ty,
so
this
is
obviously
what
I
was
gonna
hope
to
show
actually
is
a
kubernetes
cluster
as
well
running
if
I
can
get
that
up,
but
I
don't
quite
have
okay.
Well,
let's
just
skip
that.
So
I
was
able
to
run
a
couple
simulations
of
this
while
I
was
prepping.
This
presentation,
and
one
of
the
great
things
about
these
TPU
as
I
said,
is
its
performance
density.
Now,
a
traditional
Pascal
architecture
and
video
chips
gives
you
about
roughly
seven
to
ten
teraflops
per
second
here's,
a
quick
benchmark.
A
I
ran
where
you
get
about
165
teraflops
per
second,
which
okay,
that's
great
and
everything,
but
the
reliability
component
I
think
is
something
that's
still
missing
from
this
product.
But
what
I
find
very
useful
about
this
or
what
I
find,
is
a
trend
in
some
of
these
software
components
that
we're
seeing.
Is
this
tendency
for
more
and
more
components
of
the
system
to
be
made
available
to
people
using
public
pop
providers?
There's
a
rough
architecture
of
what
cloud
CPU
offers
you
can
see.
G
RPC
there
another
a
maje
to
our
cloud
native
theme
here:
Google!
A
Isn't
the
only
company
making
a
six
right
so
Tesla
is
making
a
six,
and
on
top
of
that,
oh
snap
I
have
some
missing
slides.
Sorry
about
that
and
they're,
not
loading.
It's
all
good.
The
Internet
is
not
working
for
me
today.
What
I
will
say
is
that
there
is
a
tendency
now
there's
a
trend
for
companies
now
to
offer
application-specific
chips
and
gradually
over
time.
A
What
we'll
see
is
more
and
more
chips
that
are
application
specifically
provided
to
the
cloud
and
I
think
it's
a
very
exciting
moment
for
us
to
be
able
to
leverage
some
of
the
technologies
that
we
have
that
were
able
to
build
on
the
public
cloud
of
today
and
of
tomorrow.
So
thank
you
for
everyone
for
sitting
in
and
rest
assured,
I'll
make
sure
the
rest
assured
human
beings
will
still
be
employed
by
the
end
of
the
decade,
at
least.