Hello and welcome everybody to the next lecture of architecture of supercomputers.

Today we will be talking about multicore CPUs and I'm not sure how far we get but if we

get far enough we'll also talk about the roofline model which should remind you of the previous

exercises.

When I was studying there was actually the time of transition between single core systems

and multicore systems.

You could now question why do we even have multicore systems and actually the reason

for this is physics, its manufacturing processes.

You can see here is actually an older slide.

Give me a second I'll briefly update my slides.

If you can see now the slide is pretty much the same.

I just took out some bullet points.

The reason why I did this is the slide before this contains some hints on the limits of

Moore's law.

Basically what is Moore's law?

It's actually not a law in a sense like a physical law, a law of physics, a law of nature

but it's more a statistical observation namely that the number of transistors on a densely

packed integrated circuit doubles about every 24 months.

Often you will read 18 months.

This is because one of the managers of Intel was quoted with 18 months but I think the

statistical data is more to and covers more to 24 months.

What you can see here are the transistors per die.

This is an exponential axis of course because the law says it's doubling and you can see

here data from the 1960s up to the early 2000 years.

That's a pretty good match for this.

Still it's hard to imagine that this law will continue indefinitely because we cannot shrink

structures on the die below atomic sizes.

There is the assumption that Moore's law will someday come to hold.

We don't know when.

As you know last week I've been to the supercomputing conference and I did learn a new law there

which says that the number of people who proclaim the end of Moore's law doubles every 18 months.

I don't want to say when Moore's law will end but for now it is something we can rely

on at least for the next years.

This is a different diagram.

I took this from an article on the IEEE website.

Reading this is actually highly recommended so if you've got some spare time I can only

recommend reading this.

Does anyone know the current manufacturing process Intel is using for the Broadwell CPUs?

I'm hearing some numbers.

That's exactly 14 nanometers.

This article is from 2013 so 2014 it predicts we'd be using 14 nanometers which is actually

correct.

The red curve here is the node name.

Node name in this context means the nanometer size of the process.

The question is what does this mean at all?

Actually this node name is just basically today a marketing name.

It does not necessarily mean that structures on our integrated circuit actually have this

size.

I used to think that 14 nanometers means my whole transistor is 14 nanometers in size

which it actually isn't.