평 runway

Let's start. So I think we...

So we've been through the roofline model the last time and this essentially closes the introduction to computer, to the basics of computer architecture.

Hello?

To the basics of computer architecture, do you still have questions regarding the first part?

Sorry. No? Okay. So today the second part starts with... So apologies for the old slides. So these are the old slides from two years ago now.

I didn't get around to actually update them to the new theme, to the new BIMA template, but they are still accurate. So what we are going to do is essentially...

Sorry. Today, discuss the motivation for multi-core processors. So the second part was dubbed the road to extra scale. And one of the most important components there are the CPUs.

So what we discussed in the first part was essentially how more or less a single CPU core operates.

So the question is what are the challenges in the processor designs? What do we have to do? What are the developments and how do certain physical aspects, etc., play into the design of general processors?

And what are the trends and where do we go to? So that's why I'm trying to motivate why multi-core processors are needed.

And I don't know if anyone ever heard of Moore's Law? No? Okay. A little bit. So Moore was one of the co-founders of Intel. And it's not really a law, it's more an observation.

And the observation he had in 1965 was that the available transistors on a chip double every year or every two years.

So between 18 to 24 months the number of transistors here in green double every 18 to 24 months. So this is for the Intel processor lines.

And so it's not a physical law in that sense, it's more the observation from an economic and technical perspective.

So it's not really that the advances in physics or in material sciences allow us to double the available transistors every two years, but actually the economic aspects.

So most of the considerations in modern processor designs are a trade-off between costs and actual performance increases.

So one of the biggest challenges is really the manufacturing process to actually manufacture such a chip. So do you know what a wafer is? Or how approximately a CPU gets manufactured?

Okay. So the way it goes is that you have what's called a wafer. And a wafer is a disk-like thing where silicon is on top.

And so it's usually a disk, obviously round, so this is a circle, a perfect circle. And it's a few centimeters in diameter.

And what's happening during... Hi. More to come? From Nuremberg. Oh, God. Lots of traffic?

Yeah, the train was delayed a little bit. Okay. Happens. Do more people come or are you the only ones coming from Nuremberg?

Okay. Right. So manufacturing of processors. So you have this wafer, which is silicon.

And what happens when a so-called fabric actually manufactures those chips is that on this wafer you have a few hundred processors manufactured in one step, right?

So the way how it's done is usually that, similar to taking a photograph, you burn the structure of your transistors into the silicon.

So burning is a little wrong there, but more or less with photonic beams, with lasers, you actually create the transistor structures over and over again.

And as I said, this wafer is probably, I think, about 10 centimeters in diameter or something like that.

And you try to put on structures in the nanometer scale onto that, right? So pretty complicated.

And of course, the higher density, the lower the structures become, the higher the failure rate is, right?

So usually the yield for a processor, the yield of a fabric, so the actual good processors that come out of the manufacturing line is about 10%, right?

So that's pretty low. And the higher the technology advances, the more failures you actually get.

And this is multiple reasons. The silicon might not be as pure as wanted, right?

Then you have certain mechanical issues and physical issues you can't really control, right?

And what you see usually is that you just have a spot where the produced silicon is faulty, right?

So you have to throw it away. And as technology advances, the manufacturing process, of course, gets better and better, right?

So it's more feasible to put more transistors with a lower structural length on such a wafer.

So it's economically feasible to double the transistor on a chip every two years.

So that's more or less the meaning behind it. It's not that it's physical limitations.

Well, it's also, so nowadays it's becoming also physical limitations, but back in 1965 it was definitely not physical limitations,

but more the technological slash economical limitations in the manufacturing process.

So what we also see here on this is that the clock speed of the processors is saturating, right?

As well as the power consumption and the performance per clock, obviously, right?

So that's one of the most important observations that we see for modern processors, right?

Okay, so that's just something to keep at the back of your head at the moment.

So what we also see is that what you see on this graph, so the red line here is the node name,

which is if you see marketing slides or announcement of new processor technologies,

you usually hear something like this new processor is manufactured with 14 nanometers, right?

This is the node name, right? So it goes continuously down and starts to saturate eventually, right?

At the moment we are at seven nanometers for some chips, okay? Which is actually quite small, right?

And the other lines is the gate length, right? If you look at a transistor,