►
From YouTube: KBE Insider Detroit - Ronen Dar, Run:ai
Description
We take KBE Insider on the road while in Detroit for KubeCon North America and interview Ronen Dar, Co-founder and CTO at Run:AI. We’ll hear from Ronen the challenges of running AI workloads and how Run:AI is accelerating the AI era with their cloud-native compute orchestration platform, making AI accessible and easy for everyone! Tune in and learn more about Kubernetes and AI development.
KBE Insider is hosted by Langdon White, Clinical Assistant Professor of Computing and Data Sciences at Boston University. KBE Insider is presented to you by Kube by Example (KBE) and lets you reach people deeply involved with the Kubernetes community and engage with Kubernetes contributors from across the globe.
Get started for free with our KBE AI/ML learning path: https://kubebyexample.com/learning-paths/aiml-jupyter-kubernetes-jupyterhub
A
All
right,
so
thanks
for
coming
running
and
you
know
from
run
AI,
you
know
we
really
appreciate
you
kind
of
joining
us
in
our
little
car
ride.
Hopefully
you'll
you
haven't
been
to
Belle
Isle,
so
you
can
see
that
it's
like
one
of
the
famous
points
in
Detroit
and
really
pretty
I.
Think
but
yeah
thanks
for
coming
yeah.
B
So
we're
working
with
a
lot
of
advanced
institutions
and
we
essentially
help
them
centralize
their
their
AI
infrastructure,
gotcha
and
and
allow
them
to
make
better
use
of
gpus
and
their
Computing
resources
and
essentially
provide
easy
and
simple
access
to
computing
power
to
their
researchers.
Data
scientists,
AI
Engineers,
yeah.
A
A
B
A
B
But
what's
Unix
with
us
that
we
also
we're
truly
open,
we
like
to
call
we,
we
can
integrate
with
any
tool
envelopes,
tool
or
data
science
tool
that
run
on
top
of
kubernetes
okay.
So
you
can
get
like
simple,
very
native
integration
to
our
platform
as
well,
because
we're
based
on
kubernetes-
and
we
built
it
like
that.
So
we
have
all
our
own
tools
and
then
we
can
integrate
with
any
yeah.
A
I
gotcha,
okay,
that
makes
sense
so
so
you
were
saying
saying
before
that
you're
kind
of
doing
a
lot
more
with
gpus,
like
you
know,
you
mean,
like
kind
of
surfacing
it
up
to
end
users
or
like
how.
How
is
it
you
know
what
what
is
it
that
you're
bringing
to
the
table
kind
of
that
helps?
You
know,
making
gpus
accessible,
yeah.
B
A
B
Know
it's
an
expensive
resource
right,
yeah,
I,
think
on
on
Amazon,
when
you,
when
you
take
small
gpus,
it
costs
like
30
dollars
per
hour,
so
the
newest
GPU
yeah.
So
it's
an
expensive
resource
right
and
and
and
and
and
research
is
data
science.
They
need
a
lot
of
it
right
and
it's
really
important
for
them
to
walk
so
getting
access
to
that
com
to
to
those
Computing
resources
and
utilizing
in
the
in
the
right
way
in
an
efficient
way.
B
B
Yeah
right
so
they're,
definitely
not
designed
to
that
and
I.
Think
if
you
look
at
on
gpus
and
you
compare
them
to
CPUs
so
for
CPUs
for
30
years,
a
lot
of
software
was
built
to
just
abstract
the
CPU
usage
right
right,
you
don't
when
you,
you
run
your
workload
on
CP,
you
don't
even
care
about
on,
on
which
exact.
A
Works,
it's
not
like
that
with
gpus
yeah
right
shoes
have
been
having
problems
lately
because
they're
running
M1
Max
and
you
know
a
lot
of
you-
know
a
CS
class
right.
You
know
there's
virtualization
and
stuff
and
they're
like
totally
can't
do
anything
because
they,
you
know
it's
an
arm
chip
and
they
can't
virtualize
anything
yeah.
