Computing in the Nanoscale Era
Sep 5, 2023
"Computing in the Nanoscale Era"
Naresh Shanbhag, Professor of Electrical and Computer Engineering, University of Illinois
CNST Nanotechnology Workshop 2013
So today, I would like to talk about a new research
center called SONIC.
SONIC was established earlier this year.
And its goal, its mission, is to design computing systems
on nanoscale fabrics.
So the term nanoscale appears in my talk.
And that's one of the reasons I'm here.
So what is SONIC?
SONIC stands for Systems on Nanoscale Information Fabrics.
Its goal is to enable equivalent scaling.
That is Moore's law.
If I use a populist term to describe our mission,
then it's really to enable Moore's law
in the ultra-scale CMOS and beyond CMOS technologies.
And the center has faculty from a variety of disciplines.
It's not a pure devices center.
It's a center that includes neuroscientists, system
designers, information coding theorists, circuit designers,
and device researchers.
We are spread across eight institutions,
with University of Illinois being the lead institution.
We, in fact, had bid for the center previously.
Many years back, a couple of times,
this is the first time we succeeded in getting it here
to Illinois.
It's a $30 million center with a five-year time duration
starting in January of this year.
It's funded by the Semiconductor Research Corporation, SRC,
and DARPA.
Intel is a major, major player in SRC.
And we thank Mark Bohr here for his support of the program
and for Sonic.
The time horizon for the Starnet program is 15-plus years.
And this is a very important number.
Our sponsors take it very seriously.
And our PIs take this number very seriously.
And I would say that this kind of time horizon
is one of the most farthest-looking number
or time horizon compared to the other research programs.
So what is Sonic about?
What kind of problem are we trying to solve?
I think a good way to understand our research focus
is to look at how computers are being used today.
So in 2011, there was a New York Times article
in which we showed Watson beating Ken Jennings,
the human competitor.
And this was the big news in New York Times.
But there was a side note which was not mentioned,
is that Watson is a supercomputer.
It is an 85,000-watt computer.
For it to act like an intelligent entity,
required it to consume a large amount of power.
So today's computers, we view them as number crunchers.
For them to exhibit any amount of intelligence.
And what do I mean by intelligence?
Intelligence is the ability to look at the world around you,
to make decisions based on data
that is incomplete and noisy.
That is one way that intelligence manifests itself
in this physical world.
Well, computers can do it, but they need lots of watts
to exhibit that kind of behavior.
So in Sonic, we like to think about this kind of intelligence
and we want to bring it onto our devices, handheld devices.
Are you using this?
No.
Sorry about that.
So, all right.
So today I've seen many of us carry iPhones
and we use a tool called Siri.
Siri is a voice recognition software.
Well, for Siri to work, you need to connect to the cloud.
So we bring in intelligence.
We make our handheld devices look intelligent
by putting the data processing
or the number crunching facility in data centers in the cloud.
So the intelligence resides in the cloud
and with that we are able to make our sensors,
our handheld units, look smart.
But as you look around us,
the number of these handheld mobile units, sensory swarm,
the number of those sensors are supposed to increase
tremendously over the next few years.
So the connectivity problem that we are beginning to observe
is only going to get worse.
So our view is that we need to process data locally
as much as possible and be able to demonstrate,
build what we call as intelligent computers.
Now, why did we come to this situation?
Why is it very hard for modern day computers
to exhibit intelligent behavior?
It is because we are using archaic models of computation.
Models of computation.
So the Von Neumann model was set up in 1945
and today all of our computers,
no matter where you look,
is some one way or the other,
there's one variant of the Von Neumann architecture.
The Von Neumann architecture is not really tailored
for inference-based applications.
So the applications we just talked about earlier
are all inference-based applications
in which you have massive data volumes
and then you want to process it
and come out with, at the end,
come out with a couple of decisions for the user.
That kind of processing is intrinsically imprecise.
Its output tends to be in terms of likelihoods,
probabilities, that doesn't quite fit
into this deterministic computing paradigm
that was proposed by Von Neumann in 1945.
