Thanks for the introduction. It's a real pleasure to be here. I'll present today some of the

recent work we have done in the area of ultra low power digital design. So the title of

my presentation is Toward Energy Neutral Computational Sensing. And I will explain in the next few

slides what computational sensing means to me. This is joint work from my two groups,

the one that I recently started in ETHZ and the group that I have in Bologna. Okay, so

just a couple of slides from motivations. This is the well-known Gartner hype cycle

with all the technologies coming up to the hype and then going toward more mature market

solutions. So in 2014, the peak of the hype cycle was Internet of Things. So it's clearly

almost a requirement to position your scientific work as relevant for the Internet of Things.

And of course, I do the same. Okay, so. But to be a little bit more quantitative, this

is actually one of the very interesting slides that shows how the number of the size of data

that is produced by this increasing number of Internet connected objects and nodes is

exploding exponentially faster than Moore's law. And people have started to coin new terms

and the new term is Bronto byte, which is 10 to the 27th byte, which would be the size

of the Internet of Things a few years from now, like five years from now. And so one

question you may want to ask ourselves is, yes, we produce all this data, but we want

to do something with it. So a question that we may want to ask ourselves is how much energy

do I need if I want to process just one operation per byte, one Bronto byte? And this is because

if I cannot have enough energy to process it, then what is the point in storing it?

So the answer to this question boils to another question, is how energy efficient I can be

with my computation. And this plot, which is coming from IBM, I think it's a really

nice plot, shows how humanity has been doing a very good job over time in improving the

energy efficiency of digital computation. And it shows on the, it's a bi-logarithmic

scale as you can notice, it shows on the x-axis computing density, so a unit of computations

per volume per liter. And on the y-axis shows computing efficiency, which is operations

per joule. Now you see that this is a straight line in a bi-logarithmic scale, which means

exponential improvements in both axes. And as technology has improved, things have gone

very well. And they are moving quite rapidly toward the green area. What is the green area?

Well, the green area is the biological brains. So if you can estimate roughly what a brain

does in terms of computational power, we are still a few orders of magnitude away from

what in terms of energy efficiency can a brain do. But if, you know, if you allow me just

a little bit of cheating and assuming that an elementary operation of a brain is equivalent

to a 32-bit elementary arithmetic operation, which is not. Brains do things in a different

way, but let's assume that this is it. Then to reach the energy efficiency of a brain,

we will need to have 10 to the 12th operations per joule, which basically means one picogoule

per operation or one giga-ops per milliwatt. So let's assume that we have a computing machine

that reaches the level of efficiency, the entry level of efficiency comparable with

a brain. Then how much energy we will need to process a Bronto byte with this machine,

one gops per milliwatt machine. So one Bronto byte per year on a single operation per byte,

we would just need the 10 megawatts, which is more or less the power budget of a very

large computing facility today. So essentially it means that if we can achieve this type

of energy efficiency for computation, the managing the complexity of the Internet of

Things could be, from the computational viewpoint, could be affordable. Expensive, but affordable.

So the question that I will try to answer in the rest of the presentation is how can

I come up with a machine that is programmable and delivers one gops per milliwatt. Now the

big challenge, as you know, in using silicon for delivering computational capability is

the challenge of energy proportionality. So this graph is again in bi-logarithmic scale

and shows with a blue line the one giga-ops per milliwatt target. And since we want to

have machines that go very fast and also machines that consume very little power for much less

computational capability, this is a straight line over the milliwatt and the gops. And