Technology Scaling in the Mobility Era
Sep 5, 2023
"Technology Scaling in the Mobility Era"
Mark Bohr, Senior Fellow and Director of Process Architecture and Integration, Intel Corporation
CNST Nanotechnology Workshop 2013
Thank you, Jim, for that introduction.
I guess since people may be curious about the new Intel
CEO, his name is Brian Kurzanich.
Before being CEO, he was Intel's chief operating officer.
Namely, he was responsible for all of our manufacturing
and assembly test facilities around the world.
So he really comes from the wafer fab side of Intel,
which I think is great.
And he's been with Intel, I think, 30 years.
And I know him personally, so it's a good choice.
A lot of us at Intel were surprised several months ago
when Paul Odellini, our present CEO,
announced that he was retiring, because that was about two
years earlier than the normal retirement age.
So that was a bit of a surprise.
So then the board of directors launched a search.
Didn't want to do just the normal thing of automatically
promoting one of the senior VPs.
But they did both an internal and an external search.
And they decided that the guy they had inside
was better than any of the other candidates.
Of course, I don't know who the other candidates are.
That's not been made public.
So we now have a new CEO.
He'll be taking over from Paul, I think, at the end of May
here, when Paul completes his retirement.
And over dinner last night, I teased Rashid Bashir
about how it's taking him longer
to select a new dean of engineering
than it took Intel to select a new CEO.
So come on, get things moving here.
We have too many committees that we have to go through.
Too many committees.
All right, so on to my talk.
I want to talk to you today about technology scaling
and how it's a little bit different now
in what I call the mobility era than in the past.
So I think you've all seen this graph in one form or another.
It shows how Intel and the rest of our industry
has continued to scale minimum transistor dimensions.
We develop a new generation of logic technology
every two years.
Each generation scales feature size by about 0.7x.
And today, Intel has 22 nanometer products
in volume manufacturing at multiple factories.
And we're working hard to finish the development
of our 14 nanometer technology.
We expect to start shipping the first 14 nanometer products
by around the end of this year.
Oh, and of course, just to remind ourselves,
the reason why we scale transistors
are to achieve three key benefits.
One, to improve performance.
Next, to reduce power.
And also, to reduce cost per transistor.
So those have all been very important benefits
derived from scaling.
So here's a trend going back to the 2 micron generation, which
was around 1978, 1980, sometime in that time frame.
And I'm plotting here two important transistor metrics
or parameters, the gate delay, CV over I,
and switching energy, CV squared.
So during this era, from the 2 micron era
all the way down to about the 0.18, 0.13 micron era,
we were following what I call, what I think most people call,
classical transistor scaling, following the general rules
that Bob Denard at IBM first described in a seminal paper
about 20, 25 years ago.
So in classical scaling, all you do
is you take the basic MOSFET structure
and you scale all of the horizontal and vertical
dimensions by about the same scale factor,
and scale the supply voltage by about that same scale factor.
And when you do that, you get these results,
continual and significant improvements in gate delay
and in switching energy.
So this was working quite well for several decades.
And you look at this trend and you say, OK, great.
Let's just continue doing that.
But why didn't we do that?
We couldn't continue to scale in a classical sense.
Because although we were achieving
significant improvements in gate delay and switching energy
or switching power, I think the dirty secret
is that when you follow classical transistor scaling,
you are paying a price in ever higher leakage currents.
And all three components of MOSFET leakage
were increasing during this time frame,
including the leakage through the gate oxide,
because we were making the gate oxide thinner and thinner.
Junction leakage was increasing, because we
were increasing the dopant concentration
in the substrate in the channel.
And I-off, or sub-threshold leakage,
was increasing, because as we were scaling transistor
dimensions, we were scaling operating voltage.
And as we scaled operating voltage,
we had to scale threshold voltage
to have reasonable performance.
So by the 0.13 micron, or the 90 nanometer generation,
leakage currents had arisen to fairly high levels,
where they were a significant portion of total chip power.
