How CPU Features Affect CPU Performance
When new CPUs were announced several times a year (two or three times actually), and any new architecture was indeed new by almost 100% (well, by at least 70%), it was very easy to conduct relative tests: you took two or three CPUs, compared them with each other, and you got new interesting data. Well, times have changed: when you take a look at an allegedly new processor, you see what parts are copied from what prototype (or competing solution), what features are preserved to reduce manufacturing costs, what features are added because the manufacturing process makes them almost free of charge, and what units are implemented in a mediocre way because of the manufacturing process. And then it dawns on you that comparison results of some Core 2 Duo E8500 and Core 2 Duo E8700 will be too predictable and too dull to read: you need test results of any processor, plus logic, common sense, and a little of math to calculate test results of any other model. Even in case of different CPU architectures, two or three comparisons will be enough to predict all other results even without tests.
On the other hand, good old processors must have been manufactured in the same way -- there is nothing new under the sun. But we knew less at that time. And as Solomon put it, in much wisdom is much grief. This series will include four articles for those readers, whose peace of mind agrees with the previous paragraph: for those who don't want to read dozens of comparisons, because it's easier for them to sort everything out and then draw conclusions on their own, trusting their logic and common sense. So, this series can be called "articles for IT snobs".
Besides, these articles will certainly please those users, who believe in objective laws and consistent patterns. If you don't think that it's possible to predict performance of the Core 2 Duo E8700 based on objective test results of the Core 2 Duo E8500, these articles are not for you. If you don't believe that performance of the Phenom X3 8450 determines the speed of Phenom X3 8750, you shouldn't read this article. For the rest of you we decided to give simple test answers to simple user's questions: what happens, if you take one and the same processor core and put it into a situation, when it can cope with simple user's tasks (to this or that degree)? For this purpose we've taken two most advanced cores from two popular manufacturers of x64-processors -- AMD Phenom II X4 and Intel Core i7. Frequency? What's the difference?! Cache size? It does not matter either! We've tried to evaluate exploitability and scalability of the architecture. Implementations may vary to better or worse, of course. However, all of them have common features, which are practically impossible to extirpate without changing the general concept of development. We wanted to evaluate the general concept (applied to real modern software, of course).
Speaking of reality: our method presently does not contain tests, where one application would run on the background of the other (as it often happens in real situations). We've explained our reasons many times: a multi-application test cannot possibly be representative in terms of obtained results because of the unlimited number of program combinations. For example: we decided to benchmark Unreal Tournament 3 with 7-Zip archiving and Avast scanning in the background. We have five questions:
- Why Unreal Tournament 3, and not World in Conflict?
- Why 7-Zip, not WinRAR?
- Why Avast, and not AVG?
- And finally: why game + archiving + antivirus, and not video encoding + rendering + arithmetical computations?
- And by the way, will such results be of any use outside the above-mentioned combination of programs? What happens to the results, if, for example, 7-Zip is replaced with another archiver?
An honest answer to Questions 1-4 should sound like this: "Just because we wanted to do it" -- and there is no other honest answer to that. Indeed, it's either test the entire variety of software combinations (mission impossible) or select one combination using the highly scientific method of picking out of thin air. There is only one answer to Question 5 as well: "No, they will be useless. God knows, we don't." That's exactly why we don't use multi-application tests: it's much better to obtain correct test results in a somewhat idealized environment than get vague results in a pseudo-real environment (to be more exact, in one of real environments selected from many thousands of others).
In this article we are going to evaluate performance scalability of AMD Phenom II systems depending on the number of processor cores -- we'll take a computer based on the AMD Phenom II X4 940 Black Edition and benchmark it with one, two, three, and four enabled cores. Thus, we've created an ideal test situation: nothing changes in the testbed except for the number of cores in the CPU, and there are no physical changes at all. Some of you may contradict that real single-, dual- and triple-core processors do not come with so much cache. However, we did not even try to compare processors. We put the top solution into various situations and analyze its behavior. These data will be sufficient for "IT-snobs". They will assume the rest on their own proceeding from logic and common sense.
We'd like to make only one observation about our testbeds: our way of installing memory modules. As we already mentioned, it's not possible to level LGA1366 and Socket AM2+ computers in the correct manner, because the logical configuration of three 2GB modules for the triple-channel memory controller in the LGA1366 platform cannot be reproduced on the Socket AM2+ platform, as its memory controller has only two channels. So to obtain the same memory size, one has to combine 2GB and 1GB modules. We'll describe the problem of choosing memory modes in the Socket AM2+ platform in the next article. As for now, we can only say that we used "2+1 per channel in the unganged mode".
* It goes without saying that we used one and the same processor with a different number of cores enabled in BIOS.
Write a comment below. No registration needed!