So
it's
funny
because
you
like
it
didn't
even
cross
my
mind.
You
know
when
people
started
showing
up
with
m1s.
This
was
going
to
be
a
problem.
You.
B
B
And
and-
and
they
often
utilize
them-
often
not
and
so
you're
getting
a
problem
with
the
with
the
utilization
of
those
expensive
resources
and
many
times
we
see
either
gpus,
while
other
teams
other
other
researchers,
are
waiting
to
get
access
to
gpus
right
right.
So
we're
doing
a
lot
of
work
in
in
that
sense
and
and
providing
tools
and
efficient
software
and
abstraction
layers
to
help
customers
better
utilize.
Their
gpus
gotcha.
A
Yeah
that
makes
a
lot
of
sense
and
and
I
think
you
kind
of
mentioned
it
a
little
bit,
but
I
think
that
separation
problem
is
also
really
really
important
right
because
you
know
like
I'm,
not
actually
a
researcher
myself
I'm
more
on
the
clinical
side.
So
but
you
know
I've
done
a
lot
of
software
and
it's
you
know
you
want
to
make
sure
that
you're
not
tainting.
You
know
the
the
work
of
somebody
else,
especially
on
a
platform
that
is
not
really
designed
to
be
shared.
A
B
And
yeah
I
think
today,
researchers
are,
they
have
they
deal
a
lot
with
the
infrastructure
and
and
and
and
and
a
lot
relates
to
that
or
on
how
they
shift
the
the
environments
and
software
from
one
place
to
another
yeah.
So
that's
really
difficult.
Today,
yeah
yeah,
just
shifting
your
your
environment
from
one
GPU
to
another
GPU
that
will
take
like
two
weeks
for
you
as
a
researcher,
sometimes
right
so
and
so
with
containers
and
and
our
platform
is
built
on
top
of
kubernetes.
So
you
can
simplify
a
lot
of
that.
A
B
A
A
B
A
B
B
And
what
even
is
the
utilization?
They,
you
know
I.T
teams
they
only
get
active.
They
only
get
requests
to
buy
more
gpus
right
to
increase
the
limits
in
the
cloud
to
have
more
gpus,
but
they
don't
even
know
what's
the
usage
so
right,
so
we
created
this
this
tool,
we
call
it
Alan
top
and
and
it
it
and
I
think
it
truly
simplifies
non-kubernetes
environments.
A
B
A
A
B
Helps
teams
of
data
scientists
to
share
gpus
in
an
easy,
Airways
safely
away
provision
gpus
simply
a
simple
way.
So
we
just
really
listed
like
a
month
and
a
half
ago.
I
think
it
already
had
has
like
hundreds
of
downloads
yeah,
actually
because
we
integrated
it
into
pie,
charm,
yeah
and
into
a
vs
code.
A
A
That's
pretty
fun
to
see
yeah
have
you
have
you
talked
at
all
to
an
organization
called
Massachusetts,
open
Cloud
by
any
chance?
No,
so
there's
an
organization
that
is
it's
sponsored
by
some
industry.
You
know
companies
like
so
red
hat
is
obviously
a
sponsor.
That's
part
of
why
I
know
about
it,
but
bu
is
also
a
sponsor
that
shows
Harvard
and
you
know,
and
basically
what
they're
trying
to
do
is
like.
Can
they
build
an
open,
Cloud
and
but
really
what
they
want
to
do?
A
Is
they
want
to
give
access
to
researchers
to
like
the
underneath
parts
of
a
cloud
which
you
can't
really
get
from
like
Amazon?
You
know
for
that
much,
but
then,
but
in
order
to
give
real
world
data
about
what's
going
on
on
the
under
Cloud
kind
of,
they
also
need
to
give
they
want
to
have
real
workloads.
So
they
have
researchers
kind
of
sitting
on
top
doing
data
science
research,
but
then
they
have
other
research
looking
underneath
to
see
what
it
does
to
it.
