You may remember our article series How CPU Features Affect CPU Performance that we published in 2009. That time we analyzed the performance of a number of processors to understand what affected it. Then we updated our test method and decided to do the same, but in situations and tasks closer to real life. Just like the last time, we're going to start with AMD -- with its Socket AM3 platform to be exact.

The current AMD processor lineup seems a bit chaotic, because selling partially defective CPUs is surely better for the company than throwing them out. AMD makes a lot of CPUs that vary in cache type and size, as well as number of cores, and there's a strong temptation to give partially defective processors names, disable malfunctioning cache parts or cores and eventually sell them. Thanks to this innovative policy, the company's Socket AM3 series includes as many as three dual-core solutions with different L2 cache and even L3 cache, two triple-core products with and without L3 cache, three quad-core CPUs with and without L3 cache. That and a single-core Sempron. Having compiled a table of key features, we can now see some logic in the AMD Socket AM3 lineup.

The lineup logically proceeds from one core to six cores, passing by different L2 and L3 caches. As you can see, AMD mostly plays with L2 cache in relatively weak dual-core CPUs. And then it uses L3 cache as a universal "booster." Note the two simiarly strange processors: Phenom II X2 with two cores and a huge L3 cache and Athlon II X4 with four cores and no L3 cache whatsoever. Theoretically, the former may be a perfect choice for legacy software that isn't optimized for multithreading (but then it doesn't need the second core much, right?), and the latter may be meant for optimists who believe that four cores are better than three and two no matter the cache size. Our tests will tell if that's true.

This suggests the most interesting comparisons in terms of performance analysis.

1. Increasing the number of cores while cache size remains the same:
1. 1 core to 2 cores;
2. 2 cores to 3 cores;
3. 3 cores to 4 cores;
4. 4 cores to 6 cores.
2. Increasing cache size while the number of cores remains the same:
1. dual-core CPUs -- varying L2, adding L3;
2. triple-core CPUs -- adding L3;
3. quad-core CPUs -- adding L3, varying L3.
3. Fewer cores at larger cache (per core):
1. single-core vs. dual-core;
2. dual-core vs. triple-core.

As you can see, there's much to test and analyze. Note that to focus our attention on the above and eliminate other aspects, we had to test all CPUs at the same clock rate -- 2.6GHz in our case -- whether such models actually existed or not. That's our one step towards synthetics. But it's not that bad, because there are 2.6GHz Athlon II X3/X4 and Phenom II X3/X4 processors. As for the non-existing models, there are no 2.6GHz Sempron, Athlon/Phenom II X2 and Phenom II X6 CPUs.

Tests

As stated above, we conducted tests according to the newest test method (to be published yet, sorry). But there were several changes:

1. Since our task was large-scale and interesting, and all processors performed reasonably, with no quirks, we considered all optional benchmarks as regular ones, meaning that we included their results in the total score.
2. Since some of the processors we tested didn't actually exist, we considered Phenom II X4 810 a reference CPU with the score of 100 points for the sake of convenience (see the spreadsheet with complete results).

Traditionally, here's an Excel spreadsheet with complete results.

The topic of the first review in the series may not seem the most important. It's just that we're publishing articles as soon as results are ready. Sadly, it takes quite some time to obtain results according to our test method. Logically, we should've started with comparisons involving Sempron, but then you'd have to wait for another month. So we decided to publish what we had instead.

Now let's get down to business. Today we shall try to answer the following question. What's better: a triple-core CPU with 512KB L2 per core or a dual-core CPU with 1024KB L2 per core? On the one hand, the former has an additional core. On the other hand, each core of the latter has twice as much cache available. This isn't as obvious as it may seem.

3D visualization

Like we said, it's not as obvious as it may seem. Suprisingly, only one 3D suite out of six can actually benefit from the third core. While the remaining five don't like the halving of L2 cache at all. Most likely, these suites just cannot use the third core. 3ds max developers did a nice job optimizing their software. Sadly, they are in the minority these days.

3D rendering

The gain from the third core is almost perfect for these benchmarks. But that shouldn't surprise you, 512KB L2 is quite enough in this case, because the test scene is divided into small enough pieces.

Scientific calculations

This is a more complex situation. It seems that CAD suites handle large data sets during calculations, and they cannot use the third core. Mind you, they often ignore even the second core. Optimized for multithreading, Maya, Mathematica (we're using a multithreaded variant of the MMA benchmark from 2010 on) and MATLAB help the triple-core CPU gain the lead.

Bitmap processing

The ±1% difference is well within the measurement error, so we can only point out the cache-loving ACDSee and Photoshop which is well optimized for multithreading. This makes the triple-core CPU the leader again.

Data compression

7-Zip can use the third core, while RAR obviously ignores it. As for unpacking (which is single-threaded, we checked), L2 size plays an important role in it. Still, 7-Zip performance gain exceeds what's lost during unpacking. The triple-core CPU wins.