►
From YouTube: AWG Meeting 012423
Description
Presentation by Markus Muller on Open Source Tools for Modeling: A Hands on Overview
A
B
A
B
B
B
D
B
A
F
C
G
So
my
background
is
really
memory,
design
and
circuit
design
like
Technologies,
you
could
design
and
then
I
went
more
in
the
architectural
level
as
it
was
Doc
and
then,
in
the
last
few
years,
I
took
the
lead
on
the
design
center
in
etfl,
basically
being
centralized
all
the
activities
in
media
circuit
design
for
circuit
design,
and
so
we
are
particularly
interested
in
a
bunch
of
tools,
because
that's
what
people
want
to
do
these
days
and
so
yeah.
We
are
curious
about
how
we
could
integrate
that
in
our
infrastructure.
F
Around
Alex
there
are
Professor
Brooke,
Burke
and
tatiensa.
They
are
both
digital
designers
running
our
academic
projects.
So
I
had
chance
to
talk
to
other
group
at
eth.
It's
a
Technical
University
in
Zurich
to
professor
olukabanini
and
his
design
support
team
I'm,
not
sure
if
they
managed
to
join
us,
but
maybe
later
I
will
also
share
the
contacts.
So
epfl
is
in
lausanne,
eths
in
in
Zurich
and
as
Alex
works
for
epfl.
It
would
be
great
to
also
understand
or
learn
more
about
Open
Source
activities
at
dth
in
Zurich.
B
F
B
C
E
C
B
Awesome
so
I'll
I'll
start
by
introducing
Marcus
here
I
think
he
has
his
colleague
Mario
as
well.
But
Marcus
is
a
research
Association
PhD
candidate
I,
don't
know
if
that
changed
Marcus
yeah
all
right,
so
his
interests
include
compact
modeling,
t-cat
simulation
and
characterization
of
transistors.
In
particular,
hbts
he's
been
involved
in
several
software
projects,
as
part
of
in
his
part
of
a
startup
called
Simi
mod,
so
I
invited
Marcus
based
on
semi-mods
announcement
of
their
beta
phase
of
their
bellock,
a
compiler
called
open,
vaf.
B
So
based
on
Marcus
email
right,
please
correct
me:
Marcus
the
first.
There
are
three
main
use
cases
here:
the
first
one
is
integration
of
Open
Source
pdks
into
open
source
simulators
that
employs
models.
That
cannot
be
simulated
currently.
The
second
case
is
use
of
Open
Source
simulators
as
backend
for
model
parameter
extraction
and
model
development,
and
the
third
case
is
available.
Standardization
and
development,
so
I'll
I'll.
Let
you
give
us
your
update
and
yeah.
E
E
F
E
So
far
there
have
been
some
activities
in
that
direction,
but
from
the
modeling
point
of
view,
not
so
much
there's
a
shared
document
by
Tim
answer
from
Google,
where
he
talks
about
automated
parameter
extraction,
using
open
source
tools
with
the
goal
to
improve
pdk
equality
and
push
the
open
source
Community
forward
so
flag,
Tim
Ansel
me
and
sarayuki
had
some
discussions
in
the
last
months.
So
far
we
don't
have
a
clear
Direction,
so
I
I
hope
that
I
can
help
steer
the
discussion
into
a
clear
direction
today.
E
E
Then
I
would
like
to
explain
briefly
the
software
development
steps
we
need
to
undertake
in
order
to
enable
ekv
or
piece
in
parameter
extraction
and
I
also
want
to
give
everybody
an
idea.
How
parameter
extraction
looks
in
practice,
because
not
everybody
in
this
group
has
extracted
their
model
so
far,
I
think
and
then
we
can
have
a
discussion.
E
So
what?
What
do
we
need
for
a
parliament
extraction
system?
So
this
is
a
long
list,
so
we
want
to
be
quite
modular
because
there's
many
different
models,
there's
different
sources
of
measurements
and
there's
different
circuit
simulators
and
usually
we
need
the
equations
and
also
some
way
to
simulate
a
given
model.
So
we
need
verlock
a
interfaces
both
for
model
parameters,
equations
and
circuit
simulators.
E
When
we
eat
about
real
world
modeling,
we
don't
have
one
transistor
at
one
temperature.
