AMD Phenom II 9xx CPUs With DDR3 Support
The first tests with the new test method.
June 15, 2009
Upgrade to Socket AM3 and, correspondingly, to DDR3 memory was initiated in Low-End models of Phenom II. Our further tests demonstrated that the effect of higher memory speed on processors with reduced cache was quite noticeable, although it mostly revealed itself in overclocked modes. But the evolution towards DDR3 (compatible with DDR2) has some economic reasons as well (the new memory type may be expedient to use for big system integrators even now, because producer prices for memory of various types differ greatly from retail price tags). There are ideological reasons as well, of course (to be more exact, marketing reasons, because many users still pay attention to progressive differences in specifications). So upgrading the 9xx series to Socket AM3 was only a matter of time, and all new additions to Phenom II will be designed for this socket (to be more exact, they will be compatible both with AM3 and AM2+).
Today we are going to test two 9xx series processors including the new flagship Phenom II X4 955. To differentiate new processors from the older models supporting DDR2, the number "5" is added to their index. For example, the model 945 differs from the former flagship (model 940) only by DDR3 support (and a locked multiplier, because such a bonus is only available in top models, which is now X4 955).
Dynamic overclocking and memory profiles
As you may know, with the Core i7 rollout Intel decided not to fix maximum CPU clock rates in specifications. Instead, the company introduced the so called Turbo Boost, that is it legalized overclocking in the automatic mode. Small wonder that AMD will take further measures to popularize overclocking, considering the high overclocking potential of Phenom II. The solution suggests itself: implement dynamic overclocking on the software level with AMD OverDrive, which functionality is constantly growing. In particular, advocates of classic manual overclocking will be pleased with the new option to specify a CPU NB multiplier, which could have been done only in BIOS up to now.
Smart Profiles promise the same functionality as Turbo Boost with a single exception -- users will create a profile for each application, which they want to run faster. A profile specifies a clock rate for each core (a multiplier). Programs that don't use all four or three cores efficiently, are assigned to the overclocked cores, while clock rates of the other cores are reduced to conserve power. It's possible to reach higher performance gains this way than with hardware overclocking, because this method takes into account peculiarities of a given application, not just formal CPU usage. The drawback is that you must create these profiles. But it's an excellent opportunity to combine performance gains with power saving for those who like to configure their computers for maximum performance (you can specify reduction factors for some background applications in absence of a user, that is it's finally possible to implement Cool'n'Quiet functionality for an overclocked processor). We like the option to specify a starting priority for applications: all levels are supported, from minimal to real time.
Although Smart Profiles are mostly intended for processors with an unlocked multiplier (CPU clock rates can be increased/decreased with it), you can also use them for a processor with a fixed multiplier. In this case a CPU should be overclocked in a usual way, by specifying the reference frequency corresponding to the maximum overclocking level, and then you reduce multipliers for off-line cores in the profiles.
It goes without saying that the number of ready profiles for applications will grow in future. As for now, the list includes only the most popular games. By the way, we shouldn't worry about timely support for games: it's very easy to find out how many cores a game is actively using. As a GPU company, AMD must have enough employees to keep tabs on such software to optimize ATI Radeon drivers. Ultimately, frequency of graphics processors is specified from the driver, so using the same method for CPUs is a good idea. There is only one question left from the tester's point of view: use profiles right now or wait until lots of applications (not only games) get their presets. We decided against using these profiles so far, because Phenom II looks great in games even without overclocking.
||Phenom II X4 925
||Phenom II X4 955
||Core 2 Quad Q9300
||Core 2 Quad Q9550
||Core i7 920
|Process technology, nm
|Core clock, GHz
|Number of cores
|L1 cache, I/D, KB
|L2 Cache, KB
||4 x 512
||2 x 3072
||2 x 6144
||4 x 256
|L3 cache, KB
(*) The maximum frequency supported by CPU memory controller. You can choose a lower frequency, if it's supported by the given memory standard (for example, DDR2-667 and DDR2-800 for processors supporting DDR2-1066). In case of LGA775 processors, memory frequency and type are detected by the chipset.
(**) Unlocked for overclocking.
(***) With Turbo Boost (enabled by default) the real core clock rate grows to 2.8-2.93 GHz depending on the current load, so would be incorrect to directly compare it with fixed clock rates of other processors.
||Memory (current mode)
||ASUS M3A79-T Deluxe (790FX)
||Corsair CM2X2048-8500C5D (dual-channel DDR2-1066, 5-5-5-15-2T, unganged)
||MSI 790FX-GD70 (790FX)
||Corsair CM3X2G1600C9DHX (dual-channel DDR3-1333, 8-8-8-24-1T, unganged)
||ASUS P5Q Deluxe (P45)
||Corsair CM2X2048-8500C5D (dual-channel DDR2-1066, 5-5-5-15-2T)
||Intel DX58SO (X58)
||Kingston KVR1333D3N9K3/6G (3-channel DDR3-1066, 8-8-8-19)
We should make a small remark about the memory size. Considering that the 3-channel controller in Core i7 and dual-channel controller in the other processors cannot use the same memory kit, we shall run the tests with the most natural memory size in each case (4GB and 6GB correspondingly). But what if we need to compare computers with different memory volumes directly: can an advantage of the 6-GB system be caused not by higher CPU performance, but by the fact that an application was using more than 4GB of memory at a given time, and the computer with less memory had to swap data? We actually have such an application in our test procedure (proved by our internal tests) -- Pro/Engineer, we decided to test the top Phenom II with 6GB of memory made up of two 2-GB and two 1-GB modules. It will make our comparison better posed.
Why not test all dual-channel configurations with so much memory? The fact is, when you install two memory modules per channel, you usually have to reduce timings and often even memory frequency. For example, our computer with the ASUS P5Q Deluxe motherboard failed to pass all tests in DDR2-1066 mode, and we had to soften up timings for the ASUS M3A79-T Deluxe, although these motherboards were tested with DDR2 memory kits for overclockers from several manufacturers. The MSI 790FX-GD70 was stable with the above mentioned four memory modules (Corsair XMS3-1600) in DDR3-1333 mode, even with 8-8-8-24 timings. But when we installed only two memory modules, it demonstrated stable operation with even better timings (7-7-7-20) with the command rate reduced to 1T. So it does not make sense to tune our testbeds for the i7, the only desktop configuration here (because it's not expedient to design three channels).
- HDD: Seagate 7200.11 (SATA-2)
- Coolers: Thermalright Ultra-120 Extreme (for i7), Zalman CNPS9700
- Graphics card: Palit GeForce GTX 275
- PSU: SeaSonic M12D 750W
Write a comment below. No registration needed!