And obviously, that's not a trend
that could have been continued.
So we're really facing the end of classical scaling,
and we had to do something different.
So what we did is we had to start coming up
with some very innovative or even a revolutionary transistor
ideas to continue scaling, innovations
that went beyond the classical MOSFET scaling methodology.
One of the first innovations that we introduced at Intel
was a strain silicon technique, where
we deposited silicon germanium into the source drains.
And of course, silicon germanium has a larger lattice
constant than silicon.
So it imparted a compressive stress
in the channel, which increased hole mobility,
in the case of PMOS devices.
And that strain silicon was a technique,
thus, to improve transistor performance
without having to reduce gate oxide thickness.
And again, we were limited in scaling gate oxide thickness
because of gate leakage problem.
Actually, I'll pause here for a moment.
There are lots of chairs in the front here for those of you
standing, if you want to come up and sit down, relax.
So again, the first innovation we employed,
a non-classical technique, was strain silicon.
We continue to use and improve strain silicon
on all of our subsequent generations.
And then in 2007, on our 45 nanometer technology,
we had to come back and address the gate leakage
problem with the SiO2 gate dielectric.
And we replaced SiO2 with a high-k material,
hafnium oxide-based dielectric, and combined
that dielectric change with changing the gate electrode
from doped polysilicon to special metal gate materials.
Different metal material for PMOS
and a different one for NMOS, each with a work function
chosen, optimized to get the right threshold
voltage for the NMOS and PMOS devices.
And so we now continue to use a high-k metal gate
and improve it as we move forward to the newer process
generations.
And then our most recent innovations
on our 22 nanometer technology was the introduction
of a tri-gate or a FinFET transistors.
So we're now using that for 22 and beyond.
So with the introduction of those innovative or
revolutionary transistor technologies,
we were able to reverse the trend in ever-higher leakage
currents, especially so for gate oxide leakage,
where we achieved more than an order of magnitude reduction
in gate oxide leakage with high-k.
But with the focus on strain silicon and high-k metal gate
and now tri-gate or FinFET transistors,
we are able to continually reduce these leakage components
on newer generations of logic technology
and still continue to improve gate delay and switching
energy, but perhaps not at quite as fast a pace
as we had in the earlier classical scaling era.
Another change at Intel is we are
developing a broader range of transistor types
on each generation of logic technology.
In the past, maybe 15, 20 years ago,
when our primary concern was simply
to make ever-faster microprocessors running
at ever-higher operating frequencies,
we typically had only one type of transistor,
well, two, of course, an NMOS and a PMOS.
But we simply focused on developing the fastest
transistors we could make at any generation.
But now, as we move forward to these newer
generations of technology, we are
developing on each generation a broader range of transistors
that not only improve speed, moving from left to right
on the horizontal scale, but also providing
lower transistor leakage, moving down on the vertical scale.
So we actually have a continuum of capabilities,
of transistor capabilities, in terms
of performance and leakage.
And we're really moving on two vectors, again,
the performance and leakage vectors,
and not just the performance vector as we had in the past.
And why do we do that?
Well, because we're now supporting
a wider range of products, from high-performance servers
and desktop chips all the way down
to laptops, ultrabooks, tablets, and low-power pocket devices.
So maybe another way of saying this
is that Intel is no longer a one-size-fits-all company,
no longer developing just one process technology
and one transistor type, providing
a range of transistor types to support
a much wider range of products.
So now let me talk in more detail about our latest
22-nanometer technology.
Again, this is the generation where
we introduced what we call Trigate, what other companies
call a FinFET transistor.
And I show two SEM images here, kind of top-down SEM images,
from our previous 32-nanometer planar technology on the left
to the latest 22-nanometer Trigate technology
on the right.
So looking at the left, these brighter diagonal features,
those are the gate electrodes.
Underneath, you can see the planar source drain structure.