We
have
lots
of
transistors
at
different
temperatures-
geometry
statistical
stuff,
so
we
would
have
usually
a
lot
of
data
that
we
need
to
handle
in
some
way
and
the
parameter
extraction.
Steps
are
usually
model
or
at
least
device
type
specific,
and
somebody
really
has
some
noise
going
on
here.
E
So
if
we
draw
this,
this
is
from
a
paper
of
us
long
time
ago
we
have
this
overview
of
the
system
and
I
put
some
red
terms
there.
That
denote
the
tools
we
are
using.
So
our
parameter
extraction
system
is
called
DMT
and
we
have
open
sourced
a
large
part
of
it
and
that's
basically
the
glue
where
everything
comes
together
from
data
management
we
use
pandas,
which
is
a
very
popular
Library,
especially
for
machine
learning
and
stuff
in
Python
for
circuit
simulation.
E
E
The
model
equations
and
you
can
have
all
information
regarding
model
parameters
and
interdependencies
between
model
equations.
Since
a
few
minus
internally
and
since
a
few
weeks,
we
have
released
our
verlock
a
compiler
open
web,
that's
already
integrate
it
into
the
pre-master
of
NG
spies.
So
that's
what
we
use
to
simulate
models
with
energy
spies,
because
it
does
not
Implement
many
models.
E
So
before
we
start
with
the
Hands-On
demo,
I
will
just
give
a
quick
overview
of
the
model
we
will
be
using
I.
Think
it's
a
really
cool
model.
I'm
revealing
myself.
So
as
ekv
is
a
really
good
vehicle
for
parameter
extraction,
because
it's
so
simple
and
small
that
everybody
can
understand
it
and
there's
both
analytical
and
numerical
parameter
extraction,
steps
involved.
E
So
a
lot
of
the
parameters
I
mean
it's
only
four
in
this
case
have
extraction
steps
where
you
do
not
numerically
optimize
some
circuit,
simulator
or
something
you
can
basically
Buy
analyzing
the
model
equations
get
them
straight
forward
and
I
will
show
you
how
exactly
in
a
minute.
So
we
have
these
three
equations
and
they
basically
only
specify
the
drain
current
as
a
function
of
gate,
Source
voltage.
This
is
a
simplified
overview,
so
we
have
basically
only
one
implicit
equation
for
the
drain
current
and
that's
what
we
will
be
using
now
as
an
example.
E
A
E
Yes,
that's
perfect,
so
I
prepared
two
scripts
that
I
want
to
walk
you
through.
So
as
I
told
in
the
beginning,
I
said
in
the
beginning.
This
is
just
how
I
do
it,
and
this
can
inspire
you
or
we
can
also
like
look
forward
how
we
can
proceed
with
semi
mod
or
somebody
else
does
it.
But
these
are
the
open
source
tools
we
are
using
and
we
will
just
now
demonstrate
them
live.
E
So
the
first
tool
I
would
like
to
show
is
verilog
AE,
so
very
log
AE
is
a
front
and
to
our
verilog
a
compiler
that
you
can
basically
install
with
Pip.
So
you
would
just
need
to
write
pip
install
verlock
AE
in
your
terminal
and
then
it
would
be
installed.
So
it's
pretty
simple
to
install
and
then
we
can
load
a
verilog,
a
files.
So
this
is
just
basically
a
path
to
the
very
log.
E
Yes,
so
it
was
model
card,
so
we
have
only
four
parameters
in
this
case
and
Iceberg
VT,
0
and
Lambda
C,
and
we
get
all
the
information
about
these
parameters
here,
like
the
description
and
the
minimum
and
maximum
value.
But
this
is
not,
of
course
difficult
to
parse,
but
what
more
cool
is
we
can
Mark
equations
in
the
verilog,
a
source
file
and
evaluate
them
in
Python.
So
let's
look
at
the
sekv
model,
so
we
have
here
the
drain
current
variable,
which
is
just
called
I
and
I,
wrote
retrieve
in
front
of
it.
E
E
E
We
have
to
give
it
the
temperature
and
all
branch
voltages,
and
then
we
can
evaluate
the
drain
current
in
a
millisecond.
Of
course,
this
is
not
a
circuit
simulator
but,
for
example,
local
runs.
The
series
resistances
are
not
important,
and
in
this
case
the
model
is
so
simple
that
there's
not
even
a
serious
resistance,
so
this
is
exact
and
we
can
use
this
function
directly
for
optimization
and
other
stuff.
