Overclocker DDR3 for Socket AM3
Choosing ample memory for Phenom II.
June 5, 2009
As is well known, Phenom II for Socket AM3 does not need any special memory kits. These processors have dual-channel memory controllers, which, as we have already found out in our recent tests, are quite unpretentious. There were practically no compatibility problems even with early BIOS revisions. And now you can afford extravagant solutions, like four memory modules (two per channel) operating not only at the typical 1333 MHz, but even at higher frequencies. Timings won't be much worse here (in case of high-quality overclocker modules, of course). But AMD decided to play it safe by recommending to use only two DDR3-1333 modules (one per channel). By the way, this overclocker rule still holds true for any platform. That is, it always makes sense to start with two memory modules and use the remaining slots to upgrade when necessary. But one advantage of the AM3 memory controller is operation at high frequencies with all memory slots occupied. It sticks out in test results, just because it's not axiomatic. Frequency limitations, that appear as capacitive load on the memory controller increases, are of physical nature, they have always existed on both AMD and Intel platform.
However, when we say "regular DDR3" memory, we mean memory modules tested and marked for operation with existing platforms and discrete memory controllers in Intel X38, P45, and the like. This is not very important for standard memory modules operating at typical frequencies, of course, except perhaps you will manage to choose lower timings and reduce the command rate to 1T (this parameters in chipset controllers for the LGA775 platform is usually hardcoded to 2T, only some NVIDIA chipsets allow to adjust it). However, if you have memory modules for overclockers, you can look up the recommended frequency and timings on the sticker only as background information. Manufacturers cannot guarantee stable operation in overclocked modes, of course, even on those platforms used for internal tests, because successful overclocking in real conditions depends on memory characteristics only partially. And when a platform differs principally, results will certainly be different as well. That's why we decided to perform another test of memory kits for overclockers from various manufacturers, in order to see how the recommended modes would agree with modes attainable in practice.
Here is a short explanation that will help you understand test results: physically, memory modules with typical frequencies of 1600 MHz, 1800 MHz, or even 2000 MHz can be based on the same chips. Of course, module manufacturers have an opportunity to sort and select chips that can operate at higher frequencies with reduced timings or, vice versa, chips that allow to raise memory frequency with alleviated timings. But it's all relative. By the way, it's adequately reflected on prices. DDR3-1600 modules with reduced latency can be cheaper or more expensive than DDR3-1800 depending on their cooling systems and series. The rule of thumb is to choose memory modules with the highest frequency, if you want to overclock them. And vice versa, if you don't want to overclock your memory system much, and reduce timings instead, you should choose low-latency memory modules. This holds true only for comparison of memory modules from the same manufacturer.
- CPU: Phenom II X4 955BE
- Motherboards: ASUS M3A79-T Deluxe, MSI 790FX-GD70
- Graphics card: Palit GeForce GTX 275
- HDD: Seagate ES2 SATA II 750 GB
- Cooler: Zalman CNPS9700 AM2
- Power supply unit: SeaSonic M12D 750 W
- OCZ OCZ3G18004GK, 2x2 GB (max. recommended mode: DDR3-1800 9-9-9, 1.9 V)
- Patriot PVS34G1600LLK, 2x2 GB (DDR3-1600 7-7-7-20, 1.9 V)
- GeiL GV34GB1600C8DC, 2x2 GB (DDR3-1600 8-8-8-28, 1.9 V)
- Mushkin XP3-14400, 2x1 GB (DDR3-1600 8-8-7-20, 1.85-1.9 V)
- Transcend DDR3-1600+ CL7, 2x1 GB (DDR3-1600 7-7-7-20, 1.8 V)
- Windows Vista SP1 64 bit, ForceWare 182.50, AMD OverDrive 3.0.1
- RMMA 3.8
- 7-Zip 4.65 x64
- Photoshop CS4
- GTA IV: built-in benchmark, Texture Quality: High, Render Quality: High, View Distance: 52, Detail Distance: 100, Vehicle Density: 100, Shadow Density: 16
- Unreal Tournament 3: built-in benchmark, Ranch demo (medium-sized map), all Very High except for Post FX and Ambient (High)
- FarCry 2: built-in benchmark, High settings
- World in Conflict: built-in benchmark, DX10, Very High settings
Prior to memory tests, we'll investigate into the remaining open question: how DDR3 performs versus DDR2 with a 900-Series processor. We could have answered this question for the slowest models in this series even without tests: the effect from upgraded memory must be weaker than for processors from Series 800, which have smaller cache, the other parameters being the same. But in case of Phenom II X4 955 3.2GHz it's theoretically possible that the memory exchange rate may be a bottleneck. We've got this processor in our testlab. While our CPU "department" is finishing an updated test procedure, we've taken several tests to illustrate this article. By the way, as we have taken testbed configurations and all tests, applications, files for archiving, etc. from the future CPU test procedure, our test results won't be compatible with those of previous reviews.
