The first winsome peculiarity we can see is stability of the mode with the minimum command rate (1T). We would manage to use this latency with just a few memory modules, even at the DDR2-800 frequency. And we almost always had to use 2T for DDR2-1066. And now we can use the minimum latency not only at the typical frequency, but also at 1600 MHz, as well as with four memory modules operating at 1333 MHz (we installed memory kits from different manufacturers: Apacer and Kingston DDR3-1333). We like the latter fact even better, because we heard about possible limitations in frequency, if four modules are installed. And in reality we didn't have to reduce frequency and retained the same timings, unexpectedly including 1T! Both motherboards demonstrated the same situation in our tests. To all appearances, necessary modifications were already added to their BIOS. It's good news. Even if it makes no sense to install four memory modules at once in most cases. But a user may want to expand memory capacity in future, and it can be done without sacrificing performance.

This latency does not have a radical effect on performance (perhaps you may be even forced to use 2T with some less lucky memory modules; this latency had been initially introduced for better compatibility). But the fact itself indicates that the memory controller in AM3 processors is finetuned even in the first revisions. On the other hand, it's the third generation of integrated controllers for AMD already, the company is well versed in this field.

What concerns performance comparison with DDR2 and DDR3 memory, DDR3-1333 brings only an advantage in memory bandwidth, while random seek latencies are higher than in DDR2-1066. DDR3-1600 is better than DDR2-1066 in all parameters already. And vice versa: it's clearly a bad idea to use DDR3-1066, as its bandwidth is similar to that of DDR2 at the same frequency, and its latencies are higher. By the way, it unexpectedly turned out that Apacer modules (DDR3-1333) worked well at 1600 MHz, their voltage raised only to 1.6 V.

But the real contrast between computers with DDR2 and DDR3 memory appears in the overclocked mode. As we already noted, in order to overclock our Phenom II X4 810 we had to compensate memory frequency to bring it closer to DDR2-800, even if its typical mode was DDR2-1066. There is no such restriction for DDR3 memory, and its top stable frequency is a tad higher. However, this difference will be smaller, if you deal with a processor that has an unlocked multiplier (so that you can preserve DDR2-1066 mode in overclocking). It's also obvious that an overclocked processor needs faster memory (and can use it effectively), so practically in all tests the difference is higher that we can see comparing DDR2-800 and DDR3-1600 results at the typical frequency, especially for multi-threaded access. So DDR3 memory is more justified for enthusiasts and overclockers than for those users that employ typical modes.

But before we can draw conclusions, we should take a look at results in real applications, not only in synthetic tests. Let's take the most popular tests here: archivers and games.

In fact, our conclusions after synthetic tests hold true here. When processors operate at their typical frequencies, transition from DDR2-1066 to DDR3-1333 yields only a little advantage noticeable in actively multi-threaded applications (that is when all cores read/write from/to memory controlled by different processes). Two first games in the table belong to such applications. But if we compare DDR3-1600 with DDR2-1066 (which makes sense, because both products operate at maximum frequencies available without increasing the reference clock rate and overclocking a CPU), the difference grows more noticeable. It also becomes noticeable that in a number of tests a CPU operating at its typical frequency just does not need such fast memory. So we can see performance differences better in tests with overclocking. And it's up to you to decide whether all these advantages justify extra expenses for DDR3 memory. What's important, processors for Socket AM3 give you a choice between platforms supporting both memory types. So users (as well as manufacturers) can choose proceeding from their own reasons.

As for us, we want to investigate into another theoretically justified issue: whether the new memory type affects performance of the graphics core integrated into a chipset. Motherboards with AMD 790GX are usually equipped with a video buffer (in its top modification it's DDR3-1333). However, 128 MB of video memory is not enough these days, so a part of main memory has to be used to store textures. We can try to figure out the effect by using old games, such as Quake 3, and the difference will certainly be great. But we decided against it. We'll stick to real conditions. AMD 790GX has a powerful graphics core that allows to play relatively modern games, even if with low settings. Theoretically, performance in these games should be limited by the graphics core in the first place. But what if there are aspects?

We can see a noticeable difference only in Crysis. In other cases it looks like games (at our resolutions and graphics modes) do not use more than 128 MB of video memory.

Conclusions

To all appearances, the evolutional strategy offered by AMD is adequate to the real situation on the memory market. On one hand, a small advantage in typical modes does not justify the price difference between DDR2 and DDR3 memory kits of same size and class (at the time this article was written). On the other hand, PC manufacturers can buy standard DDR3 modules for much better prices than those offered in retail stores. It seems that this migration becomes expedient even from the economic point of view, especially considering the existing performance advantage. Another category of users that might be interested in DDR3 memory (for a new computer) is consistent overclockers. Processors operating at increased frequencies may squeeze a better potential from high-speed DDR3 memory. As a result, a user may get higher performance. The difference is not big, but it's not that small either. You would get a similar performance gain from a CPU with a 100-200 MHz higher clock rate.

Write a comment below. No registration needed!