E
Of
course,
the
the
function
of
s
the
drain
current
function
of
this
model
is
super
simple
and
you
can
easily
implement
it
yourself
in
Python,
but
if
you
have
big
models
like
a
real
ekv
or
others-
and
this
saves
you
a
lot
of
time
and
what
we
can
also
do
is
using
the
parameters
attribute
of
the
functions,
we
see
all
parameters
that
influence
the
value
of
a
function.
Of
course
here
we
have
only
one
function
in
the
whole
model,
so
all
parameters
influence
it.
E
E
So
these
are
all
the
parameters
you
would
need
to
look
at
if
you
want
to
extract
the
model
and
you
need
to
know
what
they
mean
and
how
to
extract
them.
And
of
course
we
can
also
evaluate
the
much
more
complement,
complicated
rank,
current
function
of
this
model
now
without
having
written
a
single
line
of
code.
Actually,
apart
from
the
function,
call.
E
So
how
can
we
use
all
of
this
for
parameter
extraction
here?
I
have
prepared
a
small
script
where
we
extract
the
parameters
of
the
model.
I
have
two
switches
here:
automated
extraction,
because
people
have
been
talking
about
automating
all
of
this,
so
let's
put
it
to
false
first
and
let's
use
the
sekv
model.
First,
so
I
have
here
some
switches
if,
as
ekv
will
load
the
corresponding
model
card,
so
first
we
Define
the
model.
Again
we
use
the
verilog
a
defined
model.
E
E
We
have
defined
a
technology.
This
would
usually
Define
equations
for
scaling
the
models
and
we
have
also
some
specifiers.
While
we
import
the
data,
because
all
the
measurement
Engineers
have
different
naming
schemes,
we
Define
a
grammar
that
is
unique
inside
our
extraction
environment,
and
this
specifier
still
defines
this
grammar.
E
We
Define
an
extraction
object
and
then
the
core
part
are
the
steps.
So
each
step
has
the
same
structure.
A
step
is
a
class
that
inherits
from
from
a
generic
class.
It
has
a
name,
we
give
it
a
model
card,
a
library
with
measurements,
and
this
thing
here
is
what
we
call
it:
semi-mod
the
filter.
So
here
we
would
select
all
data
at
300
Kelvin,
with
drain
Source
voltage,
5
volt
gate
Source
from
0.3
to
none
and
base
bike
zero
volt.
E
We
give
it
the
verilog
a
model,
and
it
will
also
use
only
measurement
data
that
has
this.
In
its
name.
We
Define
some
parameters
to
optimize
and
some
other
specific
parameters
that
are
specific
to
each
step.
So
we
can
build
the
extraction
Flow
by
changing
these
steps
and
in
the
end
we
can
call
a
GUI.
E
E
Basically,
each
step
has
our
main
plot,
where
you
can
use
this
tool
here
to
select
data,
you
can
optimize
equations
that
are
defined
in
a
python
code
and
you
can
have
additional
plots,
but
in
this
case
the
step
has
none
and
each
step
has
its
own
local
model
card
and
there's
also
a
global
model
card
so
very
similar
to
git.
We
interact
with
the
local
model
card
and
push
only
the
parameters
here.
You
can
select
it
to
the
global
model
car
that
we
have
optimized.
E
So,
as
I
said,
sekv
is
a
really
nice
model
because,
for
example,
when
we
plot
ID
over
gmvt-
and
we
can
just
take
this
value
at
very
low
current
densities
as
n,
so
we
have
optimized
this
now
we
like
the
value,
so
we
select
push
and
then
we
press
here
push
now.
The
value
we
have
optimized
is
in
the
global
model
card
here,
so
the
next
step
Optimum,
we
optimize
only
in
the
high
current
region
there
if
you
plot
ID
over
gmvt
against
ID.
E
E
Now
we
have
the
last
step
where
we
take
the
very
log
equation
and
optimize
on
the
raw
data.
So,
depending
on
what
you
select
here,
of
course
you
will
get
different
results
and
you
can
also
use
different
normalizers
and
stuff
and
other
things,
and
what
you
can
also
do
with
I
forget
to
say:
is
you
can
also
hand
tune,
parameters
here
and
directly
see
the
impact,
and
you
can
select
the
parameters
you
want
to
actually
optimize.