|GTA IV, fps
|Unreal Tournament 3, fps
|FarCry 2, Ranch, fps
|World in Conflict, Very High, fps
There is some effect from faster memory in archivers and Photoshop (apparently in all tests with intensive computations of data, which volume exceeds the size of cache in any processor). But 6MB of L3 cache seems to be enough for games to reduce the difference in results to a measurement error. We can take into account differences in BIOS optimizations and say that DDR2 is used efficiently, while DDR3 can be several percents faster in future. But to all appearances, gamers can save on the platform costs without sacrificing anything by staying with DDR2 even after the rollout of 900-Series processors supporting DDR3. For example, they can get performance gains by spending the saved money on a more powerful graphics card.
However, prices for DDR3 are dropping, and there appear more attractive motherboards supporting DDR3 memory, so the demand for such configurations will be high. Especially among overclockers, of course, who try to squeeze maximum frequency from everything they can lay their hands on. And whatever you may say, DDR3 has an incontestable advantage in this parameter. So we proceed to the main point of our today's analysis: comparison of overclocking potentials of dual-channel memory kits from various manufacturers on the Socket AM3 platform.
Click to enlarge.
The 7-7-7 mode for DDR3-1333, just like the 8-8-8 mode for DDR3-1600, is the maximum mode specified in JEDEC. While most inexpensive DDR3-1333 modules without heatspreaders designed for the standard voltage are CL8 and CL9 models, DDR3-1600 memory kits are designed for overclockers, so we can see CL8 or even CL7 modules, which already overshoot the standard.
But it turns out, regardless of the specs, all modules taking part in our tests, except for Patriot, could work in DDR3-1600 mode with 7-7-7-20 timings, minimum command rate, and voltage not exceeding 1.8V. And the memory kit from Patriot was unstable even with the voltage increased to 2.1V, although this mode was recommended by the manufacturer. This memory kit disappointed us in the other tests as well -- we had to reduce the CPU NB frequency of the overclocked processor with this memory kit, hence inexpressive results in the tests. But formally this controversial memory kit takes the first place in frequency and timings. The effect of CPU frequencies on performance is higher than the influence of several extra MHz for memory.
The other memory modules demonstrate similar performance, so it's difficult to appoint one leader. We can see that 2GB memory kits overclock just as well as 4GB kits. That is their frequencies, memory bandwidth, and latencies do not depend on module capacity, although DDR2 modules showed this dependency. Only the memory kit from GeiL demonstrated higher latency operating at the same frequency and timings as the other contenders.
We shall not discuss whether one should install DDR3 memory for Phenom II. On the one hand, the answer is clear: if you want to save some money and not lose much (or any) performance, it will be reasonable to redistribute your budget and buy DDR2 memory, which is still cheaper. It concerns high-end processors just as low-end ones. On the other hand, in our CPU tests we'll use platforms with both memory types, and we'll keep tabs on how the situation will develop, including BIOS optimizations.
As for the relatively old overclocker DDR3 memory kits for other platforms, they work well with Phenom II, and in most cases they even significantly exceed their recommended frequencies. On the other hand, our tests show that it does not make much sense to use maximum frequencies higher than 1800 MHz. Such frequencies are not necessary even to overclock processors with locked multipliers. And it's possible to make memory run at such frequency only by reducing other multipliers and clock rates, which is bad for the overall performance. However, high-quality DDR3-1600 memory can work at frequencies higher than recommended, and DDR3-1800 covers CPU needs and can offer more.
We express gratitude to Palit and Seasonic for providing parts for our testbed.
Write a comment below. No registration needed!