Dual Channel, DDR and Intel Processors

Intel Dual-Channel DDR Memory Architecture White Paper

http://www.kingston.com/newtech/MKF_520DDRwhitepaper.pdf

 

Hello Christer,

"Dual" may be word to focus on. For "dual" the modules should be matched as closely as possibly. DDR and now DDR2 are sold in "twin" packs. These are two selected modules where the frequencies (I think) are matched as closely as possible. If you want the best dual channel operation of the RAM, get a "matched" pair.
You can get two modules of the same brand and model of RAM, but they will probably won't be "matched" for the best dual channel operation. If you use two different brands of modules you are probably only fooling yourself that they are running in true dual channel mode.

You cannot just throw to sticks of the same type of RAM into a machine and say it is working as true "dual channel ".

Read about twin or matched pairs running in dual channel (it's been a while since I read about it).

Matt

 

I approach this thread with reservation, however, there are some misconceptions that may benefit from a basic primer. Although this is dated and does not address HT technology (a related but seperate issue), a good read can be found here. Understand that this pre-dates most of today's advanced processors.

As to matched pairs vs two identical sticks - I guess this is argueable, but from experience ...........

DFI LanParty Socket A 3+ years ago - O/C'ed 2500+ Barton 333 run at 11x200
two identical modules Corsair XMS PC3200 2-2-2-5 - unbootable
matched pair Corsair XMS PC3200 2-2-2-5 - screamer

ASUS NF7s V 2.0 - 2-3 years ago - O/C'ed 2600+ Athlon Mobile run at 11x or 12x 200
Two identical sticks Samsung PC3200 (not matched) - runs like a finely tuned watch

I did about a half dozen of those DFI boards with both Bartons and Mobiles and after the above experience, never tried anything except matched pairs. The best memory I found for them was OCZ matched PC3200 2-3-2-6.

As to the ASUS referenced above, this was one of my favorite builds and I did at least 15 of these, using Samsung unmatched running dual channel at 11x for less expensive builds and OCZ 2-3-2-6 at 12x for more expensive builds.

Several other examples could be cited re: different boards and different configurations but this author sticks with has been found to work best for him.

Now, Christer - How come I get the feeling that this thread could become a very long one before you're finished with bringing all of us a better understanding of "Dual Channel" memory? Looking forward to the next comments from that inquiring mind, I remain

 

Not being an Intel type, I want to be careful about mis-stating this but I read it cautiously and reference you back to pages 4&5 of the link I provided. Logically, if the FSB is 400MHz and you have a single 400MHz PC3200 memory module it would stand to reason that another 400MHz module isn't going to do much to the overall system speed but from the article, "on DDR Dual Channel configuration, memory is acessed 128 bits at a time, instead of 64 bits, which is the normal rate. Therefore, the transfer rate is doubled." Now, if the transfer rate is doubled, the question becomes - will the system run any faster or are we talking about a bottleneck (400MHz FSB) that prohibits improved performance? I can't argue that it would run faster but I sure would like to benchmark the system both ways to determine what gives. If the difference is negligible, does this also mean that Intel has been a little less than honest in promoting their products?

Please don't introduce your intellectually curious friend to the concept of Rambus.

 

Christer, to me that is talking about apples and oranges. PC2100(266) or PC3200(400) is the speed of the modules. Dual channel, as mentioned in the Intel quote is the transfer rate. That is, dual channel improves the transfer rate of the data, it does not increase the speed of the RAM.

The transfer rate will not have any influence on the speed. A matched pair of PC2100 will have a better transfer rate than non matched pair, but the speed will still be lower than PC3200. PC2100 will not go as fast as PC3200 unless you overclock it (set the FSB to 400Mhz rather than what it is supposed to run at which is 266Mhz).

Stretching the imagination, I might liken it to the engine and gearbox on a car. You can have a good (fast) engine and you need a good gearbox to get the power to the wheels. You can have a great gearbox, but that will not improve the speed of the engine.

At least that's the way I see it.

Matt

 

OK, leaving out dual channel, that seems basically correct, but it does not really "choke ". The CPU just stops and waits for the RAM to store or fetch the data. The slower the RAM the slower the storing and fetching. I would say that it could be called a "bottleneck ", the storing and fetching will happen at 2100 divided by 3200 it's "natural" speed.

----------------------------------------------------------------------------
Dual channel, I think you will find it relates to sending the data into or out of the RAM, it is not the actual processing (storing or fetching). The data has to be sent through a "pipeline ". Dual channel improves transfer through that pipeline.

Matt

 

OK, OK I read through the article, not that I was planning too . I was a little amazed that my original comprehension seems to be reasonably(?) accurate.

Looking at the benchmarks, they use 256MB because it is the "sweet-spot ". I would like to see 2 X 256MB PC2700 dual channel vs 1 X 512MB PC3200 single channel...extrapolating I think single channel at a higher speed would come out in front. Remember, those are "perfectly" matched modules (I would expect). (The Sandra benchmark is a measure of the bandwidth itself so will naturally be better for dual channel.) You are looking at/considering PC2100.

For me, those benchmarks reinforce my understanding that dual channel will enhance the operation of the memory. The first section seems to suggest that the operation will be doubled.

The system you mention has a 2.0GHz CPU. The motherboard may work with a 1.8GHz or 2.8Ghz CPU (guess), the 2.8Ghz will be waiting for (starved of) data a lot more than a 1.8Ghz. It can also depend on the "balance" of the system and what you want the system to do (memory intensive programs?). How many milliseconds can you afford?

Matt

 

Christer, we are arguing along different lines. Me, I would not get too involved with the figures. All I want to point out is that there are many factors involved. Dual channel vs FSB is not a level playing field unless you know the RAM modules are perfectly matched. If you would like to research further I think you will find that the reason for matching the pairs is that the "modulation" of the two modules is synchronized. How closely they are synchronized depends on how closely they match. If you put in two different brands of, say, PC3200 you will be fooling yourself that they are running as true "dual channel" See the footnote in the article. The article is talking about a "perfect" situation, I am saying that if you use a matched pair you might be able to come close.

For a perfectly matched pair of modules, all the figures you mention should be accurate.

Matt

 

Unless, of course, the POST screen tells you you're in dual channel mode. My point simply being that I've never had a problem running in dual channel mode with RAM sticks as long as they are the same size and speed. But being a "good" computerist, I always buy matched, tested pairs.

I'm a little wary of what's been said about "doubling the transfer rate" when in dual-channel. If I run Memtest86 with RAM in single-channel mode and then run it in dual-channel mode, I'm quite certain that I'm not going to see a doubling of transfer speed as reported by Memtest86. But IIRC, there is "some" benefit in terms of transfer speed.