But obviously, the way things have changed on Trigate
transistors, we still have the gate electrode
moving in this direction.
But you can see multiple silicon fins
that the gates wrap around to form the active channel.
Now, you can gang these fins together
to form a larger transistor with more drive current,
as this example is.
So this example happens to have six fins in parallel,
smaller device down here with just a pair of fins
in parallel.
So you can design whatever type of transistor,
layout whatever type of transistor
you want in the circuit.
So here I show the IV curves for both the NMOS and PMOS devices
on this 22-nanometer Trigate technology.
And I'll point out the key benefit
that's derived from the Trigate architecture.
The key benefit shown here, both NMOS and PMOS
have sub-threshold slopes of around 70 millivolts
per decade.
Prior to this, the very best planar MOSFET transistors,
both at Intel and other companies in the industry,
typically had sub-threshold slopes
of around 100 millivolts per decade.
The other key benefit is the DIBL value, the D-I-B-L,
which stands for drain-induced barrier lowering.
It's another measure of the short channel control
that you have on these devices.
On Trigate, because it's a fully depleted device,
because the gate electrode has a much better control
over the channel region, you have very good short channel
characteristics.
DIBL values of around 50 millivolts per volt.
And again, to compare it to the previous best planar MOSFETs,
they typically have DIBL values of over 100 millivolts
per volt.
So again, Trigate is providing a steeper sub-threshold slope
and better short channel control.
So how do these transistor improvements
translate to circuit performance or power?
Well, the steeper sub-threshold slope
benefits circuit design in three ways.
The first obvious way is you have lower leakage current
with a steeper sub-threshold slope.
But there are two other ways we can
take advantage of that steeper sub-threshold slope.
If you have a steeper sub-threshold slope,
you can afford to tune your transistors to have
a lower threshold voltage.
If you have a lower threshold voltage,
you can either achieve a better performance
at low operating voltage.
Or if you want to use just the standard performance
so you can have a lower active power,
you can operate the chips at a lower voltage
than you would normally.
So this is a plot here of transistor gate delay
versus operating voltage.
And it's all normalized to this point
on our 32 nanometer technology.
And this just shows the normal expected increase or degradation
in gate delay as you reduce operating voltage.
But with our 22 nanometer transistors,
we achieve either of these two benefits
at an operating voltage of 0.7 volts,
or about 37% faster than our previous technology.
Or if you have circuits that aren't
concerned about achieving higher performance,
you can simply operate them at about 2 tenths of a volt
lower voltage than you would normally
and achieve the same performance.
And that 200 millivolt reduction in operating voltage,
when combined with the capacitance reduction
from having a smaller device on 22 nanometers,
those two changes combined results in about a 50%
reduction in active power, which is pretty key.
Because today, whether you're talking
about a high performance server chip or a low power
chip in a cell phone, reducing active power
is very key, because we're all striving to achieve
improved power efficiency.
Some other benefits of the TriGate structure.
It's a fully depleted device.
You need a much lower dopant concentration
in the channel to achieve the right threshold
voltage and the right short channel effects.
Lower dopant concentration means you
have less variability, which is important to circuit designers.
They want to know that they're designing circuits
with the threshold voltage all controlled
very precisely across the chip and from wafer to wafer.
So that's another benefit of TriGate.
Here are some actual cross-sectional TEM images
from our 22 nanometer technology.
The image on the left is a cross-section through the fan.
So you can see the silicon fan on that generation
is about 34 nanometers tall.
And it's about 8 nanometers wide at the middle of that fan.
And the image on the right is a cross-section
in the other orientation.
So it's cutting along the length of the fan.
And here you see the gate electrode in the middle.
This is, again, the metal-filled gate electrode,
which is wrapping around the fan, which
is in the background.
And the two tungsten-filled contacts on either side
making contact to the sources and drains.
And on this technology, our minimum gate pitch
is 90 nanometers.
