So it's a real pleasure to have.

My friend and colleague Professor and
he me with us today.

Because I got his a bachelor's
degree in electrical engineering

from Sharif University and then
continued to work as a design engineer

in Iran before coming to Georgia Tech
to get his master's and Ph D.

in electrical computer engineering He
then worked as a research engineer at

the M I R C For those of you who remember
the precursor precursor of the I even.

Before becoming a faculty member in
the C.E. in two thousand and eight

where he's now a full professor
also a associate director for

and in C.R.I. the national and
technology coordinate infrastructure for

computation and modeling he's won
a number of awards including the S.F.

career Ward as well as awards from S.R.C.
Tripoli and Georgia Tech for

teaching he last gave a presentation and
then I went Tech in two thousand and

nine so there are probably very few
of you who remember that I was there.

But I'm sure this will be very different
stuff so I'll turn it over to a zone.

Thank you for the night and
thank you very much for coming.

My pleasure.

Couple of years and so.

Without giving too much introduction we've
seen all of the talk about moves law and

how everything has been scaling down
to smaller and smaller dimensions and

the level of complexity constantly
increasing one of the things that I

sometimes like to look at is the looking
at chart which was in the log scale who

looking at the chart in the lean
years scale it sometimes

gives us a better picture of how
rapidly things are evolving and

at the heart of all this
exponential growth has been.

It.

Has been the.

Metal oxides semiconductor
transistor where you marginally the.

Potential barrier in a device
in order to control the flow of

electrons going from source to drain.

So when you applied no voltage to the gate
the energy Bear year is this high and

therefore you block all the electrons
as you increase the gate will to you

push the energy bare your down and
then you allow electrons to go from

source to drain but
in reality this doesn't work like.

Complete an ideal switch because
of the the way that energy

of electrons is distributed so if you look
at the distribution of probability of

the states in the source it
follows this very distribution and

tours this side the probability
drops exponentially so

this becomes almost like
an exponential which means that as you

gradually bring down
the the energy bare year

the number of electrons that have enough
energy to groove or this barrier increases

exponentially rich means that if you
look at the current that you get

as a result of this modulation of
the energy bill year it increases.

In an exponential fashion so for
every sixty million that you shift here

you get one order of magnitude
increase in dream current So

this picture looks very nice and

you have a really good energy Bear year
it's relatively flat inside the channel

no matter how much will drain you
apply this valid for a big divide.

But gradually as people have scaled
down the dimensions of the device

things have changed because
the control that gate has on

the channel is through this capacitance
but there are other capacitance is

involved as well there is some capacitance
to ground to do this substrate and

source and there is a capacitance
between the channel and

the drain which means that
the this potential height or

this energy profile is not
controlled purely by the gate but

also by the drain so
for a small device for

a short channel the wise the picture
looks something like this now if you

look compare this carefully with the case
over here first of all the peak of this

energy barrier is now smaller because of
the effect of the drain over there so

more electrons can go over to bear
years of the leakage increases

also right at the top here you
see that the length of the energy

barrier is small so
some electrons can tunnel through and

go to the other side so both of
these effects increase the leakage.

And the other thing is that because
these other capacitors are also involved

the change in gate world just
doesn't directly translate here so

that means that sixty
million per decade is

constantly increasing as you
make the device smaller so.

With this to summarize this affects what
happens is that this device becomes more

and more leak so the only solution
to this problem is to increase

this capacitance and make the control
that gate has on the channel better and

better and that's has been what industry
has been doing for many years so

if you look at the evolution of and
was transistor.

Started by playing or playing.

Structure then and
it was using silicon dioxide as

the insulating layer then in order
to further increase the capacitance

without making the oxide too
thin this switched to high K.

dielectric materials to change
the gate material to metallic gates

and they kept improving this
device then they went to

three dimensional geometries so
that instead of trying to reach

to the channel only from one side in
the Fin FET structure which is a three D.

structure you try to reach
the gate from three sides and

moving further people are thinking about
nano wires the channel is surrounded

by the gate from all sides to a gate
all around type configuration

either in a horizontal fashion
ordinary vertical fashion to the good

thing about the vertical sorry this one
is the vertical This is the ladder or so

the good thing about the vertical
dimension is that you can actually make

the channel lengthen longer without making
the footprint of the device larger So

there's all these changes that are
happening and this has been the evolution

and probably these will come into
production in the next few years and

the industry has a clear roadmap for
the next five or

ten years and people are now talking so
if you look at this geometry this is for

something like fourteen nanometer this is
the dimensions that people are dealing

with these are really
really tiny dimensions.

