►
From YouTube: AWG 032123 Meeting
Description
Presentation by Ken Kundert and Pascale Kuthe on OpenVAF, with Alessandro Piovaccari.
A
A
A
A
A
B
B
It
was
also
arguably
one
of
the
first,
if
not
the
first
open
source
software
project
created
and
as
a
result,
it
kind
of
spread
widely
throughout
the
industry,
and
it
was
used
as
the
background
backbone
for
all
of
or
the
foundation
for
all
of
IC
design.
And
now
you
can't
really
do
analog
IC
design
without
some
kind
of
a
circuit
simulator.
So
it's
a
it's
a
really
foundational
tool,
and
if
you
wanted
to
build
one,
there's
kind
of
two
things
you
need.
First,
you
need
knowledge
of
simulation.
B
B
B
B
There
were
also
things
that
they
didn't
really
need,
and
those
people
that
are
focused
on
the
compact
model,
extensions
oftentimes,
just
neglect
the
part
of
verilog
a
that's,
that's
not
really
suitable
for
model
device
models
and
just
ignores
that,
and
there
was
at
the
time
in
order
to
kind
of
open
things
up.
The
the
basic
philosophy
of
these
compact
model
extensions
is
that
everybody
should
try
to
publish
their
models,
although
at
least
certainly
all
the
modern
models
in
verilog
a
so
that
anybody
could
use
them.
B
This
is
a
this
is
a
big
contribution.
This
is
a
big
idea,
and
the
reason
why
is
because
writing
device
models
in
spice
is
a
horrendous
task.
It's
it's.
The
models
themselves
are
just
very,
very
complicated.
They
they
end
up.
Creating
these
gigantic
functions
with
endless
amounts
of
code
you
have
to
compile.
You
have
to
kind
of
the
models
themselves
have
to
include
not
only
the
the
equations
for
current
and
charge,
but
also
for
their
derivatives.
So
there's
a
lot
of
common
code
that
needs
to
be
factored
out.
B
It
has
to
be
done
very
carefully.
It's
very
easy
for
errors
to
creep
in
and
those
can
be
very
difficult
to
find
because
they
don't
those
errors,
don't
necessarily
cause
the
simulator
to
crash.
They
just
create
convergence
issues,
and
it's
like
hard
to
kind
of
associate
convergence
with
these
issues
with
the
with
the
line
of
code.
That's
creating
the
problem.
B
B
Come
up
with
it,
we
haven't
been
able
to
come
up
with
a
situation
where
we've
got
this
ecosystem,
where
we
generate
vanilla,
game
models
that
are
just
passed
around
and
they
get
installed
in
all
of
the
simulators.
So
the
end
result
is
there's
still
a
substantial
amount
of
investment
required
to
integrate
a
model,
and
anybody
that
supports
a
simulator
has
to
be
able
to
afford
that
investment.
B
Otherwise,
the
simulator
will
be
used
and
which
means
that
open
source
simulation
is
really
hamstrung.
In
this
whole
thing,
open
source
simulators,
they
don't
really
have
the
resources
to
support
the
models
yet
alone,
also
build
the
simulator
and
so
they're
forced
to
kind
of
divert
a
lot
of
their
resources
and
attention
to
the
models
and
and
therefore
the
simulators
themselves
suffer
verilogue
a
itself
has
kind
of
two
aspects
to
it.
B
Now
model
compilers
are
ignoring
this
and
because
a
lot
of
the
open
source
simulators,
they
don't
support
verilog
a
the
kind
of
the
whole
language.
They
only
support
these
compact
model
extensions.
Then
all
of
these
open
source
simulators
are
kind
of
missing
this
aspect
of
error,
log,
a
which
would
be
an
important,
an
important
feature
for
them
if
they
had
it.
B
So
so
that's
that's
kind
of
the
situation.
