We proceed with our series of articles devoted to analyzing various factors that affect performance of modern x86-64 processors. In this article we'll try (as far as it's possible) to answer the question about performance of a single processor core in comparison with two main competitors. For this purpose we use BIOS Setup in two motherboards for LGA775 and Socket AM2+ platforms (Biostar TPower I45 (IP45A-A7P) and ASUS M3A32-MVP Deluxe/WiFi-AP), which allow to disable all CPU cores but one.

Some of you may be surprised that we modified our LGA775 testbed. ASUS Maximus Extreme was replaced with a less functional motherboard, not with the top chipset at that. First of all, it was done to provide identical operating conditions for both cores: the motherboard from BIOSTAR works with the same DDR2-800 memory as our standard platform for Socket AM2+. Besides, ASUS Maximus Extreme cannot disable cores in Core 2 processors, that's another reason to replace the testbed for these tests.

As you can see, operating conditions are practically identical: only one core is working, memory type/frequency/timings match, core frequencies are identical as well (in case of the LGA775 platform, we had to use the feature of Core 2 Duo E7200 engineering sample to change its multiplier to a custom value). Meticulous but hasty readers may ask a question about the cache. Wolfdale core has a 3-MB L2 cache (which in this case is used only by a single core), while Agena (represented by Phenom X4 9850 Black Edition) has only a 2-MB L3 cache.

However, it's not as bad as it seems. Don't forget that Intel processors come with non-exclusive caches. So maximum volume of cached data is determined by a cache of the maximum size. And the AMD platform comes with exclusive cache (L3 is adaptive-exclusive, but theoretically, it should work as exclusive cache in case of a single core). Thus, maximum volume of cached data is actually the sum of cache sizes of all levels. In our case it's 128 (L1) + 512 (L2) + 2048 (L3) = 2688 KB. It's smaller than the 3072-KB L2 in Wolfdale by 13%. But it's the lesser evil, as the main competitors do not consult each other on the size of caches to facilitate our task.

Some of you may think that L2/L3 cache must be disabled to obtain ideal test conditions. It's a disputed point. Firstly, it's not a real-life situation – nobody uses modern x86-64 processors with disabled L2 Caches. Secondly, in this very case the AMD core would have got a very serious advantage, as L1 Cache cannot be disabled (at least in BIOS Setup). And L1 Cache size in Agena is twice as big as in Wolfdale.

Testing

As always in such cases, we are particularly but not exclusively interested in performance results obtained in real-life applications, so we replaced diagrams with tables again, and virtual points with raw test results. The "Advantage" column shows how much faster Wolfdale is versus Agena in percents (as it's the most frequent situation). So in those cases when Agena is winning, this value will be negative and marked bold and red (as it's a very rare situation).

3D Modeling and Rendering

Strange as it may seem, Maya's render engine apparently prefers the Ê10 architecture. The 5% advantage is not a deciding factor, of course. Besides, 5% of advantage in rendering speed will hardly weigh down the 20% defeat in interactive operations with this package. However, it's one of the rare cases when Agena becomes a leader. 3ds max definitely votes for Wolfdale, Lightwave joins the club (with much less ardor though).

CAD/CAM

K10/Agena is slower by 10-20% in all tests without exception. As we'll see later, this statement will be a refrain of this article.

Compile

Unfortunately, we use only one compile test so far, so we cannot compare results. We can only state that Agena is defeated again.

Professional photo processing

Rotate depends much on memory speed, and the integrated controller probably helps Agena narrow the gap, so its defeat is minimal in this test. In other respects the situation is far from peachy: Wolfdale is faster by 20% in most cases, sometimes even getting close to 30%.

Scientific applications

Agena has set its own absolute record in Maple: 17% advantage over Wolfdale. It's the biggest advantage this AMD core managed to produce in our tests. And MATLAB is full of surprises. The first surprise was demonstrated with Phenom, and here it goes again with a single core tests. This time AMD Ê10 "gave a jingle" in LU test (matrix decomposing), where it's 76% as slow as the Intel Core architecture. That's another "tragic mistake" again, of course – in a library or in a processor, or somewhere else. But there are even more phenomena in this article (in the most unexpected places). Frankly speaking, we are getting tired of them...

Web server

By the way, the second phenomenon is Real World PHP & MySQL test in PHPSpeed. Agena core has lost almost 60% in comparison with 3-4-5% in other tests. Don't you think that such a situation is typical for this reincarnation of the core? We have some ideas about it, but we'll publish them a tad later.

Archivers

This group of tests shows a traditional picture, typical defeat of Agena.

Encode

Canopus ProCoder apparently votes for Agena. But the other seven applications are of the opposite opinion. Agena is slower by more than a quarter in two codecs – OGG Vorbis and XviD. Both are open source projects. However, open source LAME and FLAC are less radical in their preferences, so we cannot draw far-reaching conclusions from this coincidence.

Games

Quite a standard situation: Agena is defeated in each test, starting from 7% (in Call of Juarez, which depends much on a graphics card) and to 32% (strange as it may seem, it's in no less 3D-dependent World in Conflict). There is nothing unusual here.

Non-professional photo processing

Over a half of applications (3 out of 5, to be more exact) demonstrate the advantage of Wolfdale over Agena by a quarter or even more. For example, Agena is defeated by 68% in ACDSee! That's the third (and fortunately the last) Phenom's phenomenon in these tests. The author of this processor's name apparently had a wry sense of humor...

Conclusion

That's a sad picture. In modern conditions, even all other things being equal, Agena core is outperformed by Wolfdale almost everywhere. Moreover, the lag is mostly significant (over 10%). However, this epithet can sometimes be replaced with "insignificant" (SolidWorks/Graphics, Photoshop/Rotate, Mathematica, many PHPSpeed tests) or "horrible" (MATLAB/LU, Real World PHP & MySQL, ACDSee). Strange as it may seem, the latter serves as a reason for our optimism.

Let's recall the recent history of x86-64 processors. AMD Phenom started its existence from such a thundering screwup, which we hadn't seen since the notorious bug in Intel Pentium. What does it tell us in the first place? It means that the processor was designed in a hurry, as a rush job, so it most likely wasn't properly tested. Then we had an opportunity to see erratic behavior of the improved Phenom in MATLAB Sparse tests. And now, in single-core mode, we again stumble upon MATLAB (LU function), PHP+MySQL, and ACDSee.

From an optimistic point of view, hoping for the best, we can assume that all these Phenom's phenomena are caused by such defects. They do not result in bugs, unlike the notorious TLB error, but they significantly slow performance. By the way, from our perspective, this category of potential strange behavior also includes all defeats with performance gaps of 25% and higher. Let's hope that all these defects can be repaired, which will be probably done in the next revision of the K10 core. It won't become the ultimate solution (too late for that...), but it will help narrow the architectural performance gap to a more or less decent level (say, below 10%) – who knows, may be it will happen.

Write a comment below. No registration needed!