And the push is to go
down to five nanometer or

three nanometer but at some point
you have to change the paradigm

you cannot keep scaling things and
trying to do the same thing that

brings us to the research that they want
to talk about today which is trying.

To think about completely new
ways of doing computation and

try to use state variables
other than electronic charge or

try to reinvent the device and
also try to come up with new way of.

Building circuits and systems out of
the same devices or out of new devices so

to motivate that if you look throughout
the history been there multiple times so.

All the days there were computers
based on mechanical switches then.

Electro mechanical sutures you know relays
will be very used to do computation then

vacuum tubes then transistors individually
and then integrated circuits and

from bipolar junction
transistors to mosques and

we've been working with this for many
years and this mean this much scaling but

it doesn't mean that this should go for
ever so there may be new devices

that can can can continue this trend for
the years to crawl.

Now when we want to do
research on new devices and

new circuits then many things will change
at all levels of abstraction and if

you look at the options that are out there
and people are studying there are a vast.

Number of options at each level
of abstraction at the very bottom

layer which is the materials that
are being used people are looking at

one dimensional materials
that are carbon nanotubes or

looking at to do materials like Graffy
like other to the materials that have

come out in the in recent years
people are thinking about using

fair magnets to do computation and I'll
be talking a lot about federal magnets or

spintronic for doing computation and

recently people are even interested in
magnets because thirty magnets have.

A aura switching time constant around
one or two or three nanoseconds But if

you are able to grow to enter for magnets
they can switch really really fast.

There's a push to use federal troops and
there's a lot of interest there there

are even people thinking about using
strain taking advantage of peas or

electric devices and apply string and
make switching happen that way and

I would be talking about these
devices in the consequence of

spintronic devices and
it's making all sorts of.

Mr devices and so forth so the number
of options at the material is quite.

Elaborate and
there are many options out there but

then once you look at all
these materials there are so

many physical phenomena that you
can potentially you do computation.

In the past it was this field effect.

And the effects that are being
used there are well known but

now people are talking about joy in
magnitude resistance effect basically

if you have two magnets and the two
magnets are aligned the resistance is lure

then the case that the two magnets are and
so that the giant magnitude exists

if you put a tunnel layer in between
the two magnets then you can even.

Further enhance the change in resistance

when the two magnets are versus
the two magnets being anti per.

And the important concept that
is being used is Magni to

many of the magnetic materials when you
are price strain to them their magnetic

properties change so
you can have a magnet that is in plane so

the two stable conditions are in play but
as soon as your prize strain this

magnet may go out of play so
this condition may become out of play

again I will be talking about this effect
and how one can potentially use this for.

For doing computation fair electorate
existence in fair electric devices

if you put two metallic contents
of different kinds and sandwiches

a year depending on the orientation of the
layer you may have higher resistance or

lower resistance so that's
a fundamentally new concept and a new.

Way of doing computation or
new way of storing information because

computation has to parse
doing the computation and

storing information so
both of them are really important so

I'm not going to go through the rest
of the physical phenomena that people

are looking at but
do number of options are quite.

Large.

Then once you have.

A nominee on a new material that you can
utilize then different ways that you

can make a device out of it you can do the
computer using these new physical concepts

and some of them are still
Field Effect devices evolutionary

type of improvement over what we half and
some of them are completely new like

magnet to magnetic tunnel junction
devices for memory applications or.

Junctions negative capacitance
devices are the case

when you insert a for
electric between the gate and

the normal insulator
then you may use some.

Spinning wave type computer which I would
be talking briefly of so there are many

options at the device level and
then once you have new devices you have

many options how to use these devices to
build the circuit and do the computation.

The way that we do competition is usually
through functions and you make an adder.

