As some of you sure know, filling all 4 DIMM slots with 512MB or 1GB modules triggers 4-way interleaving - as opposed to 2-way interleaving with only 2 DIMM slots populated with dual rank modules. This scenario potentially opens up a window of opportunity for some gain in memory performance.
On Socket939, running 4 DIMMs mostly required Command Rate 2T in order to not only achieve some decent OCs but even to fire your system up. But that on the other hand had some impact on performance drop, which destroyed any gains 4-way interleaving gave us. It was very noticeable in benchmarks but less in games.
Now, since SocketAM2 requires Command Rate 2T anyway, there is a chance that finally on this platform 4-way interleaving can give somewhat of a boost. Today I've been messing around with 4x1GB memory config on my SocketAM2 platform, running mostly SuperPi 32M, 3DMark2001 and 3DMark03. Those are the results I got so far and SuperPi 32M comparisions are particularly interesting
2x1GB @ DDR2-966 8-4-4-4.0-2T --> SPi32M 25:01 ------ 4x1GB @ DDR2-966 8-4-4-4.0-2T --> SPi32M 24:41
2x1GB @ DDR2-966 8-4-4-4.0-2T --> 3DM01 & 03 ------- 4x1GB @ DDR2-966 8-4-4-4.0-2T --> 3DM01 & 03
At DDR2-966 I got a nice 19s boost in SuperPi 32M Also 3DMark03 got 50 points extra, not that much though. Then there is 3DMark2001 that needs to be looked at a bit differently, due to a variation in CarLow. I might actually rerun this one...
2x1GB @ DDR2-1080 8-4-4-4.0-2T --> SPi32M 26:44 ---- 4x1GB @ DDR2-1080 8-4-4-4.0-2T --> SPi32M 26:11
Whoa, 33s shaved in SPi32M I did that comparo 3 times and got same result +/- 2s. This is something very interesting to see as it appears that the higher memory clock the bigger gain from 4-way interleaving. This is only my speculation, though.
I did not have enough time to run 3D at DDR2-1080 yet but I expect some good stuff as well. I will also run WinRAR benchmark, couple games and 2 more other benches. But it looks like getting 4 modules is definetely a way to go for SPi32M benchers on SocketAM2.
Bookmarks