We're
in
verilog
a
itself
is,
has
a
lot
to
offer
the
simulator
Community,
because
it's
a
standard
language
and
analog
simulation
is
kind
of
beset
by
all
kinds
of
incompatibility
issues,
their
incompatibilities
in
the
models
themselves,
but
there's
also
incompatibilities
in
the
net
list
and
the
fact
that
verilogue,
a
as
a
standard
language
means
that
we
can
move
to
it
as
as
a
kind
of
a
base
of
simulator
development
and
as
a
netlisting
language
and
resolve
a
lot
of
these
things.
B
So
let
me
let
me
talk
a
little
bit
about
this
idea
of
verilog
a
as
the
base
of
a
simulator
development.
If
we
had
a
model
compiler.
If,
if
the
industry
had
a
model
compiler
that
was
strong
and
robust
and
complete
with
a
solid
API,
then
it
would
be
a
relatively
simple
matter
for
people
to
to
create
simulators
around
it.
B
Just
you
know
somebody
could
just
sit
down
and
write
a
simulator
and
base
it
off
this
and
and
pull
in
all
their
models
from
this
model
compiler
and
that
would
kind
of
open
or
kind
of
stem
or
create
a
lot
of
of
kind
of
innovation
in
the
simulation
industry.
That's
now
being
kind
of
stifled
because
of
the
fact
that
only
the
large
simulation
vendors
can
afford
to
really
provide
the
models
that
that
people
are
willing
to
use.
B
B
Unfortunately,
because
of
resource
constraints,
it's
largely
focused
on
device
modeling
like
the
previous
model,
compilers
there's,
no
real
emphasis
on
functional
modeling,
yet
there's
they're
they're
trying,
but
they
just
don't
have
the
resources
and
they
currently
they're,
not
they're,
ignoring
param
sets.
I
haven't
mentioned.
Param
sets
yet,
but
param
sets
actually
offer
considerable
amount
of
Thomas
for
simulation
development.
So
let
me
just
speak
real
briefly
about
param
sets.
B
Basically,
when
you
go
to
model
when
you
go
to
simulate
a
device,
you
need
two
things:
you
need
the
model
equations
that
describe
that
device,
and
then
you
need
a
set
of
model
parameters
for
those
model,
equations
that
that
kind
of
fit
that
those
equations
to
the
actual
device
that
you're
trying
to
simulate
those
they're
those
model
parameters
are
there
can
be
a
very,
very
large
number
of
them.
Hundreds
of
model
parameters
to
tailor
the
device
to
a
particular
tailor
the
model
to
a
particular
device
and
currently
in
verilog.
B
Param
sets
was
an
attempt
in
miralog
a
to
provide
a
mechanism
in
the
verilog,
a
language
to
specify
these
parameters
model
parameters
now
that
by
itself
seems
very
pedestrian,
but
if
you
actually
take
those
two,
if
you,
if
you
kind
of
recognize
that
param
sets,
are
there
and
the
model
equations?
Are
there
when
you're
building
the
model
compiler
and
you
compile
them
together,
so
you
grab
the
parameters
and
the
equations
compile
them
together.
B
Then
you
can
end
up
compiling
out
a
lot
of
the
and
pruning
the
expressions
and
and
making
creating
compiled
models
that
are
much
more
efficient
than
if
you,
if
you
wait,
which
is
kind
of
the
traditional
approach
in
spice
today,
is
to
have
compiled
set
of
equations
without
the
parameters
and
then
provide
the
parameters.
Now
these
parameters
have
to
be
carried
around
in
the
simulation.
They
have
to
be
re-evaluated
each
time.
It's
there's
a
lot
of
overhead
that
goes
along
with
this.
B
A
big
part
of
that
is
the
memory
footprint
like
carrying
around
all
these
numbers,
just
really
kind
of
thrashes
the
cache,
and
so
it
makes
it
it
kind
of
slows
down
the
considerably
slows
down
the
model
evaluation.
So
if
we
can
kind
of
optimize
a
lot
of
these
things
out
create
a
device
with
a
much
smaller
memory
footprint,
then
we
could
have
much
faster
simulation.