You do computation that way but

with many of these devices it may not be
the best way of doing things you have

some new options you can make majority
against with them you can make new

type circuits with them and as you
will see some of these devices perform

much better if you go to these
nontraditional types of circuits and

then at the system level both
traditionally we had C.P.U. use the G.P.U.

but now people are trying to break the
barrier between the computation part and

the storage part so
instead of doing competition here and

trying to fetch data and restore data
into memory what if you bring memory and

computing into the same place or.

New ways of doing competition with
approximation or shine inspired computing

and at the end software and operating
systems have to look at everything and

that's the higher level
of abstraction now.

You can't really do research any of these

years on any of these topics and not
look at what is above and what is below.

Real comprehensive study
is needed if you want to

come up with something that is
completely revolutionary and

can change the computing paradigm so
in our group we try

to cross the boundaries of
these levels of abstraction and

obviously we cannot do everything but
if you look at this

chart this the size of the boxes
here is how much effort we put

on each of these so we do the smaller
amount on materials mainly

trying to develop compact physical models
with captured to keep parameters and

the key properties of various
emerging materials and

it's not just emerging materials even
materials that we are very familiar

with like silicon if you try to transport
spin through Silicon That's a new

phenomenon that is not understood so you
need to develop new models there as well.

And for devices try to develop compact
physical more those that capture

the physics of the devices and also try
to come up with better devices understand

whether the limitations and
how they can be overcome and

how we can improve them and then once
we have these compact more of those and

sit at level models we try to analyze some

circuits based on these devices try
to find what is the best circuit for

each kind of device and benchmark them and
compare with each other and

try to see at the city level
what are the limits that

devices are imposing and how we can
try to mitigate those limitations and

at the system level we try to come up
with methods that in a fast way you can

take from materials and devices and
could go to the system level and

see what is the performance of
the of the system what is the to.

Just the number of instructions per
second that you may be able to execute.

So.

You can't do all of this on your own and
the key for us has been collaboration so

we've been collaborating with many
people on the industry's side.

Being part of S.R.C. working closely
with the sponsors within S.R.C.

we have close collaboration with Intel
over the past five years we have published

a lot together and we've been working
with some of the colleagues at Intel and

we have a close collaboration with
I'm Eric that's a big institute in

Belgium which most of
the semiconductor companies you know.

Tool vendors material suppliers
are collaborating there so

working with them has been very
instrumental in our research.

We have started some new collaboration
with our especially at the system level

and also with Micron and memory
devices but another big effort within

our group in the past three years has
been this benchmarking research because

there are all these centers
that are funded by Sarsi and

NIST which are working on new materials.

This frame and new devices these
centers and these centers and

also try to come up with new circuits so
working with these.

Investigators gives us axes
to real experimental data

many tourists within these centers so
this has been a very productive

collaboration try to benchmark
various materials and

devices and compare their performance
at the circuit and system level.

So this you know they've been.

Marking effort that we took over about
three years ago before that was led by.

And in young intell and this is a summary
charge with shows you if you execute

the thirty two bit edition using different
different device concepts what would be.

The energy and what would be
the delay associated with that so

there are many device options here and I'm
not going to go through those just some

highlights one is this reference point and
this is based on the projections for

the future size of fifty
nanometer this is this was for

the eighteen technology
year this is a relatively

optimistic data point because
the international technology for

semiconductors used to
be quite optimistic.

The the other device is this
the most go voltage This is based

on theoretical projections four
point three wart get all around.

India arsenide nano Warrior do wise and
it's based on atomicity quantum transport

simulations so these two basically
Sure what's the competition

that we have in terms of seam OS
can do and then we're trying to.

Look at other devices and
see where they stand and

this is not about ranking these
devices and see which one is the best

it's because because these are all at
the very early stages of research.

It's too soon to say which one is
the winner which one is the loser but

it's the Google of this approach is
to understand the general behavior

of these devices and to understand what

is limiting their performance why some
of these devices are in the bad corner.

And how we can improve them and
try to push them to the right corner and

try to do to make them competitive so
that's the goal so this is what was done.

Back in two thousand and fifteen and then
since then we've been working on this and

I'm going to show you more of the results.

So after this long introduction

the presentation will be on
spintronic materials and devices.

