I think I just stumbled across something important.

[felinusz]

I posted this on OCF: http://www.ocforums.com/showthread.php?t=392628

I'm wondering what you guys think, and if anyone is willing to try this with their boards?

...



I just ran a quick test with my DFI nForce4 board, and am totally surprised by the results.

First off, I have noticed that when my memory is overclocked to the point where it starts to cause errors, it always causes errors at the same adresses.

Intrigued, I tried switching the two sticks between RAM slots. Stick one was in slot one, stick two was in slot three - I moved stick one to slot three, and stick two to slot one.

Strangely enough, the errors still occured at the same memory adresses.

But even more peculiar, I found that with the sticks set up one way in the same two slots, they produced more errors than they did when set up the other way in the same two slots.

With stick one in slot one, and stick two in slot three, I averaged ~5.2 errors per pass of memtest86 test #5. With stick one in slot three, and stick two in slot one, I got several hundred errors per pass. I immediately switched them back, and got my ~5 errors per pass again. I switched them back and forth twice to make absolutely sure - the difference in stability (as measured by the number of errors per pass) is signifigant and reproduceable 100% of the time.

What on earth is going on here? Why do my memory sticks appear to "prefer" one slot over the other, of the same pair of slots?

Could some of you please try exactly what I just outlined above, and see if your results line up with mine? You'll need an unstable memory overclock in order to count the errors, and "measure" stability.


...The exact same behaviour occurs with the Orange slots. The difference is smaller, but still present with stick one in slot two, and stick two in slot four, I got about 9 errors per pass of test #5 (I stuck to 15 loops of test #5). With stick one in slot four, and stick two in slot two, I didn't get any errors at all after ten loops of test #5! I tested the Orange slots twice as well - the difference is reproduceable and consistant. My previous testing of the slots had the Yellows giving me better stability, but with rearranged sticks the Oranges appear to actually be better.

I must stress that the same two memory slots were used.

Forgive me if it seems like I'm making a big deal about nothing, but I am astonished and amazed at this. With the same two slots, the memory behaves drastically different depending on which stick is in which slot. I can't believe I've stumbled across this. What the heck is going on with this?

If anyone can spend 15 minutes to test their board out in a similar manner I would appreciate it. This can't be just my board, not with both sets of memory slots behaving like this. For all we know, this has been holding back potential memory overclockability for any number of of people.


If anyone has an explanation or speculation, please share it. I am dumbfounded.

[ben805]

simple, if you test those sticks individually in single channel you'll find different ceiling in them, the gap between them may vary but the one that causes more error has to be the slower one. I don't know if this still apply to the newer board like these NF4 but personally I prefer to put the slower stick closer to the cpu to compensate the lost.

[perry_78]

On a DFI board one slot is stronger than the other, or so one of the guru's explained. Hence putting the weakest stick into the stonger slot will give overall better results in dual channel

[Tony]

On a DFI board one slot is stronger than the other, or so one of the guru's explained. Hence putting the weakest stick into the stonger slot will give overall better results in dual channel

[krampak]

Should we find that slot or it's always the same one?

[ben805]

Should we find that slot or it's always the same one?

the closer one to the cpu socket.

[Holst]

Might be to do with trace lengths.

There is usually one slot that is a little better than the others.

On a dual channel board this complicated by having to use two slots.

[felinusz]

O.K., that's reassuring

I was a little weired out by this behaviour, but it makes perfect sense. Thanks for the explanation guys.

[bachus_anonym]

I have zero explanation for it as I lack enough expertise, but this has been happening on Dual Channel boards for quite a while. Even my Abit NF7-S 2.0 shows same behaviour. Swapping sticks between DIMM slots 2 and 3 results in different level of stability at same vdimm.
So, I would say that it's normal as much as this can be normal at all

[Supershanks]

Polygon @ Rebels Haven Describes this issue well.

[EMC2]

Staying away from the more complicated aspects of the why...

Holst is spot on as to the major reason why... trace lengths. Here's a pic that shows your typical setup with a note about the time it takes a signal to travel from the CPU to the memory chips and back again. Lest you think the time stated isn't much, remember that 1) the control signals have to travel from the CPU to the memory and then the data has to travel from the memory back to the CPU (so twice the one way "flight time") and 2) at a 300Mhz FSB for example your only talking about a total of 1.66ns per data access in a burst



Here's a pic that shows the relative times for when things occur taking this and access times into account when the DIMM with the fastest access time is put in the slot closest to the CPU:



And here's one when we put the DIMM with the slowest access time in the slot closest to the CPU:



Remember in a Dual-Channel configuration, the data from both DIMMs of a pair must be valid before it's safe for the CPU to sample it

Peace

[JfRsQ]

I've been aware of this since the time I was playing with 8 KingstonX BH-5 sticks and my old nf7-s trying to make 2 of them work good in Dual Channel.
I finally came up with a set that were doing 240Mhz 2-2-2-5 @3.04V stable in memtest. The funniest part was that if you would mix both sticks (take #1 stick and drop it in the slot of #3, than I was ending with errors like no tomorrows in memtest #5

[mad mikee]

For the intuitive concise explanation

[AMD-me]

HOLY CRAP!! I thought my ch-5 was good before doing 345mhz 2-2-2-6 @ 3.2v

Now its doing DDR500 2-2-2-6 passing memtest! LOL thanks dude i prolly would have never tried that! Before at DDR500 i got 134 errors at test 5! Now i get 2, which has been stable thru 5 hours of BF2. I really think the two errors in test 5 is an anomoly anyway.

[OCZEric]

The OCZ guys have been suggesting this as a basic troubleshooting measure for some time now. If it doesn't work one way, try swapping the modules in the slot locations. More often than not this will solve a lot of your problems.

-Eric

[HiJon89]

An OCZ guy made a thread about this a while ago:
http://www.xtremesystems.org/forums/...409&highlight=

[ben805]

My mushkin redline can do 270 @2-2-2-6 rock solid, but out of curiosity I put the slower stick on dimm 2 (further away from cpu) and put the faster stick next to the cpu....it completely wreck havok the memtest, got thousands of error. switching them back eliminated the error and gain back the solid memclock I had.

If I were to run 4x512MB I definitely would put the slowest one next to cpu and the fastest on the last slot

[TEDY]

so

GOOD STICK in DIMM4 - closest orange to CPU
BAD STICK in DIMM2 - farthest orange to CPU

?

10X

[amrgb]

It's the opposite. It's the bad in the slot closer to the cpu.

[TEDY]

yeah just noticed thanks

just don't know why OCZ VX 3200 and 618-2 doesnt like yellow slots