A
But
that
might
be
a
really
good
kind
of
place,
for
you
know
you
to
get
involved
with
like
the
GM
right.
It's
like
that's
the
kind
of
thing
they're
experimenting
with
is
like:
how
can
they,
you
know,
shift
workloads
say
between
clouds,
for
example,
or
how
do
you
make
one
of
the
big
things
they
they
want
to
do
is
like
Harvard,
let's
say
has
I,
don't
know
20
machines,
but
at
any
given
point
right
at
two
in
the
morning
on
Thursdays.
A
They
don't
actually
have
anything
running,
and
so
they
want
to
loan
that
to
another
organization,
but
they
want
enough
control
that
they
can
bring
it
back
right.
So
that's
one
of
the
problems
they're
trying
to
like
solve,
but
it's
kind
of
the
same
utilization
thing
except
it's
cross,
org,
oh
yeah,
because
Harvard
wants
to
give
the
resources
to
BU,
and
then
you
know
something
later
day
when
bu's
got
free
resources,
they
give
resources
back
to
Harvard
because
there's
a
bunch
of
universities,
so
they
don't.
A
They
don't
have
a
vested
interest
in,
like
you
know,
kind
of
protecting
their
little
Silo
and
then
Red
Hat.
You
know
gets
involved
because
they
want
to
run.
They
want
to
see
how
their
software
performs
right
at
that
kind
of
scale.
So
it
just
might
be
a
really
good.
You
know
opportunity
for
a
relationship.
A
A
B
B
A
A
B
B
A
B
A
A
A
B
B
A
B
A
A
A
B
A
Makes
sense,
okay,
but
that's
probably
why
I
live
in
the
Northeast,
so
so
yeah,
so
when
so,
why
you
know
so
all
of
your
kind
of
infrastructure
runs
on
kubernetes,
which
I
like
I
totally
get,
then
why
you
want
to
be
kind
of
engaged
with
the
community
at
kubecon,
but
you
know:
do
you?
Do
you
also
have
kind
of
customers?
You
want
to
see.
B
A
B
B
B
Or
machine
learning
infrastructure
right,
so
those
people
are
usually
engineers
and
they
know
a
lot
of
what
kubernetes
and
then
they
are
building
the
platform
internally
and
and
so
they
are
Engineers,
they
know
something.
They
know
a
lot
about,
kubernetes,
right,
right
and
so
men-
and
they
are
here
as
well
in
cubecon,
Right
coupon,
is-
is
exciting.
That's
yeah!
Well,.
B
Spending
the
money-
and
you
know
that's
one
thing
so-
customers
for
sure
yeah
yeah
that
yeah
customers
for
sure,
but
also
we're
doing
a
lot
with
kubernetes
right
and
we're
not
just
building
on
top
of
kubernetes
but
we'll.
Also
we
we,
you
know
our
Engineers
are
doing
a
lot
of
things
internally
in
kubernetes.
So
right,
we've
built
our
own
scheduler
yeah.
A
B
Differently,
yeah,
so
you
know
the
the
problem
of
utilizing.
Gpus
relate
much
to
scheduling
problems
so
and
scheduling
on
gpus
that
could
be
sometimes
difficult
right
and
and
for
AI
when
you
you're
you're
running
a
AI
workout,
which
are
really
computer
intensive
and
they
are
using
expensive
resources
like
gpus,
and
what
we
saw
that
the
scheduling
capabilities
that
are
that
are
known
for
HP
the
HPC
world
I.
Don't.
B
B
A
B
A
Yeah
the
yeah
schedulers
are
like
are
really
interesting.
I
mean
one
of
the
things
they
you
know
it's
funny,
because
you
talk
about
AI
right
is
that
you
know
I,
don't
understand
why
we
haven't
seen
more
AI
kind
of
to
create
schedulers,
you
know
or
schedulers
that
are
enabled
by
their
own
machine
learning.