I mentioned before that in the past,
you may have thought of Intel as a one-size-fits-all company
just developing fast transistors to make
high-performance microprocessors.
But the products we make, the markets we serve,
have gotten much broader, much wider since then.
So now we develop what we call a system-on-chip version
of each logic technology generation, where we offer,
really, a menu of features for our circuit designers
to choose from.
They can choose them based on exactly what type of product
they're trying to make.
So for transistors, we offer four types
of transistors from high-performance
all the way down to ultra-low power.
We offer some special high-voltage I-O transistors,
where they may be needed.
We offer a range of copper interconnect stacks,
anywhere from 8 to 11 total layers
with different pitches, different minimum pitches
at each layer.
We offer some precision passive devices, resistors, capacitors,
and inductors, and a range of embedded memory,
high-density SRAM, high-speed SRAM,
and a one-time programmable electrical fuse device.
So again, we offer this menu of feature options.
And each product design team will
pick what is best for their particular product.
Question, yes?
I have a question on your previous slide.
On the fins, you already have, at the top,
something like 6 or 7 nanometers dimension.
Yes.
So that is on a 12-inch wafer?
Yeah, we make these 8-nanometer diameter fins
across a 300-milliliter wafer, yeah.
So the next generation, the next few generations,
it's going to be changing the height of the fin?
Well, you know I can't answer that.
But yes, there will be changes on the next generation.
And I forgot to include it in my slide set,
but I have an image where we kind of,
a higher magnification image of that fin,
and you can count the atoms across it.
Yes, another question?
What kind of Q do you get on your high-Q inductor?
Oh, we have a whole range of inductors.
Some of them, Q value is up around 20.
So I showed this graph before, but just making a point,
then, on the SoC technology, we offer
these four different transistor types or flavors
from a high performance all the way down
to the ultra-low power.
And these are some of the actual ION versus I-OFF,
or IDSAT versus I-OFF curves from our 22-nanometer SoC
technology.
These are all measured at 0.75 volts NMOS on the left,
PMOS on the right, showing how the I-OFF values span more
than four orders of magnitude.
So the highest performance logic devices
are up here with leakage values, I-OFF values
of around 100 nanoamps per micron.
The lowest leakage devices, the ultra-low power devices
are down here with the leakage values of around 10
picowamps per micron.
As I mentioned, then, we offer a range
of copper interconnect stacks.
The initial stack on our CPU chips is nine layers of metal.
And you can see cross-sections here of the copper wires
and the vias that connect from one layer to the layer below.
And right away, you'll notice the hierarchical arrangement
of interconnects that we use.
The tightest pitches down at the lower part of the stack,
thicker, coarser wires up at the top end of the stack.
So obviously, some layers are optimized for good density.
Other layers are optimized for good low RC performance.
And an interesting note, the transistors down here,
these little tiny things at the bottom.
So actually, today, I think we spend more money
making the interconnects on a wafer
than we spend on the transistors.
And of course, for the SoC products,
we offer different interconnect stacks
with both a different number of metal layers,
but also, in some cases, where more of them
are the dense, tight-pitch layers versus the coarser
pitch layers.
These are some of the added features on the SoC technology.
I mentioned a high-voltage transistor,
just showing how we use a thicker gate oxide on the fin,
so they can support a higher operating voltage where needed.
These are some of the inductors we also provide.
And we have a very high-density metal insulator,
metal capacitor here.
So there's a hafnium-based dielectric sandwich
between two titanium nitride capacitor electrodes.
All right, so now, how do we pattern these things?
Here's, again, the trend I showed earlier,
decreasing minimum feature size by about 0.7x per generation,
superimposed on a graph that shows
the trend in the wavelength of light
that we use to pattern these features.
Back in the 1980s, early 1990s, life was simple.
We were patterning features that
were larger than the wavelength of light used.
We're now patterning features that
are much smaller than the wavelength of light.
And of course, I had a great education myself
here at the University of Illinois,
learned lots of things.