But this is what we use today.
There's no other, we don't consider any other alternative.
Complicating our lives further
is the fact that Moore's Law has slowed down.
You know, the number we saw in Mark's plots,
we saw plots going all the way down to seven nanometer
and then dot, dot, dot.
So scaling has become difficult
and one of the primary reasons for that difficulty
is because device researchers are using metrics
or device design which were driven
by the Von Neumann architecture.
You have a deterministic architecture,
you have a deterministic logic network,
and therefore you need a deterministic switch.
And a deterministic switch has to operate
exactly the same way every time you switch it on
and switch it off.
Your threshold voltage should have a very narrow tolerance.
So where are all these specifications coming from?
They're all driven top down by the need
for building a Von Neumann architecture.
Sonic takes the position that let's take Von Neumann out
and look for other ideas, look into other disciplines,
bring in perhaps statistical information processing
concepts and build systems out of those.
In that way, we can enable really novel nanotechnologies,
very novel devices that otherwise you would consider
them to be of no use.
So why not look, so in Sonic we ask this question,
why not look into the area of communications?
So if you look at the 20th century,
the three major inventions or discoveries
of the information age, one was the transistor,
the other one was Von Neumann's architecture,
and the third one, I would argue,
is the paper that Shannon wrote in 1948.
In fact, this paper was written in the same year
that the transistor was invented.
In fact, this paper was written in the same place
that the transistor was invented.
So there's a lot of spatio-temporal coincidence
that resides in that year, 1948.
Prior to 48, communication systems were designed
by just transmitting a large amount of signal power
and overwhelming noise.
So you have a high SNR channel.
If you want to talk to somebody,
you shout across the room to get your message across.
But after Shannon's work, he showed,
Shannon showed that you can encode the data in smart ways
so that you don't have to really send
a large transmit power, but instead use the structure
of your signal to recover, to infer what was being said
or what was being transmitted.
So that's the birth of modern communication systems.
All our cell phones, the optical communications,
all the way to wireless, wireline wireless,
all of them use this basic concept that Shannon proposed.
But Shannon's ideas have never had any influence
in the design of computing systems thus far.
So what Sonic is trying to say is let's mine
this 60 years worth of technology that resides
in the world of communications and bring it to bear
on the design of computing systems.
So one of the things Shannon did was to distinguish
between the notion of data and information.
I'll just take a couple of minutes
to illustrate that difference.
So data is bits and bytes.
And a computer today would, a computing engine today
would look at those bits and bytes,
look at its current state, and go to another state
and output something, right?
That's a deterministic von Neumann form of computing.
Information is totally a function of the probability
with which those data bits or data bytes occur in practice.
And the nature around you is generating those probabilities.
You don't have to generate it.
It's available.
We just ignore it.
When we build computing systems,
we just ignore those probabilities.
In fact, Shannon quantified the notion of information
through this term entropy.
So if you look at that expression,
the data values don't appear,
but the probability values appear.
So everything depends on how the probability distribution,
the density functions look like
when you describe the data that needs to be processed.
So in summary then, today's computers process
what I call as the form.
It works on the form.
But we, as users, are interested in information.
Statistical inference applications
are trying to dive into this massive big data,
massive volume of data,
trying to extract information from it.
So we need to build processors
that are focused on information rather than data.
And that's a key distinction between the sonic computer
and the computer of today.
Most large centers have many research themes, many tasks,
and sonic also has a similar feature.
But we are lucky that all of our tasks and research
is kind of driven, guided by a philosophy,
a research philosophy that says
that if information can be described by a density function,
and if you can characterize the nanoscale non-idealities
with another density function,
perhaps we can put them together
and create a common mathematical framework
that would allow a device researcher
to talk to a system designer
and perhaps discover new modes,
new technologies for processing information.
So we believe that moving from the world of determinism
and embracing statistics is the right way to go.
And we call this, we call our computers the sonic computer
as a statistical information processor.
Instead of a data processor,
we want to be designing information processors.
All right?
So putting it all together,
what does it mean for a device researcher?
What it means is this.
