So, hello and welcome to the lecture on computer science in this week.
We started last week with chapter 2, which is about homogeneous and heterogeneous multivicam processors.
In chapter 1 we have already dealt with basic things that are in the modern microprocessor nowadays,
so from procedures to the CISC and so on, so I have to get everything back.
And now in this chapter 2 it's about feeling the existing microprocessors on the teeth,
so to take a look at how they actually look internally.
But first of all, I started last week with, yes, I wanted to motivate again why we actually got multicam architectures.
And in this context I also told you a bit about technology in general,
about microprocessor technology and how it has developed there.
Yes, that's what we've done, no lecture in technical computer science without the Mursche law.
I went into it again, so a very interesting graphic that shows us where the performance increases came from in the last few years,
either due to the technology, that's this green curve here,
technology-wise, that means we can just make it faster,
or just because we make our architecture smarter, and that's this blue curve here.
Okay, I went into it last week, then, yes, why is it that this green part has less and less influence,
you can see it especially very nicely here, so really steep rise here,
and then, after the turn of the millennium, it actually didn't go any further with the speed frequency, why?
It would have just gotten too hot, our chips would have melted under circumstances,
if we just kept turning the speed wheel, yes, you can see it again here, exactly, 11th, 12th,
exactly, we have now arrived at everything, I did that last week, let's quickly take out the data here.
Okay, exactly, and we want to take a closer look at that today, then we go back to the model,
what should help us in the first approach, how we can find out what the multicore system is,
if we can't get any further with the power, there could be two reasons,
one, we simply have too many, yes, our processors are simply on strike, or they are stopped for some other reason,
and these other reasons are the memory, that means we can't get enough quickly with the supply of data.
Okay, but now again here, so I have now taken out a newer graphic from the book of Patterson Hennessey,
Computer Organization and Design, the last edition, the RISC-V edition,
and yes, you can see that very nicely, over the years the different microarchitectures have been shown here
by interprocessors, and you can see, yes, there was a real increase to 3.0,
so here it was logarithmically applied, that means if it runs linearly here, then it is actually exponential,
so we had a real exponential increase here at the speed wheel,
quickly within only a few years from 200 MHz to 2 GHz, unbelievable,
and then the whole thing flattened out, then it went back to 3.6 GHz, and since then, yes, in this area we are moving,
but nevertheless we have managed to, yes, the black curve here, just as steep, even steeper,
is the increase in the loss of power, so the power consumption,
and yes, that then increases to the top value of 100 watts,
and 100 watts, yes, we are here today, the Beamer over there, who did not like me, did not work,
and then it increases to this value, and then there is already the thermic danger limit,
and from there we have managed to reduce the whole thing a little again,
yes, among other things, with cleverer techniques, for example turbo boost,
that one then always adjusts the voltage or the frequency.
Yes, in the first approach, the InSemos circuit technology applies,
that the loss of power is calculated from the product of capacitive load,
what I have done here, so a couple of lines and so on, to build up capacitive loads,
and times the operating voltage to the square, that is exactly the crux,
that it goes into the square, I have already told you about the Hensch's scaling last week,
that this is then problematic, times the frequency,
and yes, you can calculate here how it comes to 30, so the voltage,
the voltage is in the course of the years from 1982 to the present time,
from 5 volts to 1 volt is reduced, so we already have a saving for factor 5,
Presenters
Zugänglich über
Offener Zugang
Dauer
01:25:41 Min
Aufnahmedatum
2017-12-11
Hochgeladen am
2019-04-30 23:39:03
Sprache
de-DE
-
Organisationsaspekte von CISC und RISC-Prozessoren
-
Behandlung von Hazards in Pipelines
-
Fortgeschrittene Techniken der dynamischen Sprungvorhersage
-
Fortgeschritten Cachetechniken, Cache-Kohärenz
-
Ausnutzen von Cacheeffekten
-
Architekturen von Digitalen Signalprozessoren
-
Architekturen homogener und heterogener Multikern-Prozessoren (Intel Corei7, Nvidia GPUs, Cell BE)
-
Architektur von Parallelrechnern (Clusterrechner, Superrechner)
-
Effiziente Hardware-nahe Programmierung von Mulitkern-Prozessoren (OpenMP, SSE, CUDA, OpenCL)
-
Leistungsmodellierung und -analyse von Multikern-Prozessoren (Roofline-Modell)
- Patterson/Hennessy: Computer Organization und Design
-
Hennessy/Patterson: Computer Architecture - A Quantitative Approach
-
Stallings: Computer Organization and Architecture
-
Märtin: Rechnerarchitekturen