B
B
I
think
that
would
both
that
would
really
kind
of
make
it
a
force
in
the
industry
and
give
it
kind
of
a
momentum
that
it
wouldn't
have
if
it
if
it
were
just
able
to
kind
of
replicate
what
other
signal
is,
but
if
it
could
go
beyond
what
other
simulators
could
do
as
far
as
performance.
That
would
really
make
it
a
force
in
the
industry.
B
The
simulator
vendors
themselves
are
are
in
it
for
them
their
own
cells,
so
they
just
want
to
in
their
mind,
all
they
need
to
do
is
is
to
get
a
partnership
with
The
Foundry
get
models
that
are
compatible
with
their
simulator
and
as
far
as
they're
done
that
as
far
as
they're
considered
they're
done.
That's
that's
beautiful,
but
that
just
kind
of
leaves
these
other
simulator
venues
out
out
in
the
you
know,
struggling
to
get
that
same
kind
of
compatibility
and
it
just
kind
of
completely
locks
out.
The
open
source
simulator.
B
D
Sure,
yes,
Rob
Bain,
so
just
curious
with
open
source
simulators
and
what
is
the
engagement
level
with
foundries?
You
know,
for
example,
like
tsmc
I
guess:
I'll
use
that
as
the
the
poster
child
so
to
speak,
but
are
they
amenable
to
working
on
this
because
they,
of
course
have
to
devote
quite
a
bit
of
effort
in
terms
of
modeling
and,
of
course,
for
advanced
process
Technologies
the
Silicon
development
companies
often
have
internal
resources
as
well,
so
I'm
just
curious
what
the
stance
is
with
Eda
suppliers
on
that
excuse
me.
D
Yeah
in
general,
this
seems
to
be
a
problem
right
I
mean
it's
a
it's
I
attended
a
seminar
on
machine
learning
and
Eda
a
week
ago,
Friday
and
it's
the
same
kind
of
thing
relative
to
that
people
are
just
extremely
sensitive
about
their
data,
whether
you
know
going
to
simulators
or
into
Eda
applications
or
into
machine
learning
models
right.
So
it's
kind
of
a
a
problem
across
the
board.
So,
okay,
thank
you.
E
And
make
a
quick
comments:
yeah
I,
think
Rob,
you're,
you're,
correct,
but
I
think
the
biggest
problem
with
The
Foundry
Now
problem
number
one
is
the
trust.
As
Ken
said
right.
They
want
to
be
sure
that
the
model
that
they
deliver
simulate
correctly,
because
if
the
stuff
is
not
simulated
correctly,
the
customer
can
complain
and
they
will
have
to
throw
away
wafer
and
maybe
ask
for
refund
Etc.
E
So
the
real
trust
is
number
one
income
of
cost
and
then
the
second
element
is
what
you
said
about
security
or
whatever
you
know,
but
in
that
case
the
compilation
of
the
model
actually
already
can
help
in
the
direction
as
well
so
I
think
if
the
verilogy
model
would
be
trusted
can
be
trusted,
then
it
can
be
really
be
revolutionary
in
that
direction.
In
my
opinion,
anyway,
that's
my
comment.
Thank
you.
B
Yeah
from
my
personal
experience,
what
I
found
is
that
if
you're
not
the
one
of
the
top
simulation
vendors,
if
you're
not
bringing
in
a
lot
of
Revenue
to
The
Foundry,
then
they
won't
talk
to
you
either.
So
even
the
vendors
don't
get
much
access.
Only
only
kind
of
the
primary
vendor
gets
a
lot
of
that.
B
All
right,
so
let
me
just
kind
of
finish
my
last
slide
and
then
and
then
we
can
kind
of
move
on
to
Pascal.
So,
from
my
perspective,
openvaf
is
the
first
project
in
25
years
that
that
has
the
foundation.
