►
From YouTube: 003 ONNX 20211021 Tzortzatos ONNX AI on IBM Z Mainframes
Description
LF AI & Data Day - ONNX Community Meeting, October 21, 2021
ONNX and the AI on IBM Z client journey
Speaker(s): Elpida Tzortzatos & Andrew M. Sica (IBM)
A
B
Thank
you
alpeda,
so
I
want
to
talk
a
little
bit
about
ibm,
z
and
really
what
it
is
right
and
what
it
is
is
the
flagship
server
that
powers
a
large
number
of
critical
core
workloads
that
are
really
the
backbone
systems
of
record
for
clients
across
numerous
industries
and,
as
we
say
here,
there's
a
good
chance.
You've
probably
used
one
today,
especially
if
you've
just
done
some
basic
things
like
use
your
credit
card,
for
example,
and
as
you
can
see
from
the
numbers
on
the
charts
here,
these
systems
are
really
still
in
widespread
use.
B
B
B
Middleware
and
applications
interact
with
each
other
and
get
a
benefit
from
co-location,
so
we're
working
with
our
clients
on
a
number
of
use
cases
for
ai
in
this
environment,
and
many
of
them
need
to
operate
within
the
bounds
of
these
workloads,
which
are
often
low,
latency
high
volume
transaction
environments
to
fit
ai
into
that
environment.
A
key
part
of
our
strategy
is
to
allow
our
clients
to
build
and
train
ai
models
anywhere,
including.
A
B
And
the
deploy
to
z
for
production,
the
benefit
of
onyx
and
the
onyx
ecosystem.
Here,
I
think,
is
very
clear.
It
allows
our
clients
to
leverage
their
existing
investments,
train
on
other
architectures
or
on
ibm
z
or
linux
one,
but
then
have
that
model
portability
without
having
to
deploy
that
same
full
ecosystem
used
for
training
or
development.
B
B
It
automates
the
process
of
compiling
an
onyx
model
and
deploying
it
so
the
models
are
compiled
using
onyx
mlir,
which
gives
us
that
highly
portable
minimal
inference
program
wmlz
then
deploys
a
serving
instance
which
exposes
endpoints
for
the
client
application
use
on
z
os.
This
allows
us
to
provide
different
paths
to
integrating
calls
to
the
models
in
the
business
workloads.
B
One
such
capability
is
a
cobol
scoring
service
that
greatly
simplifies
integrating
the
call
outs
into
a
client's
application.
This
is
a
really
great
feature
because
it
helps
to
minimize
the
code
changes
required
in
the
client's
business
application
and
minimize
the
risk
of
impact
to
those
critical
workloads.
B
So
let
me
touch
here
on
the
value
of
onyx
to
our
clients
and,
as
we've
talked
with
them,
really
what
seemed
to
resonate
with
them,
and
I've
highlighted
some
of
this
as
I've
gone
through
the
deck,
and
I
think
my
main
message
here
is
that
the
core
values
that
you
have
on
the
community
page
of
interoperability
and
hardware
access,
do
really
resonate
with
users
for
ibm
z.
Our
strategy
is
built
around
providing
ai
capabilities
in
a
way
that
aligns
with
the
core
strengths
of
our
platform
onyx
and
the
ecosystem
around.
B
The
adaptability
to
different
architectures
is
also
important
and
has
come
up
as
a
key
feature.
As
we've
talked
about
onyx
with
our
clients,
the
enterprises
we
talk
to
do
not
want
to
have
to
redesign
their
models
or
pipelines
for
deployments
different
environments,
whether
it's
edge
device,
z
or
x86.
So
the
consistency
is
key.
B
We
also
know
from
these
conversations,
of
course,
that
data
scientists
are
not
focused
on
the
point
deployment
platform
in
an
environment
like
ours.
We
want
to
make
sure
that
the
assets
they
do
develop
then
will
be
portable
because
forcing
them
to
re-architect
or
recreate
for
deployment
right
can
cause
the
project
to
fail.
B
B
Okay,
thank
you
all
for
your
time
and
thank
you
for
letting
me
present
here.
There
are
some
links
here
on
the
right
hand,
side
that
talk
a
little
bit
about
what
a
mainframe
is,
and
some
recent
blogs
that
we
put
out
around
our
hardware
announcement
to
tell
them
chip,
as
well
as
our
strategy
on
the
platform
around
ai
and
some
of
the
related
developments
that
are
going
on
there.
So
with
that,
thank
you
again
and
take
care.