A
You
know
it's
like
I
I
was
toying
around
with
an
idea
for
a
while
of,
like
you
know,
you
know,
there's
a
whole
organization
at
Red,
Hat,
there's
whole
organizations
elsewhere
too,
that
are
like
putting
out
perfectly
optimized
tunings
for
like
databases
on
Rel,
for
example,
but
it's
like
well
how
expensive
would
it
be
to
actually
just
run
a
little
machine
learning
model
on
the
machine
itself?
That's
kind
of
watching
the
kernel
and
saying
hey
we're
getting
spikes
like
this.
A
Why
don't
we
twiddle
this
knob
and
try
it
for
a
while
right
and
then
go
on
to
the
next
one?
Oh,
that
didn't
work
very
well.
So
let's
just
turn
it
back.
You
know
I
I,
wonder
why,
like
I
feel
like
we
could
be
using
the
machine
learning
techniques,
we
know
to
kind
of
modify
schedulers
based
on
what's
literally
happening
at
the
moment,
which
I
always
think
is
kind
of
interesting,
but
you
have
to
have
the
capability
right.
Yes,.
B
They
came
with
also
with
their
own
schedule,
so
schedule
is
like
the
heart
of
the
system,
right
of
orchestrators
right
the
platforms.
So
it's
an
important
aspect
of
a
system
working,
so
scheduling
can
be
done
in
different
layers,
so
you're
speaking
about
the
the
low
level
layer
right
and
and
for
sure,
I
think
I
totally
agree
with
you.
You
can
do
things
with
the
with
machine
learning
just
to
to
optimize
certain
certain
specificities,
specific,
like.
A
Yeah
like
why
do
I
have
to
tell
the
computer
there's
a
database
running
on
it,
it
can
figure
it
out
for
itself
and
then
tune
itself.
It's
kind
of
what
I've
I've
been
wondering,
and
so
with
like,
but
kubernetes
schedule
right.
You
can
distributed
system.
Schedulers
are
super
interesting
as
well,
because
you
have
all
these
other
kind
of
variables
at
hand
right,
because
you
not
only
have
the
obvious
stuff
like
load,
but
you
also
have
things
like
our
resource.
Like
the
amount
of
resource.
A
We
have
changes
right,
as
you
add
machines
to
the
cluster
or
remove
them
from
cluster
and
stuff.
Like
that,
you
know
you
have
all
these
other
things
going
on
in
a
distributed
system,
environment.
You
know,
that's
Way
Beyond,
just
you
know
kind
of
a
simple.
You
know
single
kernel,
environment
of
it,
so
I
think
schedulers.
A
It's
totally
a
mistake
yeah
because
you
know
it's
like,
and
this
is
where,
like
you
know,
kind
of
we're
already
kind
of
doing
it
or
recognizing
it
right.
That's
what
Auto
scaling
is
you
know
is
like
okay,
so
the
humans
can't
actually
predict
in
advance
what
the
scaling
looks
needs
to
look
like.
So
we've
introduced
these
Auto
scalers,
but
so
can
you
just
do
that
with
all
the
things
you.
A
B
A
B
A
B
B
A
A
B
A
Like
you,
the
utilization
get
on
Street
Hardware
was
on
CPUs,
where
it
was
terrible
like
20,
you
know,
and
then,
even
when
we
started
doing
virtual
machines
getting
utilization,
you
know
you'd
start
to
get
utilizations
in
like
60
70,
but
I
mean
that's.
Why
containers
are
so
awesome
because
you
can
like
get
ridiculous
density
and
you
can
over
provision
such
that
if
there
is
even
some
utilization
room,
you
can
shove
something
else
in
there
really
easily.
You
know
which
I
think
is
you.
A
Amazing
right:
well,
it's
funny
because
you
know
it's
like.
Even
in
my
advanced
classes,
you
know,
because
I
teach
like
I
teach
very,
very
like
first
semester
freshman
year,
students
and
I
also
teach
like
ones
who
are
like
seniors
and
graduate
students,
and
so
I
have
both
ends
of
the
spectrum,
but
even
the
seniors
and
graduate
students
like
trying
to
explain
to
them
like
because
they
didn't
live
through
it
right
so,
but
trying
to
get
you
know,
show
them
it's
like.