That's
strong
enough
to
really
address
these
issues
and
I.
Think
if
given
some
more
funding,
they
could
provide
a
behavioral,
modeling
capabilities,
kind
of
functional
level,
modeling
capabilities
in
the
language
as
well
as
this
highly
optimized
models
based
on
param
sets.
B
That
would
be
powerful
enough
so
that
they
could
conceivably
become
basically
the
standard
model
compiler
for
the
entire
industry,
not
only
the
open
source
simulators,
but
also
for
the
commercial
simulators
I.
Don't
think
the
commercial
simulators
have
a
really
good
model.
Compiler
Solutions
I
think
they.
B
If
we
want
to
stay
in
the
analog
design
game,
we're
going
to
need
to
do
something
to
kind
of
make,
to
kind
of
bring
excitement
back
to
the
field
and
to
make
it
easier
for
other
people
to
kind
of
join
in
and
I
think
that
open
source
simulators.
You
know,
that's
not
everything,
but
it's
kind
of
an
important
component
to
it.
B
E
Man
I
want
to
write
a
little
bit
on
the
question
I
see
like
in
the
in
the
chat
and
that
Rob
and
Ken
has
been
first
of
all.
I
agree
with
Ken's
opinion
here,
maybe
I'm
a
little
bit
biased
in
the
direction
but,
for
example,
not
for
the
analog
tool,
but
for
the
digital
tools.
I
have
seen
some
document.
The
European
community
of
the
European
chip
type,
where
they're
trying
to
build
their
resilience
they're
actually
pushing
a
lot
through
openness
in
the
order
in
the
future.
You
know
you
design
a
stay
cheap
with
the
company.
E
You
need
to
be
able
to
take
the
chip
again
with
another
tools,
company,
etc,
etc.
So
there
is
really
a
lot
of
push
in
that
direction
for
resilience.
I
would
say
not
for
the
analog
and
this
model
by
this
for
digital
design,
So,
eventually,
I
think
it
can
be
pushed
in
the
direction
as
well.
So
anyway,.
A
F
Yeah
one
of
the
things
is
that
there's
a
legacy
approach
of
like
building
like
compact
models
and
then
shooting
parameters
for
those
for
to
plug
into
the
simulator,
but
that's
partly
because
everybody
used
the
compiled
simulator
and
the
models
are
compiled
in,
but
there's
no
real
reason
to
compile
the
models.
In
these
days
you
you
can
just
make
a
model
directly
from
the
the
Fab
data
you
have,
and
you
know
they're
the
the
simulator
can
do
that
for
you
to
some
extent
but
optimizers
and
just
give
you
a
model
for
what
your
process
says.
A
A
F
F
For
this
simulator,
you
know
it's
just
a
big
data
problem
and
and
then
for
a
lot
of
the
stuff
you
you're
looking
at
like
standard
you're
looking
at
things
which
are
circuits
which
you
can
build,
and
you
want
to
meet
behavioral
models
for
blocks
rather
than
devices
a
lot
of
the
time.
I
mean
device
level
simulations
like
horribly
slow,
but
you
know
the
Block
Level
design.
G
D
A
C
Can
you
can
you
see
my
screen
now
yep,
perfect,
okay,
then
yeah
so
I
I've
prepared
to
talk.
You
know
now
about
open
VFS
itself:
I'm
Pascal
culture,
I'm,
the
primary
author
of
openvf
I'm
kind
of
going
to
present
what
openvf
can
already
do
today,
because
it
can
already
do
more
than
existing
compilers
and
then
kind
of
give.
You
know
an
outlook
on
development
where
we're
hopefully
had
it
in
the
future,
for
the
starting
for
motivation
can
really
provide
it.
A
really
good
introduction
here
so
I
just
keep
this
brief.
C
You
know
very
lock
a
is
already
the
industry
standard
for
the
Distributing
device
models
and
standardized
by
the
CMC,
but
also
has
great
potential
for
Behavioral
and
functional
modeling
of
an
existing
vertical,
a
compilers
or
not
just
app
to
just
not
up
to
the
test.