And then we'll be talking about one form
of the room or fix their kit basically

networks and you see that some of these
devices that were performing very poorly

brilliant function is can potentially do
much better if you go to something that

takes advantage of the physics of
the device and finally summarize.

So for
Sprint tronic device is just a very.

Brief Introduction is that
when you pass current through

a conductor the electrons have
random spin orientation and

as a result there is no net spin in
the current So there is no spin Porter.

But when you pass
electrons through a magnet

depending on the orientation of this
magnet the electrons get polarize around.

Depending on that magnetic.

So the majority of electrons not all of
them but the majority of the electrons

coming out of that magnet will
become aligned with the magnet and

then if you pass this spin
porous current on another magnet

you and if this current is
strong enough then you can

actually rotate your intuition of this
magnet because these electrons have.

Angular momentum that they can transfer
to the magnets and reoriented and

make it aligned with the spin for
ization overlap.

It's so here it.

Is happening and here spin
transfer torque is happening now.

One advantage that these devices have
is that this can be all metallic and

therefore you can have
the low voltage operation and

also magnets can be nonvolatile
can store information

even if you remove power so
these are some of the benefits but

the question is how would you use
this to do computation with it and

how would be what would be the performance
of it and I will use this as a case study

where you see how going through the levels
of hierarchy starting from materials to

devices to circuits will give us some
insight into the operation of these

devices one of the proposed devices
was not was all spin logic by.

Purdue and the idea is that
you pass electrons through

this magnet and polarized electrons and

and inject them into non-magnetic
material here like copper so

here spin accusation is generated
because elections coming from this

magnet gets put get poor as according
to your intuition of this magnet so

then this is electrons real diffuse and.

Move to the other side and then on this
side they will inject a torque on this

magnet and will flip this according to
this so this would behave like a buffer

Now if you connect three of these together
then you can make a majority because

these spin poor areas ation of electrons
that are going to appear here will depend

on the majority of the inputs and then
this one can drive the next stage so this

can potentially be computing pair of them
where this would be the building block

the majority gate communication happens
the same way as the computation and.

The channel materials that you
can use here are any material

any metal like copper or aluminum or
in some cases even graphing.

To analyze this you need to look
at the dynamic of the magnets so

the creation describes this and

you need to look at the interface between
the magnet and the channel of the mixing

conductance matrices you need to look at
Spindrift diffusion in the channel and

then look at the dynamic here and you can
write all these equations in matlab and

try to solve themselves consistently
which would be time consuming and

it would be very messy and
you need to worry about convergence and

if these are going to work at the at

what as electrical engineers we are much
more comfortable with circuit inspires and

do simulations there so
we took all of these physics and

for each one we came up with
a circuit model where or

basically subcircuits which exactly
replicates the function of each.

Each block So
here are this is the dynamic of

the magnet stimulated by these
by this sub circuit and.

These are the interface matrices and
then within the channel

you need to have electrical signal
because this is a conductor and

has a like a normal in a clinic
you have a distributed R.C.

circuit but at the same time
you need to account for X.

Y.
and Z.

components of the spin porous current so

you have these branches which account for
the X.

and Y.
Z.

and then you can just
set it up inspires and

then Spice will handle this on its own
you don't need to worry about the.

The convergence issues and at the same
time it gives us some insight because we

are very familiar with these
circuits these are distributed R.C.

networks so we know how they behave so

well as a sick a designer when you see
some elements like this you can get much

more intuition rather than some Matlab
codes that are solving some equations.

But at the material level you also need
to be careful because material parameters

highly dependent on what kind of
dimensions you're dealing with it

what kind of interfaces you're
dealing with what kind of impurity and

dropping you're dealing with so
part of our work has

been to develop a compact more of
those for some important primary

like spin diffusion land spin
relaxation and we've been calibrating

these models read a lot of experiments so
that will give us some important.

Some confidence in the results one thing
that I want to mention is that for

example if you look you can go and
look it up and

see that what is the spin diffusion Lenten
copper in materials around five hundred or

six hundred nanometer but
you can't use it in a real circuit because

at nano scale electrons get scattered
at the surface at the ground boundaries

manage scale you need to worry
about the size of facts so

this model captures size effect and based
on that gives you the spin relaxation.