If you want to build information processors,
if you're reaching out into the world of communications,
into the world of neuroscience,
extracting information processing principles,
and then building computing engines,
the kind of devices we need
are not the kind of devices that are built today,
in the sense that we don't need an ideal switch.
We may not need a very high eye-on to eye-off,
something that looks like a switch.
In fact, what we would like is to move from a switch
to a slightly larger computational unit,
perhaps a majority gate, perhaps a full adder.
We want, we use the term nanoprimitives.
So not a nanodevice, but a nanoprimitive.
Nanoprimitive definitions come about from the applications.
Applications define what nanoprimitives make sense.
And then we don't need these devices to be ideal,
that that is deterministic.
They can exhibit statistical attributes,
and we know how to handle statistics
at the architectural and system levels, all right?
So this is what I think we need.
So the device researchers in Sonic
are those who are actually willing to take the next step
and build small modules,
working with our circuit designers and system designers,
actually building some small subsystems
to demonstrate their ideas.
And in fact, this idea has legs.
We have shown in the last six years
that using statistical estimation and detection techniques,
we can, in fact, compensate for tremendous amounts
of errors that might occur
in even current day process technologies.
So this was a work that was done
as part of a research team I was leading
called Alternative Models of Computation.
GSRC was one of the previous generation,
I won't use the term StarNet,
but XCRP centers based in Berkeley
and then later in Princeton.
And I was leading a research team
where we were asked to kind of
investigate these kind of ideas.
And it turned out to be a very good, successful theme.
And this illustrates that we can take some of these ideas
from algorithms, architectures,
down to integrated circuit implementation.
We can demonstrate with measured results.
So here we reduce the supply voltage
until we start getting errors.
Something similar to what Mark said earlier
when I asked the question,
if I want to get energy efficiency,
but I can tolerate variations,
what is the right way to do it?
Well, reduce the supply voltage
and you'll start seeing errors and variations.
This is exactly what we did with our CMOS chips.
The chip on the right hand side
is in fact designed by Rami Abdullah,
one of my students who is currently at Intel.
And this illustrates, this gave us a lot of confidence
that the ideas we are talking about,
the theoretical statistical inference based ideas,
in fact, have legs.
And in fact, they have a lot of practical implications.
So the Sonic Research Agenda is
the structure of our research program
is to have four themes.
The first three themes are systems driven themes.
So theme one on stochastic information processing,
theme two is called information based
mixed signal interfaces.
So the first one is about information processing,
the idea that if you have non-ideal
statistical building blocks,
how do you put them together to process information?
The second theme two is looking at interfaces.
If the components in my interface
are also statistical in nature,
how do I transfer information from it?
So these two, theme one and theme two,
are communication inspired, Shannon inspired.
Theme three is looking into the world of neuroscience,
asking the same question.
How does the brain, how does Ken Jennings
process information so efficiently?
And we want to distinguish between
kind of the neuromorphic kind of research
that has gone on for so many years.
This is not neuromorphic computing.
It is neuro-principled.
We want to understand the principles of computing
that reside in the brain, model those mathematically,
and then bring it onto an implementation platform.
Theme four is a very important theme.
It is a experimental theme.
We have device researchers in theme four
and circuit designers in theme four.
And you'll notice there is a sub-theme
called nano-functions, which is more of a circuit-centric
theme, whereas nano-primitives is much more
of a device-oriented theme.
So all of these three, the top three themes
are anchored firmly on a foundation
of an experimental theme, which is theme four.
That is, you would like very far-reaching ideas
to find their way all the way down into silicon
or beyond CMOS implementations.
All right, this is our team.
It's the best of the best right here sitting in Sonic.
I can tell you that each of these individuals
can succeed on their own.
They don't need Sonic.
However, what Sonic brings to them and to our sponsors
is the fact that these smart people
are going to work together.
If you have a neuroscientist working with a device
researcher, imagine what sort of ideas they can come out with.
And imagine what kind of students
they will educate and send out to industry.
Illinois is the lead institution in Sonic.