They
are
slow.
They
are
hard
to
use,
especially
the
existing
open
source.
A
compilers
are
very
difficult
to
set
up.
C
You
know
in
in
the
past,
the
old
payload,
a
integration
and
NG
spice
required
you
to
recompile
the
entire
simulator
and
change
the
simulator
Source
Code
by
hand
to
get
in
a
new
model.
So
really
cumbersome
and
even
worse
the
models
they
output,
the
candle.
We
mentioned
that
they're
inefficient
they're
two
times
slower
than
if
you
write
them
by
hand
and
okay
also
only
cover
a
small
part
of
the
language
than
that
again,
as
Ken
mentioned,
the
functional
part
of
the
language
standard
is
often
neglected
and
they
are
simulator
specific.
C
So
every
simulator
needs
to
kind
of
go
through
the
effort
again
of
building
their
own
verlocker
compiler
and
that's
kind
of
where
we
build
openvf
at
semi
mods.
To
address
these
problems
right
now,
openvf
is
released
publicly
open
source
and
a
GPR
license
it's
integrated
into
with
ng-spice,
so
you
can
right
now
use
it
to
simulate
with
NG
spice
and
the
old
verilog.
A
supporting
in
g-spice
has
been
deprecated
so
and
will
be
removed
in
one
of
the
next
releases.
C
So
we
have
will
be
the
only
barricade
integration
in
g-spice
and
the
focus
has
been
easy
setup.
So
you
don't
need
to
don't
do
some
complicated
steps.
You
just
need
to
download
the
binary
funnel
from
our
website,
compile
your
model
and
put
one
statement
in
your
netlist
and
you're
good
to
go.
Another
Focus,
then
kind
of
going
the
same
direction
has
been
good.
Ui
I've
just
put
an
example
down
here,
so
we've
been
inspired
by.
C
You
know
traditional
compilers
for
traditional
programming
languages
like
C
by
the
error
messages,
so
that
it's
really
you
can
tell
what's
going
on,
whereas
traditionally
with
where
like
a
compilers,
you
often
didn't
really
have
a
clue.
What
was
even
going
wrong.
Your
simulation,
just
didn't
work
and
kind
of
this
ease
of
setup
has
already
spawned
a
lot
of
activities
around
NG,
spice
and
openvf.
C
A
compilers
itself
works
like
a
c
compiler,
so
it's
really
inspired
by
the
field
of
compiler
Construction
in
computer
science
and
that
allows
us
to
achieve
some
I
think
really
good
results.
So,
firstly,
compilation
speed
I,
mentioned
before
traditional
very
compilers
are
quite
slow,
and
here
you
can
see
these
are
just
some
example
models.
You
can
do
the
same
tests
for
most
other
models
where
you
see
the
compile
time
and
seconds
between
openvf
size,
adms
and
ads
inspector.
C
These
are
just
the
compilation,
speeds,
not
simulations,
and
you
can
see
it's
10
times
faster
across
the
board,
really
and
even
more
in
some
cases-
and
you
compare
these-
you
know
open
VF
as
a
open
source
tool,
and
these
are
tools
with
very
expensive
licenses,
but
even
more
crucially,
you
know
you
compile.
When
you
compile
a
model,
you
often
run
simulations
very
very
often
and
what's
crucial
is
that
these
simulations
are
fast,
and
here
openvf
also
has
advantages.
C
Now.
The
comparison
here
is
a
bit
more
difficult,
because
comparing
simulation
times
across
simulators
is
a
bit
more
nuanced
and
challenging.
So,
instead
of
comparing
the
simulators
directly,
what
we've
done
is
we've
done
a
relative
comparison,
so
we
use
the
built-in
model
built
into
the
simulator
as
a
Baseline
and
then
the
verilog
a
model
compiled
with
the
respective
very
like
a
compiler
kind
of
as
the
benchmark.