E
So,
for
example,
I,
like
the
value
of
n,
so
I
will
deselect,
it
I
will
optimize,
and
now
we
have
a
pretty
good
fit
here.
So
usually,
what
we
do
in
the
end
is
look
at
the
circuit
simulation
in
this
model,
because
the
model
here
is
so
simple
does
not
make
difference.
So
if
you
go
to
the
Circuit
simulation
object,
which
is
called
xverify
in
our
infrastructure,
we
pull
the
parameters
and
we
see
we
have
a
pretty
good
model,
at
least
if
we
look
into
the
linear
scale.
E
E
So
what
we
can
also
do
is
if
we
select
all
the
ranges
in
the
data
that
we
want
to
optimize
on,
and
we
put
them
into
the
script
like
this
and
we
can
do
something,
but
that
comes
close
to
automated
extraction.
So
we
would
activate
automated
extraction
and
then
all
of
the
steps
take
their
default
ranges
for
the
data
and
call
their
Optimizer,
and
then
they
push
the
optimized
parameters
and
pull
off
the
previous
step.
So
this
is,
as
far
as
we
have
come
to
automated
extraction.
E
So
now
all
the
steps
have
optimized
themselves.
So
using
this
data,
which
we
have
put
into
the
script
manually
once
so
I
looked
at
the
data
as
a
modeling
engineer,
I
selected
the
right
and
put
it
into
the
script,
and
now
we
can
do
circuit
simulation
again
and
it
looks
perfect.
I
mean
not
perfect,
but
not
bad.
Let's
say
it's
a
very
simple
model
too.
E
E
Everything
else
stays
the
same
and
actually
all
the
remains
of
the
script
stay
the
same.
Just
these
steps,
which
are
model
specific,
like
the
extraction
of
n
at
low
current
densities,
we
will
not
do
now
with
the
full
ekv
model,
so
I'm
no
expert
for
ekv
I've
never
extracted
a
full
ekv
model,
but
just
we
change
the
model
now
and
let's
deactivate
automated
extraction.
E
E
C
E
F
E
Yeah
sure,
but
I
mean
what
I
want
to
say
is.
These
are
not
necessarily
good
parameters.
We
have
now
adjusted
a
very
stupid
numerical
optimization
and
it
kind
of
works
for
this
one
characteristic,
but
if
we
would
now
go
through
the
full
measurement
data,
we
must
make
sure
that
the
model
parameters
actually
make
sense.
We
will
see
the
issues
if
they
don't
make
sense
later
so
actually,
this
is
all
I
wanted
to
show
with
respect
to
our
graphical
user
interface
and
the
extraction
flow
so
for
in
semi
mode.
E
We
only
extract
at
the
moment
hbt
models.
This
is
our
specialty
and
we
do
software
engineering,
but
we
could
easily
also
put
here
by
defining
the
appropriate
steps
for
our
mosfet.
We
could
build
the
same
thing
for
our
mosfet,
for,
for
example,
the
ekv
model
and
let's
go
to
the
slides
bank,
so
I'm,
sorry
before
we
continue
the
discussion.
E
Then
we
have
model
geometry
and
temperature
scaling.
We
want
to
model
the
almost
over
the
whole
range
where
you
can
buy
the
transistor
without
destroying
it
so
inside
the
safe
operating
area.
So
a
we
must
make
sure
that
the
extracted
model
parameters
somehow
fit
into
the
pdk,
and
we
have
also
modeling
over
a
frequency
and
transient
effects
like
trapping
and
stuff
statistical
modeling
and
in
the
end,
in
practice
of
modeling,
at
least
from
our
perspective,
some
parameter
tuning
is
necessary,
so
we
extract
many
parameters
using
analytical
steps.
E
E
E
There's
some
guys
in
Australia
working
on
fully
automated
extraction
using
artificial
intelligence
and
I
think
that's
the
right
way
to
go.
But
this
is
really
not
our
problem.
Right
now,
so
at
semi
might
be
really
good
with
the
open
source
tools
I
have
just
presented,
but,
for
example,
we
are
no
ekv
experts
and
also
we
must
must
rent
a
lab.
If
we
want
to
measure,
for
example,
the
university
people
might
have
better
access,
so
we
must
split
the
project
between
the
partners.