Another way of doing computation
is wave computing and

one approach is to use spin waves so
if you have a magnetic layer you

can create spin waves on it basically
these magnets will start oscillating and

you can generate wave and then you
can potentially do interference and

based on interference you can make
a majority cage if the phase of

the inputs are the same there's going
to be constructive interference if

they are out of phase you're going to have
destructive interference and there's been

proposals to do this week
computing based on spin race so

there are some issues about these
the proposals that spin waves are weak and

you need to regenerate them every
time you need to have gain.

If you just want to use
conversion back and

forth between electrical signals and
spin of waves then you need to have

complicated since amplifiers thermal
noise must be accounted for and

these spin waves are nonvolatile
whereas magnets are volatile and

one advantage of magnets is
non-relativistic How can you use

spin waves and build circuits
that are still nonvolatile so

to do that our approach has been to
use magnet to stricture and to use.

Piers to electrics to create
spin create strain and

reorient magnets based on that so
the idea is that.

You have a magnet to electric So

when you apply normal to the magnet
is in place there is no string so

the magnet is in play and
it has to stable conditions in positive or

negative effects as soon as you apply to
the peers electric you create string and

then the magnet may go out of plane if the
design has been done right so with this

approach what we can do is we can take
the state variable of this magnet being.

So the orientation of the magnet
would be our state so

it can be either in positive or
negative X.

in order to read the stage
variable what we do is we

apply the magnet to electric So now
the magnet wants to go out of place but

depending on the initial condition
if it's positive or negative X.

the spin wave that you create is going
to have one hundred eighty degree fish.

Now you have generated this spin with
you have read your state here and

now the receiver needs to detect
the phase of this coming spin if

the where did we do it is the opposite so
on the receiver side.

Where you want to do
the right operation really.

Initially put the magnet
in the out of planes.

But when the spin with
arrives we turn this world.

Now the magnet wants to fall in place but
the direction that it's going to fall

really depends on the defeat of
the spin that it's coming so

this way we have generation regeneration
every time the circuit is nonvolatile and

we can control the direction of flow of
information by clocking in the right way.

So one of the things that
we did was try to see

if this is reliable on the thermal
noise and one thing that you would see

is that on the if you put the magnet
initially at the high energy and

now the spin rate wants to fall this
in either positive or negative X.

you will see that this is not reliable but

the reason is that when you put the magnet
in a high energy state it's going to make

many oscillations before it damps and
it falls on positive or X.

directions but
you can use Exchange spring system or

use the built in string to change
the energy profile here and by changing it

you change the profile here such that this
because this zero point becomes the other

point now this is going to be
a very damped switching and

as soon as it goes into one port it's
going to converge around that port so

it's not going to make so many switches
between positive and negative X.

which you do here and here it's going
to have a deterministic stitching so

this is the simulation that we did try
to see what would be the error rate and

it seems that we can get to really
good error rates for this circuit

so and this is one that we did with I.

Became interested in this proposal and
they wanted to simulate an entire.

Majority gate and recall aberrated with
them and did the simulation and this

nano scale majority get based on spin
waves under simulation side is work.

So this is the summary of
the benchmarking results for

spintronic devices so
these are this is the spin rate logic it's

very low energy it's low because you're
dealing with magnets and magnets or

that they have time constants in demand
of can't reach it so this is a low energy

relatively slow device this is
a magnet to electric M.T.G. device for

the sake of time I'm going to skip
to the main vol device this is using

spin Hall effect of devices that
can be made out of that and

one of the things that you see here
is that how we started from this

original device and
gradually try to make it better

this is the all screen logic that there
showed and depend if this is the in plain

version of it this is the out
of plain version of it and if.

It's become available at room temperature
then this would be to perform

the potential performance of
this device but what you see

is that these devices are very energy
hungry Compared to see Mars and

that's a problem that we need to solve and
we can try to make these materials better

to make it better or
we can try to use this materials for

something other than bullion logic
that would be talking about so.

So feel defective Weiss's there are many
options again I have to skip the details

of these devices and I'll just show you
the latest benchmarking that we have.