But one thing I was taught that was maybe not quite right,
and I can blame maybe the physics department.
I seem to recall a physics professor,
or maybe it was a TA, telling me,
well, you can't pattern features smaller
than the wavelength of light.
So this later learning has given me
maybe a healthy level of skepticism
about what is possible and what's impossible.
So anyway, nonetheless, we are patterning features
much smaller than the wavelength of light.
Initially, we did that with techniques
such as optical proximity correction, where
you add special features on the mask
to make sure that the pattern on the wafer is what you want,
using phase-shifting masks, using immersion lithography
to provide a higher numerical aperture,
and in many cases, layout restrictions.
What I mean by that is in the past,
circuit designers could lay out the transistors and the wires
any way they wanted, wide wires, narrow wires,
one orientation, rotated orientation, diagonals even.
But as patterning gets to be more difficult,
we've had to impose more restrictions on layouts,
more uniform grids of how transistors are laid out
to make sure that they can be patterned.
And now we're patterning 14 nanometer feature sizes,
and one of the ways we've gotten down to that level
is to use something called a double patterning technique,
where you have to send the wafers through the lithography
tool twice to build up the pattern that you
want that has these 14 nanometer minimum features.
And then looking ahead, our industry
has been working for quite some time
now on the next big wavelength change
to reduce wavelength down to the EUV level, about 12
and 1 1⁄2 nanometer wavelength.
Obviously, that would serve us well
at these 14 or 10 nanometer generation.
But unfortunately, despite many years of effort
by many companies, EUV is still not ready yet.
I think one of the latest challenges
is we simply can't build a light source with enough output
power to have good throughput through the manufacturing
tool.
So yes, we are making test patterns with EUV.
They're nice and small.
They provide the resolution improvements
that we're looking for.
But a key metric in the lithography world
is how many wafers per hour can you process through the tool.
Typical immersion stepper today can probably
do close to 100 wafers per hour.
An EUV stepper can do maybe two wafers an hour.
And when the tool costs $100 million,
you want more than 10 wafers an hour out of that tool.
So we still have a lot of work to do
to get the manufacturability or the cost effectiveness of EUV
to where it needs to be.
So luckily, we still have other approaches.
I'm pretty impressed with the lithographers at Intel
that they keep coming up with tricks and ideas
to extend a 193 immersion.
And maybe one of the obvious ones
is to add more double patterning layers.
But that does seem to be a very viable approach, at least
down to the 10 nanometer generation
and possibly even the 7 nanometer generation.
OK.
So I've talked about some of the innovative things
we've done on transistor evolution
for the past decade or so.
So what's next?
What comes after this?
I did mention that, of course, we've
evolved from a planar transistor structure
to a trigate or a FinFET structure.
So we're now here.
I think an obvious next step or an obvious next option
is to further improve the electrostatics
of the transistor by going not to this three-sided fin
structure, where the gate is wrapping around and forming
conducting channels on three sides of the fin,
but to eventually go to a gate all-around structure
or a nanowire structure, where the gate electrode completely
encompasses the wire and provides really
the very best electrostatics.
So that's one option that we're exploring.
And these are some early device results
published by our research group up in Oregon,
showing the IV curves on the left for both NMOS and PMOS.
First thing you notice that the subthreshold slopes are already
improved from the 70 millivolts per decade
that the trigate device offered down close to 60 millivolts
per decade, at least on the PMOS device in this case.
And the short-channel characteristics,
the DIGL values, are improved from 50 millivolts per volt
down to somewhere between 10 and 20 millivolts per volt.
So these gate all-around or nanowire-type structures
are demonstrating the improved electrostatics
that you would expect them to provide.
But then the next question is, OK,
if you want to make a nanowire, how small can you make
or should you make that nanowire?
So we've done some simulations on different nanowire
diameters, from 10 nanometer diameter down
to sub-5 nanometers.