I can also tell you that during the proposal preparation
phase, the process itself was a grassroots process,
in that all the PIs contributed their inputs.
We had an executive committee.
We had team leaders all set from day one.
And the process was, I would say, seamless.
It was not as painful as I thought it might be.
So just the idea that the process
of going through preparing the proposal
gave me a lot of confidence that we all will work together
and produce great results.
The management team is composed of,
so we have taken each of the team leaders
and created an executive committee.
Andy Singer from Illinois is leading team one.
Yandabai from Berkeley leads team three
on cognitive information processing.
Boris Murman, who comes with a very strong background
in mixed signal analog design, leading team two.
Team four is very important to us because the nano team.
And we have allocated circuit designer David Blough
from University of Michigan and Philip Wong from Stanford
as the two core team leads.
We have other EC members who have been kind of selected.
They're PIs, of course, but they've been selected
to be on our executive committee
because they bring in lots of directorship type experience
from previous FCRP centers.
So Rob Rutenbar, for example,
is former director of C2S2.
Sharad and Jan are both former directors of Music and GSRC,
the previous generation FCRP center.
So I got a lot of people who can give me advice
and I can brainstorm with them.
So I'm delighted to have them on the team and on our EC.
So the takeaways of my talk are,
number one, we need to embed information-based metrics.
Not an ideal switch, eye-on to eye-off, these things.
Maybe they're secondary metrics.
The primary metrics are how much information can flow
from the input to the output of a device.
If you have a transistor, a noisy transistor,
or a noisy nanoprimitive, how much information does flow
from the input to the output?
Information is statistical, noise is statistical.
Maybe we can optimize the transistor
in a very different way.
We want to move from von Neumann to Shannon.
I don't take von Neumann out, bring in Shannon,
takes out this deterministic brick
that sits between inference applications at the top
and nanotechnologies at the bottom.
In fact, I would argue that if von Neumann were here today,
he would applaud SONIC.
Because in fact, he did applaud SONIC in 1956
in his paper in which he was talking about
building reliable systems out of unreliable components.
But right in the front page,
the first page of his paper says,
I wish my work were just like Claude Shannon's
because that's the right way to treat errors.
The manner in which errors should be treated
is as information.
I believe that SONIC's research program
and our vision is doing exactly that.
I want to stop here by pointing out
that SONIC brings together von Neumann as a background,
Shannon, and Bardeen, the three major components,
the three major innovations
that occurred in the 20th century.
I believe that we have a chance of making history.
This is the first time that all the pieces
have lined up properly,
we have the right set of researchers,
the timing is right, industry needs it,
society needs it, and we have funding today.
So with that, I'll stop.
Thank you.
Any questions?
Who could argue with a dead von Neumann?
Yeah.
Have you considered, like, including quantum computing?
Quantum computing, yeah.
So we thought about it,
but we consider that we don't have to go to that level
right now.
Shannon is a sort of material that we can mine
from the world of communications.
And, you know, I think that's the right way to go.
Yes.
I'm assuming completely, sorry, I'm assuming,
so the proposal of designing a slightly imperfect processor
in order to treat information at a certain level.
So what is the major immediate benefit of such a processor?
Like, is it in terms of power consumption?
Yes.
Or in terms of more useful information?
That's right.
So the major benefit is that at the end,
we want robustness and energy efficiency, right?
And the goal of a statistical information processor
is to hit those design points
where energy efficiency and robustness are both optimized,
and which would not be optimized to that level
if we stick to the von Neumann paradigm.
So think about a communication link.
The communication link today with your turbo code,
the LDPC codes, the energy per bit of information
that's transferred, it has already hit its optimum value.
The minimum value has been hit.
Shannon capacity has been reached, all right?
After 60 years, we have reached it.
We want to do the same thing for the world of computing.
Computing is way behind when you look at,
in terms of sophistication,
when it comes to energy efficiency and robustness,
it's way, way behind communications.
So to bring it on par,
the world of computing needs to come
and be on par with the world of communication.
So yes, the short answer is,
we get energy efficiency and robustness in one shot.
Thank our speaker again.