C
So
in
this
case
we've
used
the
Visa
master
I
model
it's
been
around
for
a
couple
of
decades
been
and
has
been
around
in
g-spice
for
a
long
time.
It's
very
well
optimized
and
which
you
can
see
here
is
the
model
compiler
of
open.
Bf
is
only
six
percent
slower,
it's
only
marginally
slower
than
the
you
know.
C
Op
model
that's
been
optimized
for
decades
and
hundreds
of
human
hours
put
into
it
and
inspector
by
comparison
to
build
a
model
performs
well,
but
the
verlock
a
model
is
170
percent
slower,
so
they
do
not
have
they
do
not
have
this
Advanced
optimizations
with
openvf?
Yes,
now,
how
do
we
accomplish?
This
can
already
alluded
to
this
again.
C
You
know
when
you
handwrite
a
compact
device
model
into
a
simulator.
You
need
to
do
this
laborious
test
of
hard
coding.
All
these
symbolic
derivatives
into
the
simulator-
and
you
know
it's-
it's
labor
intensive.
It's
cost
derivative
extremely
expensive
and
I've.
Just
have
one
interesting
data
point
to
add
here:
70
of
the
C
code
in
NG
spice.
C
So
if
all
of
these
were
implemented
in
NG
spice
as
well,
then
the
the
percentage
which
would
rise
significantly
and
openvf
is
able
to
come
this
close
to
handwritten
code
and
really
therefore
displace
the
need
to
write
models.
By
Hand
by
using
Innovative
algorithms
to
generate
derivatives
more
or
less
similar
to
human
words,
because
openvf
Works
more
like
a
c
compiler,
we
have
a
much
deeper
understanding
of
the
model
source
code
in
the
in
the
compiler
and
can
therefore
apply
heuristics
to
more
efficiently
generate
these
derivatives.
C
C
In
fact,
the
osdi
is
already
integrated
in
g-spice
and
has
been
designed
in
close
cooperation
with
the
size
team
and
therefore
it's
all
definitely
also
possible
to
integrate
into
size
in
the
future,
which
has
an
internally
very
different
architecture
than
NG
spice.
So
really
openf
also
opens
up,
as
can
already
mentioned,
the
device
models
to
any
simulator.
In
fact,
I've
also
been
already
had
people
reach
out
to
me,
which
have
researched
project
simulators,
which
they
are
now
using.
C
Openvf
as
the
tool
for
getting
the
models
into
the
compiler,
so
they
are
already
research
projects,
they're,
not
public,
but
research,
projects
that
are
integrating
this
oci
interface
into
their
research,
simulators
now
kind
of
at
the
last
Point
kind
of
giving
a
road
map
where
we're
headed
in
the
future.
So,
as
can
mentioned
right
now
for
our
initial
release,
you
know
you
have
to
start
somewhere.
We
started
with
focusing
on
compact
modeling.
C
We
can
already
compile
all
CMC
industry
standard
models
and
many
more
the
one
thing
kind
of
missing
right
now
from
the
compact
modeling
perspective,
noise
simulations,
which
is
what
we're
working
on
right
now,
which
we're
hoping
to
release
this.
Sometimes
here
and
then
you
know,
in
the
long
term,
we
have
some
more
Ambitions,
because
openvf
is
internally
very
flexible.
It
has
very
robust
architecture.
That
means
it
has
been
designed
from
start
with
the
functional
modeling
aspects
in
mind
as
well.
C
So
in
the
future
we
also
want
to
integrate
the
functional
modeling
features
and
behavioral
models.
The
modeling
features
into
open
vaf,
so
that
also
behavioral
modeling
referralog
a
becomes
possible
and
then
even
further
down
the
line.
You
also
want
to
support
nested
modules
and,
as
can
mentioned,
parent
set
statements,
and
what
these
really
allow
you
is
to
Define.
You
know
a
good
example
is
the
skywater
BDK
the
skywater
pdk?
