They say that a whole stadium of people wishing the same thing at the same time can make it come true. For example, if they all start meditating with a photo of a certain tree in a forest, picturing it on fire, that very tree may really ignite. We don't know whether this psychological experiment has to do anything with reality. We just never tried it. :) But imagine how many reviewers and interested users said that it made no sense to upgrade from DDR2 to DDR3. So it must be really hard for the new memory type to justify its existence, even if there are some technical prerequisites for it already. We understand that we cannot expect any qualitative leap in performance, especially in case of similarly clocked memory modules. However, we decided to publish a thorough analysis of the situation.

Theoretically, DDR3 memory has two pros and one con (not in terms of cooling systems, of course; DDR3 needs lower voltage, so it dissipates less heat than DDR2 memory operating at the same frequency; and it offers at least no worse economic parameters in the same class, for example, if we compare DDR2-1066 and DDR3-1600 or DDR2-667 and DDR3-1066, etc). For one, this memory runs as higher clock rates than DDR2. And for two, even if we take a look at frequencies of memory modules for overclockers, DDR3 apparently takes higher frequencies by storm. That's what it was designed for. And what concerns drawbacks, it's easy to guess that it's higher latencies. To be more exact, less aggressive timings available to users. But the real latency values equally depend on frequency and timings. So having a sufficient advantage in frequency, DDR3 must in the very least match DDR2 in latency or outperform it, also offering higher memory bandwidth (determined almost solely by frequency).

That should be the reason for higher performance (if a CPU can utilize higher memory bandwidth, of course) in those applications that used to suffer from inefficient data exchange between a processor and memory. But this is theory. Let's see how it happens in practice.

Testbed:

• CPU: Phenom II X4 810
• Motherboards: Gigabyte MA790GP-DS4H, ASUS M4A78T-E (for integrated graphics tests), MSI 790FX-GD70
• Memory: 2x2GB Corsair CM2X2048-8500C5D DDR2-1066, Apacer DDR3-1333 CL9, OCZ 3G18004GK DDR3-1800, unganged mode in all cases
• Graphics card: ASUS HD4870X2 (ATI Radeon HD 4870 X2, 2GB GDDR5, 750 MHz / 900 (3600) MHz)
• HDD: Seagate ES2 SATA II 750GB
• Cooler: Zalman CNPS9700 AM2
• PSU: SeaSonic M12D 750W

Software:

• Windows Vista SP1 64 bit, Catalyst 9.2, AMD OverDrive 2.1.6, AMD Fusion for Gaming Ultility 1.0
• RMMA 3.8
• 7-Zip 4.65 x64
• WinRAR 3.80
• GTA IV: built-in benchmark, 1680x1050, Texture Quality: High, Render Quality: High, View Distance: 52, Detail Distance: 100, Vehicle Density: 100, Shadow Density: 16
• FarCry 2: bundled benchmark, 1680x1050, Ranch demo (medium-sized map), Very High settings except for Post FX and Ambient, which were High
• Lost Planet Extreme Condition: built-in benchmark, 1440x900, maximum settings, DX10, AFx16, Cave scene
• World in Conflict: built-in benchmark, 1680x1050, DX10, tests were run at Very High and High settings
• PT Boards Knights of the Sea: demo benchmark, 1680x1050, DX10, maximum settings

The typical mode for the memory controller in processors for Socket AM3 is currently DDR3-1333, but we also ran our tests at lower (1066) and higher (1600) frequencies. Processors for Socket AM3 support a frequency multiplier (to raise it without increasing the reference clock rate), which is available in BIOS. Besides, we tested operating stability in DDR3-1333 mode with two memory modules per channel (all four memory slots are filled) and compared results with the increased CPU and NB frequency.