A
Okay,
so
I
have
straight
Hardware
right
and
then
I
have
virtual
machines
and
then
I
have
containers,
because
what
we
want
is
we
want
every
iota
of
computing
power
to
get
used
right
to
maximize
the
amount
of
you
know.
If
we
spend
a
certain
amount
of
money,
we
want
to
use
as
much
of
that
as
we
can,
and
we
don't
do
that
in.
In
most
other
scenarios
we
do
get
very
good
utilization
right.
You.
B
Know
and
I
think
that,
like
now,
a
big
story
that
is
happening
in
the
world
right
now
is
around,
like
even
increasing
those
abstraction
layers.
In
terms
of
like
the
delay
is
that
the
the
the
level
of
the
layers
of
abstraction
with
hybrid
cloud
and
multi-cloud
right
story,
and
that's
an
amazing
story
right,
we're
hearing
it?
The
love
from
customers?
Who
are
you
know,
and
they
are
now
they
are
building
the
platforms,
the
services
and
they
strategize
around
multi-cloud
or
hybrid
Cloud,
Solutions,
and
so
we're
hearing
it
a
lot
from
customers.
B
That's
that's
big,
that's
big
and,
and
it's
a
lot
enabled
by
kubernetes
right
I
believe,
like
I,
believe
that
no
all
companies
all
startups
and
and
now
on
will
build
their
services
or
their
platforms
as
a
hybrid
cloud
or
multi-cloud
solution.
If
now,
I'm
building
a
new
service
I
will
build
it
as
a
multi-cloud
service
on
top
of
kubernetes,
there's
no
reason
not
to
do
that
right,
right,
right,
yeah,
so
because
the
value
is
so
so
so
big
right.
If.
B
A
A
A
It's
like
it's
as
good.
You
know
you
have
your
perfect
data
center
if
it
even
exists
right
and
you've,
gotten
that
really
high
utilization
with
containers
and
kubernetes
or
whatever,
and
then
you're
like
oh,
but
we
want
to
do
hybrid
cloud,
and
so
now
you
have
like
I
said
you
just
brought
that
utility
problem
back.
You
need
to
optimize
for
it
all
over
again.
A
A
A
A
My
grand
plan
was
I
was
going
to
do
philosophy,
undergrad
and
then
go
to
grad
school
and
do
AI
by
doing
CS,
because
in
those
days
because
I'm
old
people
thought
of
AI
as
a
hybrid
of
philosophy
and
computer
science.
Now
they
think
of
it
as
more
like
psychology
and
computer
science,
okay,
but
yeah.
So
that
was
my
original
plan
and
then
I
I
I
took
a
I
took
a
CS
class
in
my
like
sophomore
year,
like
second
semester
sophomore
year
it
was
the
intro
to
CS.
A
A
A
B
B
A
A
Mean,
but
even
just
like
you
know,
navigating
Maps
right,
like
I
mean
you
know,
like
Auto
navigation,
I
mean
it's
just
it's
in
there
all
over
the
place.
You
don't
even
realize
it,
which
is
the
promise,
in
my
mind,
of
like
good
Computing
good
software
is
like
you
shouldn't,
even
see
it
right.
It
should
just
work.
Oh.
A
That
revolutionary
Revolution
and
help
right
right,
yeah,
the
I,
think
the
you
know
it's
a
little
shout
out
to
the
Ford
right.
I
mean,
like
you
know
this
car
right.
It's
like
it'll
do
like
Auto
parking
and
you
know
Auto
driving
to
some
level
and
stuff,
and
you
know
I
have
the
both
the
homogeny
aspect
too.
It's
like
you,
know,
I
actually
like
driving.
You
know
like
I,
like
I,
like
kind
of
the
control,
and
you
know
and
I
missed,
having
a
stick
shift.