Is
this
huge
amounts
of
spice
net
lists
and
essentially
what
these
parent
sets
allow?
C
C
If
you,
you
know,
start
having
more
widespread
use
of
it
like
eight,
and
you
also
want
better
tuning
for
a
log
a
that
means.
You
know
tooling
we're
used
to
from
other
programming
languages
like
a
formatter,
an
IDE
which
gives
you
Auto
completions
in
your
editor
or
linter,
and
in
fact
all
of
these
tools
already
have
like
Baseline
small
implementations
in
openvf.
It
was
designed
from
the
start
to
be
used
as
an
IDE.
C
It
was
we
already
have
a
linting
framework,
so
we
already
have
a
couple
warnings
that
pop
up,
when
you
do
something.
That's
not
technically
narrow,
but
might
not
be
intentional
and
really
expanding
on
this
and
building
a
language
service.
What
could
be
done
down
the
line
and,
of
course,
I've
LED,
this
kind
of
as
last
point,
but
would
also
be
great
to
you-
know,
integrate
with
other
simulators.
Do
projects
like
that
into
a
push.
C
C
You
know
we've
invested
into
this,
but
it
requires
funding
because
we
are
do
not
have
the
resources
to
alone
to
push
all
these
features,
but
we
think
that
openvf
can
bring
significant
value
out
to
the
industry
and
therefore
hope
that
we
might
be
able
to
find
partners.
Thank
you
for
your
attention.
Do
you
have
any
questions?
Anybody.
B
You
know
I'd
like
to
make
a
an
observation,
and
that
was
there
was
one
particular
slide
that
Pascal
presented
where
he
showed
the
verilogy
model
like
a
little
block
diagram
or
the
verilogy
models
are
on
the
side
and
then
they
were
compiled
and
they
got
stuck
in
this
open
Library.
The
source
Library
not.
B
I
forget
what
I
called
it.
Basically,
it
was
a
shared
Library
and
then
that
could
interface
to
all
the
various
simulators.
So
once
the
whole
param
set
stuff
is
working,
the
the
foundries
can
create
those
shared
libraries
and
then
distribute
them,
and
then
any
simulator
could
use
them.
So
that
would
really
think
I
that
would
that
would
that
would
basically
compile
the
the
pdks
and
so
that
they're
not
visible
to
the
outside
world,
making
the
foundries
very
comfortable,
but
it
also
opens
them
up
to
all
the
various
simulators.
E
A
A
A
A
G
G
So
now
we
can
use
this
compiler,
together
with
Andreas
bias,
to
run
our
simulations
without
needing
to
buy
licenses
for
Cadence,
and
so
now
we
can
do
everything
we
wanted
this
tool
to
do,
but
we
see
so
much
potential
and
we
really
don't
have
the
resources
to
go
further
and
I
would
be
super
happy
if
we
can
go
further
because
it's
a
really
like
it
can
help.
Everybody
I
think.
A
A
F
C
C
E
G
C
C
A
C
G
B
C
It's
like
I
have
implemented.
Documentation
can,
of
course,
always
be
improved,
but
there
is
like
a
50
page
PDF
or
something
so
so
we
already
have
quite
a
bit
of
documentation
and
in
fact,
like
I,
said
a
researcher
reached
out
to
me
that
he's
using
this
document
and
has
a
working
prototype,
so
it
can
be
used
to
to
to
to
integrate
into
simulators
excellent.
B
C
I
I
We
have
some
solid
funding
internally,
but
you
know
we
have
lots
of
internal
priorities
that
have
been
pushing
things
and
that
we,
you
know
this
kind
of
the
development
side
just
keeps
sliding
and
sliding,
and
it
it
has
become
a
source
of
anxiety
for
me
that
we
keep.
It
just
keeps
pushing
forward.
So
you
know
I,
you
know
we
I
I
really
feel
like.
We
want
to
start
pushing
this,
though,
and
it
makes
it
hard
as
I.