And what the models and some of our early experimental data
are showing us is that the mobility is dropping off
rapidly as you go below 5 nanometers.
So we need to understand that better
if we go this route to select the right nanowire diameter.
Another area of very active research
is to explore different channel materials.
Of course, I'll remind you, we've already
changed the source drains to silicon-germanium source
drains.
We've already changed the gate dielectric to a high-k.
We've already changed the gate electrode
to special metal gate materials.
So what's left?
Well, let's change the channel.
Why stay with silicon for the channel material?
So these are some of the materials
that are being explored or have some interesting properties,
including germanium, gallium arsenide, indium arsenide,
and indium antimonide.
Of course, the main reason for exploring different channel
materials is to provide improved carrier mobility,
both electron and hole mobility.
This is, I think, a nice plot showing the relative electron
and hole mass, as shown by the size of the circle compared
to silicon, but also the lattice constants and the band gap.
So this kind of illustrates the range of interesting options.
But obviously, a larger lattice constant
is going to give us some manufacturing challenges
and also, obviously, a small band gap
is going to give us some problems with leakage currents.
Question?
Yes?
So does this mean that we're never
going beyond 0.6 nanometer technology?
And when do we drop that designation?
Does this mean we'll never go beyond a 0.6 nanometer?
Oh, just because of the lattice constant?
I haven't thought that far ahead, sorry.
Do you anticipate there's a theoretical?
Well, obviously, whenever I'm asked,
when is the end of Moore's Law, and I'm
sure nobody will ask that question today.
But clearly, once you're down to atomic dimensions,
the way we've done Moore's Law is going to come to an end.
But something else may replace it.
So the nomenclature will probably have to change as
well.
What's that?
The language that we use, the technology known language.
Yes, yes.
Even the naming system has evolved
and is not quite as meaningful as it was 20 years ago.
20 years ago, when you talked about a quarter micron
or half micron technology, it meant something very explicit.
It meant the gate length in the MOSFET.
But today, most companies are deviating
from that very simplistic naming system.
They still use names like 22 nanometers and 14 nanometers.
But even I will admit, on our 22 nanometer technology,
I cannot point to the one feature that is 22 nanometers.
Because some are bigger than 22, and some are smaller than 22.
But what is true, at least at Intel,
is when we talk about a new generation of technology,
such as moving from 32 to 22, it
does indeed reflect an average 0.7x reduction in feature size.
And it does reflect an average of about a doubling
of transistor density.
So again, different channel materials
is an active area of research to enhance mobility.
Graph on the left shows some of the benefits.
It shows at least how, not only how
germanium provides better home mobility than silicon.
That's nothing new.
But also showing how germanium responds much better
to channel strain or channel stress
to further improve home mobility.
So that's pretty attractive.
But the ongoing challenge with some
of these different channel materials
is the need or the difficulty in forming a high quality gate
dielectrica on that channel.
So that's, again, another area of active research.
Of course, we've been looking at carbon materials,
whether it's carbon nanotubes or graphene,
which does hold promise for very high mobility, very high drive
currents.
But I think as many of you probably know,
we have this challenge, both on graphene and on nanotubes,
of yeah, you can get great mobility.
But when you have great mobility,
you usually have a pretty small band gap.
And that can give you a leakage problem.
So still, at least that challenge
with carbon-based materials, finding a solution
that gives you both high mobility as well as
good off-state leakage.
Here's kind of an illustration of that challenge
for carbon nanotubes.
We're plotting here nanotube diameter
versus the on current.
But I'm also showing that this colored contour
a plot of a leakage current.
So unfortunately, the highest on currents
also provide the highest off currents.
So there's not presently an acceptable solution
for nanotubes or even graphene.
Another class of device we're looking at
is called tunnel FETs.
Of course, MOSFETs work by electrically modulating
the barrier height between the source and the drain.
Tunnel FETs operate by electrically modulating
the barrier width between source and drain.
