Memory Question

[Bozo]

On an Intel system, which is better if you are installing 1GB of RAM; 4-256MB sticks, or 2-512MB sticks??
The only reviews/test that I found are for "Super XXX Extream memory with flashing LEDS and a 3 pound heat sink are the cat's ass in your bitched out windowed case....." "Be the envy of all your Uber-Geek friends with these 2GB pink memory sticks with dancing girls..." " get women with blah blah blah..."
Not what I was looking for.

Thanks,

Bozo :mrgrn:

 

[ddrueding]

2-512MB sticks, no question. I don't know of any benchmarks that say 4-way interleaving is any better than 2-way. Considering a price/performance ratio, 512MB sticks are on their way out these days; consider 2x1GB.

 

[LiamC]

Interleaving is not dependent upon the number of DIMM's in a system, it is dependent upon the internal configuration of the chips themselves.

http://www.realworldtech.com/page.cfm?ArticleID=RWT110401204523

 

[Bozo]

Interleaving is not the issue I was wondering about.
The Intel boards do what Intel calls Dual Channel. The chipset has the capability to read and write to two memory slots at the same time. Therefore, it doesn't fill one memory stick and then go to the next. It uses two memory sticks at the same time, theoritically doubling the speed.
There are four memory slots on the board. Two are labeled 'A' and two are 'B'. If both 'A' slots are populated, both will be used in 'Dual Channel' mode. But what if all four slots are filled? Does it fill all the memory in the 'A' slots and then go to the 'B' slots, or does it use all four at the same time. I would speculate that it fills 'A' and then goes to 'B'.
I contacted Intel support, but have not gotten a reply.

Bozo :mrgrn:

 

[ddrueding]

That is the technology I was talking about. All the AMD S939 boards use it as well. AFAIK, there is no further advantage to using 4 slots versus just the one. Furthermore, if you use AMD, the efficience of it's memory controller makes even this overkill.

 

[LiamC]

Yes.


What you are seeing is two channels with the ability to have two DIMMs per channel. This is only to facilitate mamory expansion. When only one DIMM slot is populated, the data path looks like:

NB -> Slot 1 -> NB.

With both DIMM slots populated, you can view the data path as:

NB -> Slot 1 -Slot 2 -> NB.

I can see at least two things wrong with this explanation, but in this context it will suffice. Having two slots only aids capacity.

With two channels, your data path can be either twice as wide or you can have two independent paths. PC's implement 1 x 128-bit wide data path as opposed to 2 x 64-bit. If you are operating in dual channel mode, then theoretically, your memory is twice as fast, or at least, twice as much data can be fetched at any one time. This holds up when you are doing image processing/multimedia/scientific/engineering work as these softwares work with huge data sets. This falls down when dealing with most other desktop software as the memory access are small (< 128-bits), frequent and random. In this situation, a 2 x 64 implementation would be faster.

Why don't people choose 2 x 64? How fast does Word or your browser need to run? On the other hand, when dealing with large data sets (10's, 100's, 1 000's of Megabytes), a halving a the time to return it is noticeable.

Unused bandwidth can also be used by branch and stride predictors in modern memory controllers to pre-populate caches, so excess bandwidth scores again. Database access is about the only area I can think of off the top of my head where a 2 x 64 pattern might hold advantages, but even that would depend on the contents of the DB.

 

[LiamC]

Yes, I noticed the typo–but I quite like the idea of mammary expansion
:mrgrn:

 

[Mercutio]

Only up to a point, my friend.

 

[LiamC]

Didn't Mae West say too much of a good thing is great?

 

[Bozo]

Got a reply from Intel:

"Thank you for contacting Intel(R) Technical Support.

The Intel(R) Desktop Board D915GAV has support for Dual Channel Memory. In this case, the 4 x 256MB memory sticks will provide more performance boost, however they will likely cost more than 2 x 512MB.

Also, if you use 4 x 256, you will not be able to upgrade system memory unless you replace a RAM stick.

Please do not hesitate to contact us again if you need further assistance.

Sincerely,

Jose Pablo C.
Intel(R) Technical Support

Intel(R) Desktop Boards Support Web Site"


Looks like four sticks are better than two.

Bozo :mrgrn:

 

[ddrueding]

Looks like four sticks are better than two.

In any practical application to any noticeable degree? I think not.

 

[Tannin]

An interesting response from Intel, Bozo. Where did Liam go? Now we need him to explain the theory behind it for us.

 

[LiamC]

No idea about Intel's claim, and I would be loathe to dismiss it—after all, they should know.

None of the documentation I have seen indicates how this would work in practice, though the last time I looked was early 2003. You could (in theory) do a finer grained bank interleave, but it would require intimate details of all known memory module configurations and be updated as new ones are released; and given how often module manufacturers fudge their SPD's, I'd say that it is, in practice, not feasible.

NVIDIA used a caching algorithm in their memory controller, but this wasn't dependent upon the number DIMM's populated.

It could be just marketing talking out their arse, but I'd discount that given the litiguous nature of the U.S.A.

I'll ask around.

 

[mubs]

Server class motherboards have typically had a high amount of interleaving. My memory has gone dim, but I think a HP NetServer I configured / installed / supported required 8 sticks in each of two banks just to start working. Max capacity was 32 sticks.

 

[CityK]

Memory chips are currently divided divided into 4 internal banks. These internal banks are what are referred too when talking about memory interleave. Using memory interleave has a lot to do with saving on power, localization on storing data, and umm other things that I can't remember.

Only one internal bank is active at a time. Within that bank, only one row/page is open active at a time.

Reading and writing to an open page involves the lowest of latencies for accesses to RAM. Having to close that particular row and then opening another row in the same internal bank has a greater latency penalty, but it is, however, still quicker (i.e less penelized) then having to close up that internal bank and switch to another internal bank and open up the row with the desired address.

With 4 modules you have (surprise, surprise) 2 times as many internal banks that are open in the system as you do with 2 modules in the system. It goes without saying, you have 2x as many open rows too. The chances of a page hit are increased.