Think
Eric
has
mentioned
too,
that
you
know
we.
I
We
are
very
hamstrung
in
terms
of
what
we
can
do
with
our
funding
in
in
terms
of
interactions
with
outside
groups.
You
know
zeiss's
release
open
source,
but
you
know
we
don't
have
quite
the
it's
very
hard
for
us
to
accept
external
contributions
into
the
code.
There's
lots
of
impediments
to
that
because,
due
to
certain
internal
drivers,
so
you.
E
I
You
know
just
because
we've
it's
kind
of
been
quiet
on
our
front,
I
guess
doesn't
mean
the
interest
isn't
there
and
that
you
know
we.
You
know
we.
We
are
very
driven
to
do
this
and
you
know,
maybe
you
know,
like
I
said
you
know
having
this
exchange
with
Aditya.
Maybe
we
can
change
some
of
our
where
people
are
focused
a
little
bit
but
anyway,
that
that
was
the
comment.
I
guess.
F
I
H
H
G
I
left
this
question
is
for
us
at
semi
mod,
so
the
models
we
extract
are
always
CMC
models
where
you
have
a
verilog
a
code
and
for
some
of
the
really
old
PC
models
there
are
like
verilog.
A
files
sent
by
emails
that
are
not
official
I
have,
for
example,
a
psm4
or
there's
another
model
from
a
keysight
for
Indian,
phosphide,
hpts,
but
I
know
we're
not
actively
working
on
that
I
think
it
might
be
nice
to
have
it,
but
the
funding
for
these
modeling
activities
is
really
not
where
it
should
be.
C
G
H
G
I
Just
just
real
quick
since
Sandia
is
a
member
of
of
the
CMC
I
have
a
little.
You
know.
I
occasionally
get
an
Insight
on
things,
but
just
just
a
comment.
It
actually
came
up
in
a
recent
meeting
about
you
know.
Would
it
be
worth
trying
to
port
to
be
some
for
to
verilog
a
and
basically
the
answer
was
well.
You
know
from
the
developer
side,
you
know
if
you
want
to
spend
the
money
on
that.
You
know
it's
possible,
but
you
know
it's
considered
a
legacy
model.
It's
pretty
much.
I
I
E
Hey
Jason,
so
some
of
the
open
source
communities
based
on
a
little
bit
older
Technologies
like
for
example,
now
we
have
skyworks
skywater
and
then
Global
foundries.
So
this
model
will
be
measured,
extracted
using
Boulder
models
right
So.
Eventually,
when
they
come
in
I,
don't
think
those
Foundry
you
know,
will
probably
re-characterize
their
technology
using
like
different
models,
they've
already
done
so
kind
of
in
the
corner,
and
we
need
to
kind
of
develop
all
the
models.
So
we
are
ready
for
this
technology
when
they
get
released.
G
A
Yeah
yeah,
so
I
think
we
are
about
an
hour
now,
but
I'd
like
to
catch
up
with
you
Marcus.
If
you
have
some
time
after
this
meeting
and
next
week,
we'll
have
a
discussion
about
specs
extraction
from
Magic.
Some
of
my
students
were
working
on
a
tape
out
in
GF
180,
and
we
just
noticed
that
you
know
the
extracted
results
are
completely
off
compared
to
close
tools
and
I.
A
Think
that's
a
good
time
to
discuss
with
with
magic,
which
is
the
only
tool
to
use
for
extraction
right
now,
pix
extraction
in
open
source
community
and
see
what
are
potentially
the
work
that
can
be
done
there
or
what
are
the
possible
tools
that
we
could
use
in
addition
to
Magic?
So
if
anyone
has
any
comments
or
any
suggestion,
please
reach
out-
and
you
can
arrange
that
but
yeah
thanks
Ken
thanks
so
much
for
attending
and
Pascal
really
nice
talk
and
hopefully
we'll
get
some
funding
for
semi-mod.