The following content has been provided by the University of Erlangen-Nürnberg.

Good afternoon everybody. Welcome to our today's Invasive Computing Seminar.

It's my great pleasure to host today a guest from very, very far away.

It's Professor Moteza Biklari Apari. I hope I pronounced it almost well.

And I know him also personally from my visit last summer at the University of Auckland in New Zealand.

And he is here on a trip to Europe and I'm very happy also he had time to pass by

and tell us about his research work on energy efficient real-time computing.

Thank you very much. Thank you. Okay, thank you Jürgen for the introduction and also thank you for the invitation.

It's my pleasure to be here to visit Professor Jürgen Thay and his colleagues.

What I'm presenting here is one of our ongoing research in the Embedded Systems Research Group in our department.

And this is a joint research by Professor Zoran Salcik.

So first I talk about the motivation of this research.

Just looking at the CPU and performance improvement in the recent years,

we see first the single core error, we could achieve single thread performance improvement.

This red circle indicates the current state.

So as we see the further performance is diminishing from single thread performance point of view.

Application for this case usually use general purpose language like C++ and also Java in some cases.

And some special accelerator where program using assembly.

So this technology, this performance improvement has been enabled by Moore's law.

And as we could have smaller transistor voltage scaling also would be possible.

But this progress has been limited by power consumption as well as the complexity of the single chip processor.

Then moving to multi core error, again the technology was enabled by Moore's law as well as using symmetric multiprocessor architecture.

So we saw that we could achieve again performance improvement up until some point, for example, which is the current time.

But here some specific language to exploit the parallelism in the application have been used.

But again, the constraint was power consumption limited parallelism in the software and also more importantly, scalability.

Then some heterogeneous systems have been developed, especially by exploiting the data parallelism available in some specific application and relying on power efficient GPU.

And this we are in the almost early stage of this progress.

So based on the ITRS prediction, there are still some potential for further improvement in this case.

But here this relies on some specific language like OpenCL and CUDA.

And again, we are constrained by power as well as using a specific purpose language.

So this is one side of the story. We know that Moore's law might be still valid until maybe 2020 or some people say a little further.

But the problem is with power consumption on the chips.

Denar's scaling, which was valid until 90 nanometer technology, would predict that as we have a smaller transistor scaling toward a smaller transistor,

the power consumption of the transistor also would be scaled based on that. But that was valid until 90 nanometer technology.

And since then, this scaling doesn't work, which relies on something which is so-called dark silicon,

which we have many transistors on the single chip but cannot be used at the same time.

So this is one obstacle of relying more on multicore scaling as was the case here.

The other side is from the application point of view, especially in our case that we are interested in embedded system application.

We see that there has been significant increase in embedded system complexity.

Today's typical embedded system application have a number of concurrent behavior,

which process data, communicate with each other, as well as interact in some way with the environment.

In addition to this, we need also to consider support for distributed embedded system design,

as well as considering the importance of power energy optimization,

and more importantly, for real-time embedded system to guarantee the worst-case execution time and worst-case reaction time.

So these are the challenges that should be considered by using the available technology in a proper way.

So basically, our challenges would be how to model and implement such complex embedded system,

and also how to achieve the required low power consumption and time predictability in these cases.

So the approach that we are using to tackle this problem is relying on GALS system,

globally asynchronous, locally synchronous system.

In this case, individual component of the system will operate at different clock domain,