These are the so this is the most lowered
the most high performance are shown

here but a few highlight devices this is

a theme this is to lose
of to dematerialize and

you make it tunneling device out
of them the potential to tunneling

happens between the two layers and
you module that kind of searching.

This is base fit which based on
the boards on condensation and

the efforts to demonstrate this has not
been successful this was supposed to be

done by and by layer graphene So
at this point this is just theory and

has not been physically demonstrated.

This is negative differential resistance
device based on tunneling between two

to delay years again but
this new kind of circuits didn't require

special clocking to do that and
also we need to

keep in mind that CMOs itself is a moving
target it's going to become better and

better one up one example is as I
mentioned looking at nano warriors and

either in the lateral case or the vertical
case and we've been collaborating with

benchmarking these and and
evaluating their potential performance and

this is at the level of
a processor an ARM processor and

the conclusion was that
the lateral device was better for

high performance because you can still
apply the strain and get higher mobility

whereas the vertical device was better for
power applications.

Now said that building CMOs in brilliant
applications it's very difficult and

that takes us to places where
we try to take advantage

of the physics of the device and
make circuits that are more efficient so

to do this kind of benchmarking for
nontraditional circuits we had to.

Have some requirements the circuits should
have many applications it shouldn't be

a very niche kind of circuits it
should be something that can be

used to implement many beyond
CMOS devices efficiently so

that you can have
a meaningful benchmarking and

it should have well defined boundaries and
specs so for that purpose

networks are a good candidate
because they can do many

times it's a universe of computers so
it can do any it's a complete.

Function it can do anything that you
want with it and it's very efficient in

doing image processing and associative
memory and tracking targets and.

And at the heart of it is dot product
which is used in many other circuits.

So the way that cellular neural
networks work is that it's a two D.

at a very.

Cells where cells
are connected to each other.

And the dynamic of each cell is
dictated by a equation like this so

this is the state variable of J.

and there is a feedback factor and

there is a linear combination of
all the inputs to the circuit and

the linear combination
of the states nearby and

some biasing condition so
did the normal way of implementing

this by making by using our power amps for
the neurons and

all the synapses which determine what
is the weight is dictated by the.

Operational transconductance.

So basically you can use mosfet or you can
use to fret Superman these amplifiers on.

Steps Up there should devices would be
good because they're going to give you

better gain lower power but

at the same time you could do a sixty and
you can do digital.

Design and for each you have on a six
circuit which does exactly the same so

the digital option is also there and
then we wanted to do

spin tronic implementation of this
in a more efficient great and

if you look at the L.G. equation which
dictates the dynamic of a magnet

it's very similar to the dynamic
of the C.N.N. So the same kind of

equation is applied and
if you think about it if you look at this.

Cell the current that you pass through
this will generate spin current ritual

Suge this magnet now this
behaves like an integrator and

you see that as you increase this current
you are going to get faster searching but

the key is that this
equation is very similar to.

This equation and
we can take advantage of that so.

You can.

Put a reference resistance years so
this is an empty J.D. orientation of this

magnitude determine the resistance of this
device there is a reference point here and

then there is a normal CMOS inverter
rich amplifiers the signal and

drives the nearby cells and
then this is the driving circuit and

the size of this or is this is the true
means what is the weight of each synapse

and this is the output voltage versus
Magna's magnetization direction.

This is how one cell one
completes looked like and

here is the simulation result
taking account of thermal noise and

here we are showing one function
which is noise filtering.

Here is an associative
memory application so

it's supposed to correspond one to two and
three to four and

here you see how one is
being transferred to two so

these are the simulations of all
the magnets within the circuit

now to do the benchmarking the case study
that we looked at is this very simple

associative memory case these numbers
these digits must be corresponded or

transferred to these digits and
we look at various way of doing this

so this is the benchmarking result
this is this network benchmarking.

One of the things that they want to
highlight is this light blue data points

does the spintronic devices so for
example this is something similar to.

Spin diffusion or
all spin logic this is spin hall

device and these devices now
can perform better than C.

MOS whereas if you remembered
the brilliant functions

these devices were many orders of
magnitude worse than c mass and the reason

here is that we are taking advantage
of the physics of the device to do.

The actual circuit which is that.