So in the on state, carriers are tunneling through the barrier
from source to drain.
So why would we be interested in this device?
Well, because it offers the possibility
of a device with sub-threshold slopes
that are much steeper than 60 millivolts per decade.
Now, I already described earlier how
the Trigate device with a sub-threshold slope
of around 70 millivolts per decade
provides some pretty interesting benefits
in terms of low voltage performance or lower
active power.
Well, if we can come up with a TFET device that
has even lower, steeper sub-threshold slopes,
so we can further reduce operating voltage,
further reduce active power compared to today.
So again, the benefit is that compared
to sub-threshold slope on normal CMOS,
you can get much steeper with a TFET.
But one of the challenges is that to really make
a reasonably good TFET, you have to have a heterojunction
device.
So again, we have to get into some 3-5 materials
to make a good TFET device.
So I've already briefly described
TunnelFET, which is not a MOSFET device any longer.
So that's one device I think we could
describe as beyond a CMOS.
Another example might be graphene-based pn-junction.
And beyond that, there's a much wider range
of spin-based devices being explored
at various universities that may also
provide interesting beyond-CMOS options.
So I don't have time today to go through all these devices.
But again, they're spin-based.
Some of the names include spin-torque-majority-gate,
spin-torque-domain-wall, spin-FETs, spin-torque-triad,
spin-torque-oscillators, all-spin-logic,
spin-wave-device, and nanomagnetic-logic.
So all of these devices look intriguing in that they
can be much smaller and denser than CMOS, very likely
lower power than CMOS.
But as an industry, as we explore these different
options, we have to have some way to benchmark them,
to understand what exactly is the performance and power
capability of each one of these.
And can we, using such a metric, such a benchmark metric,
can we down-select, can we eliminate
some of these devices as being uninteresting
and focus more of our research dollars
on the more promising or more interesting devices?
So in an interesting paper by two Intel authors,
they have put together a circuit benchmarking technique
to explore the merits or demerits of some
of these beyond-CMOS options.
So here, they're plotting the energy of a, in this case,
we're using a very simple circuit, a two-input NAND
device.
And the vertical axis, we're plotting the switching energy
in femtojoules versus a gate delay on the horizontal scale.
And I start by showing two points, a high-performance CMOS
point and a low-power CMOS point.
So this is what any new beyond-CMOS technology would
have to be compared against.
So starting off with some of the electronic options,
they're not CMOS, but they're some of the TFET, spinFET,
and graphene p-injunction devices.
As you can see, initially, they have pretty good delay.
They are faster than low-power CMOS.
And the spinFET and the graphene devices
initially appear to use more energy
than high-performance CMOS, so thus, they're
not very interesting.
But the TFET devices, they seem to be
in a pretty interesting place on this graph.
Below, certainly, the CMOS line,
not quite as fast as high-performance CMOS,
but still in a very interesting regime.
So initially, we think TFET devices have some promise.
But now, as I continue and add some of the spin-based devices,
the upper-right cloud there are spinTorque.
The lower-right cloud there is a mixture
of spin and electronic devices.
It's called magnetoelectric devices.
As you can see, all of these spin-based devices
are initially slower than the low-power CMOS,
so not very promising in terms of performance.
And only a few of them are lower energy than low-power CMOS,
so initially, not as promising as we would like.
But they have one other advantage.
They're non-volatile, and that can
be a pretty interesting benefit.
So if microprocessors can be designed
to take advantage of that non-volatile characteristic,
there may be some significant performance or power
advantages.
So we need to give that some more thought.
Now, so these previous studies were all
done on a very simple two-input NAND circuit,
but the authors of this paper extended it
to a more comprehensive 32-bit adder circuit.
The general conclusions are not much changed,
that the spin-based devices are still generally slower
than low-power CMOS and not much lower energy,
whereas the TFAD devices are not quite as fast
as high-performance CMOS, but they
are faster than low-power CMOS, and they
have about an order of magnitude lower switching energy.