E
If
we
want
to
have
a
success
in
my
opinion-
and
another
point-
is
the
stuff
we
built
here-
the
the
extraction
flow
can
be
reused
for
the
next
Technologies
for
sure.
So
actually
that
was
my
talk.
These
are
the
references
and
now
I
hope
we
can
have
some
kind
of
discussion
to
see
how
how
we
can
move
forward.
F
First,
I
really
would
like
to
thank
Marcus
for
the
presentation.
It's
really
a
huge
amount
of
work
and
impressive
developments.
I'm
a
little
bit
biased,
so
I
will
not
talk
anymore
and
I've.
Seen
matia's
Booker
is
joining
Carl,
so
he
works
on
newest
AKB,
free
model
implementation,
very
local
coding.
So
he
can
comment
also
on
the
extraction
and
extraction
flow,
as
Marcus
asked
for.
H
H
Glad
to
cooperate
with
you
on
on
ekd,
three
model
extraction
and
and
and
in
some
in
some
extent-
let's
say
automation
of
that,
but
I
I
definitely
agree
on
on
your
point
of
view.
Full
automation
is
maybe
not
not
really
so
much
possible
it.
It
takes
a
lot
of
data,
and-
and
maybe
maybe
at
that
moment
this
is
maybe
not
the
main
goal
but
but
still,
let's
say
to
to
some
extent
this.
H
H
So,
yes,
I'd,
be
I'd,
be
glad
to
to
interact
with
with
anyone.
We
have
already
done
some
work
on
on
EK
V3
extraction,
using
using
your
your
tool
for
on
the
EK
V3
implementation
within
g-spice,
I.
Think
in
the
last
meeting
my
co-worker
and
PhD
student
Nicos
mcgrizz
had
presented
some
things
on
honey
kv3.
We
we
would
be
very
glad
to
cooperate
with
you
on
this
and
on
extraction
as
well.
Are.
H
H
E
B
B
B
H
F
B
F
B
F
E
I
will
be
glad
to
look
at
the
current
structure
and
then
see
how
it
fits
to
our
structure.
I
mean
basically,
what
we
do
is
we
have
one
directory
for
each
technology.
Then,
under
each
technology
we
have
per
wafer
One
Directory,
then
per
die
per
device
per
temperature
and
then
inside
the
measurement
data
we
use
a
specific
naming
scheme,
which
is
very
useful.
E
So,
for
example,
if
we
go
here
in
in
the
tool
we
have
defined
it
once
and
we
only
use
these
python
objects
so
that,
if
you
use
them,
you
have
always
the
same
naming
scheme.
So,
for
example,
that's
just
the
10.
and
that's
just
the
drain.
Source
VDS,
but
in
a
specific
way
I
mean
some
people
write
VDS.
Some
people
write
this
and
then,
if
you
automate
it
and
you
have
different
sources,
it's
a
big
headache.
So
you
have
to
somehow
make
sure
it's
always
named
similar.
F
E
E
I
forgot
one
thing:
there's
different
websites
for
the
tools,
so
DMT
core
you
can
find
just
by
Googling
for
DMG
core,
and
then
you
will
find
this
gitlab
page.
So
we
maintain
this
on
gitlab,
not
on
GitHub,
but
I
think
there's
a
mirror
on
GitHub
and
there's
also
a
website
for
it.
So
if
you
go
here
to
the
readme,
you
can
go
to
the
DMT
website
and
the
installation
and
stuff
is
explained.
E
E
C
B
B
I
J
I
J
Not
yeah,
but
the
thing
is
what
I'm
saying
is
like:
if
you
get
the
open
source
tools
into
people's
flows,
then
there's
more
incentive
for
them
to
do
the
other
piece.
So
is
it
kind
of
chicken
and
egg
thing,
but
you
can
say
okay,
you
know
from
my
view
what
you're
doing
with
characterization
is
the
kind
of
digital
twinning
you're
just
doing
transistors
to
transistor
models.
J
You
can
also
do
digital
twinning
of
Block
Level
models
to
you
know
verilogy,
so
you
can
take
something
which
is
a
bunch
of
transistors
and
stuff
and
you
know
like
an
op-amp
and
turn
it
into
a
behavioral
model.
That's
similar
digital
twinning,
if
you,
if
you
do
that
stuff
and
get
people
to
use
the
open
source
tools.
C
J
Yeah
yeah
I
understood
what
he
was
saying.
I
find
it.
