How CPU Features Affect CPU Performance
|
Video encoding
|
DDR2-800 4-4-4-12 unganged |
DDR2-800 4-4-4-12 ganged |
DDR2-800 6-6-6-18 unganged |
DDR2-400 4-4-4-12 unganged |
Canopus ProCoder ↓ |
00:03:46 |
00:03:46 |
0% |
00:03:48 |
-1% |
00:04:00 |
-6% |
DivX ↓ |
00:05:04 |
00:05:03 |
0% |
00:05:08 |
-1% |
00:05:30 |
-8% |
Mainconcept ↓ |
00:07:54 |
00:07:55 |
0% |
00:07:58 |
-1% |
00:09:27 |
-16% |
x264 ↓ |
00:10:48 |
00:10:47 |
0% |
00:10:55 |
-1% |
00:12:33 |
-14% |
XviD ↓ |
00:05:46 |
00:05:37 |
+3% |
00:05:49 |
-1% |
00:07:33 |
-24% |
Results of switching between controller's modes and raising timings are traditionally dull. We cannot say that XviD is well-optimized for multiprocessing, and the ganged mode improved its performance by 3%. All our tests show performance drops with the halved memory bandwidth. And there is almost complete correlation between the performance drop value and multiprocessing optimizations (the better the latter, the higher the former), except for XviD.
Java
This test has good multiprocessing optimizations, and it slowed down in the ganged mode. In other respects, these results resemble the situation with archivers.
3D games
|
DDR2-800 4-4-4-12 unganged |
DDR2-800 4-4-4-12 ganged |
DDR2-800 6-6-6-18 unganged |
DDR2-400 4-4-4-12 unganged |
STALKER: Clear Sky ↑ |
54.4 |
54.6 |
0% |
53.4 |
-2% |
48 |
-12% |
Devil May Cry 4 ↑ |
194.1 |
194 |
0% |
193.6 |
0% |
184.7 |
-5% |
Far Cry 2 ↑ |
41.1 |
41 |
0% |
41 |
0% |
31 |
-25% |
Grand Theft Auto 4 ↑ |
60.8 |
61 |
0% |
60 |
-1% |
46 |
-24% |
Lost Planet ↑ |
42 |
42 |
0% |
42 |
0% |
38 |
-10% |
Unreal Tournament 3 ↑ |
133 |
133 |
0% |
128 |
-4% |
109 |
-18% |
Crysis: Warhead ↑ |
47.4 |
48 |
+1% |
47 |
-1% |
42 |
-11% |
World in Conflict ↑ |
40.4 |
39 |
-3% |
37 |
-8% |
28 |
-31% |
The ganged mode is traditionally boring, only two games responded to timings -- these are the best games in terms of multiprocessing optimizations. But the most interesting situation occurs, when the memory bandwidth is halved: the normal performance drop is about 10%, 20% is not rare, and World in Conflict breaks a record in this test dropping its fps by almost one third. However, it did not come as a surprise to us: games and archivers were proved the most "memory-intensive" groups in our previous tests.
Conclusions
We've obtained lots of various results, but the bottom line can be expressed in a couple of words: neither memory controller's modes nor messing with timings have significant effect on performance of systems based on Phenom II. We've found only one application that slowed down by 8% in the 6-6-6-18 mode. But general results with "good' and "bad" timings differ by a value within the measurement error. Ganged/unganged modes also turned out to be just a funny toy for enthusiasts: there is again only one application that differentiates between them. However, we can at least establish a fact that the unganged mode gains dividends more often. You can see it even in the total score: it's higher by one unit than that for the ganged mode, all other parameters being equal. Thus, practical recommendations to owners of Phenom II systems are quite short: enable the unganged mode in BIOS and forget about this option for good. It looks like low-latency memory is useless as well, it may attract only overclockers. However, only these people usually buy such modules anyway, so it's OK.
On the other hand, we cannot say that memory does not affect performance at all: when we degraded all its parameters by twofold -- performance generally dropped by 11%, and some applications slowed down by one fourth or even more (by the way, they were mostly 3D games). But in order to actually see the real slow-down, we had to install very slow memory modules -- it's hardly possible to find such memory modules these days (probably only in servers, but they usually use register memory DDR2-400 ECC). Only in this almost synthetic situation we managed to see some effect of the concept described by designers of new memory standards: the more cores in a processor and the better multiprocessing optimizations in applications, the higher memory bandwidth you need.
We'd say that our test results are slightly disappointing, but they are really useful -- to all those users, who thought to upgrade from DDR2-800 to DDR2-1066 or even faster DDR3 (needs a new processor). Performance gains after the upgrade from DDR2-400 to DDR2-800 are in front of you (these are actually performance drops caused by switching from DDR2-800 to DDR2-400, but it does not matter. Any performance drop becomes a performance gain, if you look from the other side). These are the consequences of twofold difference in memory speed. That is even if we assume that Phenom II is perfectly scalable, you will have to upgrade to DDR3-1600, if you want to repeat this effect with DDR2-800. It's up to you to decide whether the game is worth the candle.
In next articles we'll analyze the effect of various factors on performance of another popular platform, Intel Core i7.
Write a comment below. No registration needed!
|
|
|
|
|