So again, this initial benchmarking effort,
this initial screening effort, suggests
TFADs may have more promise, but it should not
stop us from further benchmarking and further
improving our understanding of some of these other
beyond-CMOS options.
OK, looks like I'm just about on time here.
It comes to my summary slide.
Transistor scaling has evolved in revolutionary ways
over the past decade and will continue
to do so going forward.
With the increased interest in mobile computing devices
and power-efficient computing, the goal of scaling
now favors lower power over higher performance.
And as we explore a variety of beyond-CMOS devices,
the selection criteria should focus on higher density
and lower power.
So that concludes my talk.
Thank you for attention.
We have time for a few questions,
but first I'd like to present Mark with this plaque
that we've made for him in appreciation of his willingness
to come and participate as one of our keynote speakers.
Please, your applause again.
Time for a few questions, a couple of questions.
I'm glad you asked questions as we were going along,
but maybe a couple more.
Mark, you didn't speak about nanomechanical logic.
What are your thoughts on actually that is?
Yeah, good question.
There is some interest in that.
Some individuals in our research group
have been looking at that, not making devices,
but doing some of the modeling and simulations.
It's problematic, I think, that the problem
is some of these devices, these section forces
are pretty difficult at these very tiny dimensions.
But I won't totally preclude it.
So I could have added that, should
have added that as another beyond CMOS option.
Way in the back, Rashid.
In the nanowire structures that you were showing some results,
is there strain induced because of the observation?
Does that help?
Strain should help.
Those devices, I don't think were strained.
We'll have to piece things together eventually,
but those were probably unstrained devices.
Coming from the circuit and systems area,
when I look at these devices, I noticed
that the device community is presenting
what is called an ideal switch.
That you're trying to shoot for an ideal switch,
and that is being driven by our need
to build standard logic circuits,
one-nominate architectures.
Suppose we take that out of the picture
and say, give us a functional block that
may behave statistically.
So the process variation envelopes
are expanded significantly.
If such a relaxed specs were to be
provided to a device engineer, what benefits can we get?
I mean, can we use that relaxed specs
to reduce energy significantly?
Or are there other ways of exploiting it?
Well, I think clearly the best way we can exploit that
would be to operate at lower voltages.
What restricts us mostly on operating at lower voltages,
not because things get slower.
Yes, they do, but in some cases that's OK.
But because random dopant fluctuation
will make some transistors non-functional or more
susceptible to errors.
So that's probably the first thing
we would do, is lower operating voltage.
And that would automatically increase the variations?
Yes.
One last question.
You mentioned that Intel spent more time designing
the modern transistors.
So my question is, is Intel trying
to solve the interconnection structure using metals?
Or is Intel sitting on the front of the line integrating?
Yes, so the question is, what means
are we exploring to solve the interconnect problem?
We do a lot of evolutionary things,
learning how to make interconnects pitch tighter,
how to minimize resistance.
That's an odd phrase, minimize resistance increase.
But just to try to do everything we can
to avoid significant increases in resistance, lower
K dielectric.
So those are all kind of evolutionary things.
A more revolutionary idea, such as optical interconnects,
is something that we are exploring
as a chip-to-chip solution.
But I don't foresee it being an on-chip solution
for a very long time.
Out of all the Beyond CMOS devices,
which one would you say is the most promising,
from Intel's perspective?
Well, I already indicated.
I don't know where we're left.
So far, by this, this is what's called a very preliminary
benchmarking study on a very early understanding
of how these devices work.
So a lot of caveats on my response.
But so far, the only one that shows promise
is the TPET.
But let's go stare harder at some of these other options.
But these are simulation results?
These are mostly simulation results
based on what experimental data we know about these devices.
But as a community, as an industry,
we need to really stare hard at this type of benchmark
and apply it to the different ideas
so we can narrow down this range of options
and focus on the ones that have more promise.
Thank our speaker again, please.
Thank you.