You
know,
I
agree
people,
you
know
somebody
needs
to
pay
for
that,
but
unless
they're
using
the
open
source
tools
at
a
higher
level,
they
have
no
motivation.
So
what
you
want
to
do
is
to
replace
the
proprietary
tools
at
a
higher
level
to
then
get
an
incentive
to
do
to
get
people
to
pay
for
the
lower
level.
B
Well,
I
think
it's
kind
of
orthogonal
right
now.
The
models
from
skywater
are
kind
of
broken,
even
for
the
commercial
use
right.
So
you
know
just
getting
this
Loop
here
as
part
of
this
open
source
movement
right
just
measurements
in
general,
in
for
each
pdk,
Skyward,
130
or
GF
180,
and
see
how
things
can
be
done
differently.
J
C
E
J
Well,
people
have
people
have
built
chips,
you're
not
get
you're,
not
getting
the
direct
data
from
the
Fab,
but
you
have
a
chip.
That's
been
built
on
the
process,
so
you
can
you,
you
know
just
run
things
through
the
pads
and
stuff
and
see
see
what
they
do
and
infer
stuff
about
the
process
and
transistor
models.
From
that
I
mean
it
seems
like
an
AI
job.
I
mean
it's
something
you
should
be
able
to
do
with
any
chip
and
any
process.
Yeah.
B
J
B
I
I
have
a
question
for
Marcus
right
and
vladic
and
all
other
people
involved
in
measurements.
So
what
kind
of
funding
are
we
looking
for
here?
I
think
it's
more
much
more
of
a
practical
question,
so
you're
requesting
some
funding
here,
Marcus,
and
maybe
you
can
add
that
to
the
Google
Notes
from
or
the
document
from
Tim,
maybe
as
well,
so
other
people
can
have
a
look
at
it
and.
B
F
F
They
are
very
highly
motivated
to
contribute
to
open
source,
but
the
characterization
lab
is
part
of
the
clean
room
process
line,
it's
really
overloaded.
So
it's
question
of
time
and
resources
and
I
I
talked
to
my
old
friends
at
Institute
of
Electro
technology
in
Warsaw.
There
is
a
clean
room
and
characterization
lab
and
they
they
are
quite
open,
so
they
can
dedicate
some
time
and
share
their
resources
for
on
the
wafer
measurements.
F
B
F
K
I
just
wanted
to
just
inform
you
that
we
will
actually
start
measurements
end
of
this
week,
so
measurements
will
be
done
on
skywater
the
technology
and
we
have
selected
a
bunch
of
devices
nmos
pmos
and
we're
we're
gonna
start
measurements.
So
we
now
have
received
the
probe,
so
we
can
automate
automatize
the
the
measurements,
and
so
this
will
be
started
the
end
of
this
week.
Hopefully
we
can
come
back,
I,
don't
know
when
is
the
next
meeting,
but
in
one
month
or.
B
G
K
Up
to
you
guys,
yeah,
so
so
we'll
maybe
we
can
come
back
with
some
first
results.
I
mean
after
isscc.
We
will
already
have
done
part
of
the
the
measurements
will
be
done.
Then,
once
we
have
the
measurements,
we
will
do
some
extraction,
we'll
do
extraction
for
the
simplified
dkv.
First,
that's
the
easiest
one
and
then
we'll
do
also
some
extraction
for
ekv
2.6
great.
B
A
C
K
K
K
C
K
B
So
one
other
additional
thing,
I
requested
from
vladic
and
said
that
UK
to
review
the
test
style
we
made
for
in
mpw5
and
sadiuki
started,
making
some
notes,
which
are
very
you
know
extremely
valuable.
So
maybe
others
can
take
a
look
as
well
and
add
their
notes,
and
you
know
it
is
just
an
improvement
for
the
next
shuttles.
Maybe
you
can
make
another
test
tile.
B
K
C
A
H
Yes,
many
many
of
us,
we
missed
at
least
part
of
it
so
hi
to
Christian,
also
simply
because,
because
it's
just
an
opportunity
to
to
to
to
maybe
connect
a
little
bit
the
things
we
might
be
doing,
because
we
we
have
also
started
doing
measurements
at
TUC.
On
the
on
the
test
chips,
we
do
have
I
I
think
it
would
would
be
nice
if,
if
we
somehow
could
could
coordinate
things
so.