(continue from last post)
The net is
1. For current nforce2 dual channel, little or no advantage until AMD fsb is twice as fast as the memory speed (in the future ?).
For Intel P4 chip set, since fsb is twice as fast as memory speed, minus some overhead in the memory controller, the effective bandwidth is about 70-80% of dual channel bandwidth, (guessing some what here, have to look up for details). It is clear 2 modules in dual channel should be used whenever possible.
2. Overclocking memory module(s) vs dual channel vs bank interleaving
IMHO,
one single bank memory module, can clock few MHz higher than
one double bank memory module, can clock few MHz higher than
two modules in single channel mode, can clock few MHz higher than
two moudles in dual channel mode
For AMD nfoce2, I found that for two memory modules, putting in as dual channel (dimm1+dimm3 for A7N8X) cannot be clocked as high as putting them in as non-dual channel (dimm1+dimm2), by a few MHz. (This does not mean the dual channel is slower in nature, actually should deliver almost twice the bandwidth. I suspect it is due to the current memory controller implementation, driver, ..., in a way that the fsb is lacking in dual channel SYNC mode. Exact reasons TBD, it is unknown to me???) ANY DIFFERENT OPINION ON THIS ??
So for AMD MB, the few MHz gain by overclocking a single memory module along that line translates into higher memory bandwidth and 3D performance and outweighs the intrinsic advantages for dual channel and multi-bank memory modules.
Dual channel, pro: the double memory bandwidth minus some overhead loss (if the fsb can use it as in P4), and also the speculative prefetch cache in the NF2 memory controller.
Bank interleave using 1 or 2 modules, pro: to reduce RAS overhead when read/write large blocks/pages of data that cross over memory banks, as in video applications
This is just a tradeoff between the performance gain in overclocking (say 5-10 MHz), and the few % performance gain from dual channel (AMD NF2) and bank interleaving. It seems to me that the MHz gain by overclocking has a slight edge over the dual channel/bank interleaving for AMD MB (exclude the NF2 version that has integrated video).
But if for non-overclockers or like manufacturers who run at a fixed frequency, then using two modules in dual channel and bank interleaving will clearly get the advantages mentioned above.
Again, for Intel P4, it is clear that using two modules win.
3. Whether running as duall channel or not, should use memory of the same size (MB), same number of banks/module, same RAS/CAS timing (those magic numbers that look like 6-3-3-2 1T) into the different dimms (basically, the same part number).
4. 256MB vs 512 MB vs 1GB
Currently, most modules are using 32Mx8 = 256 Mb DRAM chips, regardless of speed (6ns, 5 ns or 4.5 ns).
1 module with 1 bank overclocks the highest, currently it is 256 MB. But 256 MB is usually not enough to run smoothly in XP (too much paging).
256 MB = 1 bank in 1 module (which contains 8 chips on 1 side (nonECC), (9 chips for ECC module))
512 MB = 2 banks in 1 module (which contains 16 chips on 2 sides (nonECC), (18 chips for ECC module))
If you want 512 MB total, two choices:
2 modules 256 MB (1 bank/module) ---- prefer for Intel P4 for dual channel advantage
1 module 512 MB (2 bank) ---- prefer for AMD since dual channel no advantage yet !!!
If you want 1 GB total, you have to get 2 modules of 512 MB (no choice)
5. To test whether dual channel is active:
For Intel P4 MB, you will see a noticeable drop in memory bandwidth when one channel is taken out.
For nforce2 MB, you won't see a big difference w/ or w/o dual channel when running in SYNC mode.
In order to see the dual channel effect in nforce2, you can test by running ASYNC at 50% or 75% with memory speed being slower than fsb. Then you will see a noticeable drop in memory bandwidth when one channel is taken out.