The following content has been provided by the University of Erlang and Nuremberg.

No, no, no. Thank you so much. Thank you. Thank you for inviting me. And I'm really glad to

be here and to share with you some of the work that my research group is doing at National

University of Singapore. So in case you have never been to Asia, Singapore is called the

gateway to Asia. So it's the easiest introduction you can get to Asia. So I invite you to come

to Singapore if you do get a chance to be in the region.

By the way, feel free to ask me any questions as we go along in the talk. I would prefer

that. So just feel free to ask me questions. So today I'm going to talk about heterogeneous

multicores in the dark silicon era. And I guess I don't need to give that much introduction

for this group because invasive computing is all about heterogeneous multicores. But

I guess the kind of heterogeneous multicores I'm going to talk about today is somewhat

different from the kind of heterogeneity that you are exploring in this chair. So we are

of course firmly in the multicore era and we have already hit the power wall and we

are going to get more and more cores on chip. But today what is happening is that most of

these cores are really kind of homogeneous in nature so you have the more of the same

thing on the chip. But with that what is happening is that one of the problems that's creeping

up is this problem of dark silicon. And what is happening is that of course due to Moore's

law you are getting more and more transistors on chip, right? So Moore's law is not failing

yet and so you are going to get more and more transistors. But what is happening is that

there is an associated law which is kind of denard scaling that's failing. And what happens

because of that is that even though your transistor size is getting smaller you are not reducing

the power per transistor. So when that happens here is what is the problem that you are going

to see. So let's suppose that this is a chip that you had. It has a single core. In 2008

it was at 45 nanometer and it was consuming one unit of power. It was running at unit

frequency and it has unit area. Now when you try to scale it down, so if you go down to

2014 and we are looking at 22 nanometer design, so this particular core that was occupying

this entire area now is occupying only one fourth of the area. So that's really nice

because now we can put four cores on the chip and that's what multicore is all about. Unfortunately

what is happening is that your power requirement for this core still remains the same. That

means now if you put four of these cores together then they are going to consume four units

of power instead of this particular chip which was consuming just one unit of power. Unfortunately

your thermal cooling capacity and things like that are not improving. So your thermal design

power budget still remains the same. So if you had one unit of thermal design power at

this point, you still have one unit of thermal design power at this point. So what happens

because of that is that, so let's suppose I had one watt here so I still have one watt.

Now if I switch on all the four cores then I am going to consume four watt and so clearly

I cannot actually do that so I have to switch off some of the cores. So in this case what

happens is that you are going to switch off all the three cores and you are left with

just 25% of the silicon area that you can exploit at any point in time. And this is

going to get worse and worse if you are looking at 2020. This is a prediction from ARM based

on the roadmap for frequency and power and things like that that you are going to get

only 10% usable chip area. So that means that even if you put 100 cores on chip you are

only going to be able to use 10 of these cores at any point in time. So what can you do about

this? Well you can be a very pessimistic person about it and you can say that well I can have

this entire very big chip and I can put 1000 cores on that chip but since I can only use

hundreds of them why bother? Let me just go with a smaller chip and let's just put 100

cores on it. You can do that but that's not nice both from the industry perspective because

then you don't have anything to do and of course from academic perspective because we

have no research to do. So we would like to be more optimistic about it and we would like

to come up with some solution to this challenge and get kind of something out of this very