The the dynamic of the neural networks so

to benchmarking result is completely
different when you do it for

this particular saloon or

network applications compared to the
boolean case that we talked about before.

Now one other.

Explanation of why these.

Spintronic devices can do much
better than this is that if you look

at the circuits here each of
these synapse you're going to be.

With many transistors and

this is an analog circuit is going
to dissipate power all the time.

There is this which is a much
more complicated circuit and

it's going to be dissipating
power whereas here each

synapse is just two transistors and
a magnet and

the neuron is just these two M.P.G.
and just two inverters.

So one other thing about the lunar
new network is that people have shown

that you can use similar neural networks
to efficiently implement convolution or

neural networks and convolutional neural
networks are much more used these days for

many applications that requires.

Machine learning and inference from

a big large set of data so
this is the work that our collaborators.

And Sharon who are at Notre Dame
have done and be sure that

all the leaders in a convolutional
neural network they can be done.

With.

That works and that So
these are for example an example of

those circuits that can be done
via it cellular and it works so

the results that we have over there
are quite come can be translated here and

we're working with it to
benchmark their results for

four convolutional neural
networks based on this so.

Throughout the benchmarking
what Observateur is that

as you are more those become more and
more accurate than you.

For more limiting factors then
the benchmarking results become more

pessimistic but
at the same time if you really think about

what are the issues whether the problems
or the devices and try to modify and

reinvent these devices the data points
gradually move to the right corner but

still it's going to be
quite challenging to be.

In.

A brilliant domain and novel system and

circuit concepts are quite
necessary to utilize

the new properties of these devices and
one example is the cellular

neural networks that have been implemented
these beyond Sue Moss devices.

So.

Before I and I want to acknowledge
the people who have done the work.

Research engine in my group he's done
a lot of work here the current and

former Ph D.

students some of the names are listed here
and they've done a lot of the work and

the results that you saw mostly from
them we have some collaborators that

Intel which we have been working very
closely the the benchmarking effort

is guided by you know
representatives from the four major.

Sponsor companies and all the.

Students within the Star net assets so
thank you very much for

your time and I would be happy to answer
any questions thank you very much.

Questions.

But the answer.

Is.

I would say most of them we have
major material challenges So

for example all the devices.

The.

Street did suffer through
suing are much of what

the theory predicts and in most cases
is due to the traps so those trap

states are limiting the minimum current so
you can't go be long below that.

Due to the material as the D.

fabrication a material product which is
all very strange and some of the magnetic

devices still the the materials
are not as mature.

So I think most of them are suffering you

know they need a lot of
improvement in the material part.

Yes I.

Think it was.

Guys yet the.

Process.

Is there.

So.

That.

Works.

Out OK.

So I'm trying.

To.

Change.

You guys.

In many cases nothing has been
fabricated completely you know for

example if you look at this C.S.L. device.

Pieces have been demonstrated
the physics is well understood but

to integrate the entire device
we're still not there so

most of the changes that you see over
these trends is redesigning the device

itself it's not restarted from
this experimental point and

we said what if we can make this part but
there what if we can make this material

but that it's basically starting
from this original proposal and

looking at OK when you put it in the
circuit what is really limiting you for

example for this case was that the size
of the magnets had to be big because of

the two contacts that you had to make and
now is there a way to break this

device into smaller pieces and reuse the
current that is coming and therefore not.

Needing as much current so that.

Takes you may be here or for
you to for example here you.

You add another layer to funnel
electrons into the magnet saw So

you clicked more spins and funnel them
into the device So these are not.

These are the result of
a redesign of the device

after learning what is limiting them.

Yeah so.

It's become.

Actually nothing more than not and
not always so

sometimes because you are coming
up with a better design.

And you actually make it better and

the error bar should get smaller so
it depends if you become to speke So

if the trend here is just to become too
speculative and say that what if I have a.

Spin relaxation that can be ten microns
then yes that's that's true but

when you are making this device better
by coming up with a better design and

in many cases actually you
make the device simpler then

know the error bar is
not going to increase.

At.

This.

Point.

What with.

The rise right I think that's
a very good point I think for

some of these devices would be great
because some of these devices have larger

error rates so that would be one factor.

Another factor can be complex city of
the device how difficult it would be.

So.