Very weird RAM problem

[James01]

I?ve been having some pretty weird problems with my RAM. I spent pretty much all of yesterday trying to diagnose what was causing it. This is what I?ve concluded:

*Each individual stick is stable
*Running any two sticks in dual channel mode is stable
*Running certain combinations of 3 sticks is unstable
*Running all 4 sticks is always unstable
*Running two sticks next to each other in single channel mode is unstable.

Type of RAM:
4x1GB A-Data Vitesta Extreme Edition
Rated at 800Mhz with 4-4-4-12 timings and 2~2.1 volts

RAM settings:
Running at 876Mhz with 4-4-4-12 timings and 2.1 volts.

And here?s something else weird that I found. Running all 4 RAM sticks at 876Mhz with my CPU at 438x8 (3.5Ghz) fails Memtest. Running all 4 RAM sticks at 888Mhz with the CPU at 266x9 (2.4Ghz) and a 3.33 memory multiplier is stable.

Now you?d think that would point to an unstable CPU, right? Then why are two sticks running in dual channel 100% rock solid stable with the CPU at 438x8? The CPU has also withstood 10 hours of the Orthos small FFT test and blend test stable.

What on earth is going on here? I've tried upping MCH, FSB, CPU, and RAM voltages. Nothing works. Someone please help!

 

[Idontcare]

On my Striker Extreme I initially loaded it with 4x1GB of Mushkin Redline DDR2-1000 stuff (way too much $, way to little common sense at the time) and I suffered VERY similiar issues as you.

They'd run fine in many different combos, but as soon as I populated all 4 Dimm slots with the ram then I'd get errors. But when put into dual-channel config only they managed just fine to pass memtest.

On a whim I just pumped the DDR2 voltage to 2.3V (with active air cooling, as we discussed in your other thread). They became memtest stable but way to high voltage for my liking.

To get all 4 sticks of ram to play nice and not cause me errors I eventually had to clock them all the way down to DDR2-800 at 2.2V with completely relaxed timings (basically SPD).

So if I had to guess what the culprit is with your ram and populating all 4 Dimms I would bet somewhere you need to massively relax one or more of your DDR2 timings in the BIOS.

 

[James01]

It would seem the motherboard definitely is the problem then. I got a similar response to yours at the OCAU forums as well.

Is there any explanation as to why this happens? I can't believe it effects a Striker Extreme, isn't that one of the most expensive boards available? I've got a Gigabyte 965P-DS3P, I'd expect things like this to happen on it

Why isn't this a more commonly known issue?

Oh well, looks like I've got three options:

1. Run 2GB RAM at 876Mhz 4-4-4-12
2. Run 4GB at 876Mhz with very loose timings (even 5-5-5-15 isn't stable!)
3. Run 4GB at 778Mhz (CPU@398x9)

Hopefully 778Mhz will be stable with 4-4-4-12 timings. I'll go for option 3 I reckon.

Thanks for all your help

 

[n7]

I believe that like on my P5B Deluxe, your mobo's NBCC (northbridge core clock) increases when using lower multipliers, which means more stress on the NB, not to mention the extra stress 4 dimms are already putting on the NB!

You can use Sisoft Sandra to check this on your mobo perhaps.
On my mobo, the lower multi [than default] you use, the higher the NBCC runs at: Pic
That's 400 MHz when using 10x350, 432 Mhz when 9x389, & even higher when 8x*** etc...
Those arrows point to the mobo's NBCC, not FSB in case it's confusing. For reference, my FSB is actually 389 MHz.

Basically, try using your default 9x multi instead, & yes, you might have to loosen timings a bit, &/or increase vdimm/vNB.
P965 is weak with 4 dimms...pretty much all P965 mobos are, so you have to tweak things in most cases.

 

[James01]

Yes! You're a genius! With 9x390, Sandra reports the NB speed as 390Mhz. With 8x438, it reports it as 493Mhz. That's exactly what's happening. The NB just can't handle 4 RAM sticks at that high speed. Finally, my problem is solved!

I'm pretty sure the 4 sticks are now stable at 780Mhz 4-4-4-12. Using a memory multiplier, I might even be able to get them running at higher speeds anyway. I'll be running Memtest and Orthos overnight to make sure they're 100% stable at this speed first.

Thanks everyone

 

[n7]

Heh, thanx.

I'm not the genius though, these guys are: http://www.thetechrepository.com/showthread.php?t=30

I found on my mobo that lower multi (& the resulting higher NBCC) weren't really an issue with my E6400, E6300, but started to notice it with my E6600, & when i started running 4 dimms, it became ridiculously apparent.

Under HCI Memtest load, my temps on the NB get insanely hot @ 9x389 vs. 10x350 simply due to the higher NBCC speeds (even tho vNB is set the same).

I believe the NBCC are tied to the straps as well, which is why P35 is so nice.
On most P35s, you can adjust what strap is being set, which removes the "issues" we P965 users have.

 

[BonzaiDuck]

My two cents: you've acquired a quantum increase in understanding these things.

I've got the Striker board. I was able to get to about 390 Mhz x 2 = 780 DDR with a 1:1 ratio and an E6600 CPU. Replacing it with the Q6600 quad meant an entirely different ball-game.

I've also determined that my system needs less RAM voltage to run at DDR2-800 with a 5:6 divider, than what's required for 1:1 at DDR2-667 and tighter timings.

But I'm too prone to digress here.

While your musings about the NorthBridge stress may be correct, I myself would not count on the RAMs running at 876 Mhz with the stock timings.

The rating of memory has always been a manufacturer's choice, but they're not out-of-bounds in "truth in advertising" if the spec represents an upper limit, given timings and voltage.

What I try to look for is a "holy Grail" module tested by some reviewer in his lab, showing that you can run the bus-speed up beyond spec at the same timings while keeping the voltage below the maximum spec. It's hard to find something like that.

With four modules, and without a-priori knowledge of your experience, I wouldn't bet much on getting beyond 800 Mhz at the stock timings -- if it were a matter of Las Vegas odds-making.

On the matter of the NBCC -- that's very interesting, and I never thought of it. It makes explicit what we only sensed intuitively about the drawbacks of using lower CPU multipliers.

So now that I think of it, the OP's observations are similar to those I also had today. Beyond some point, you're actually better off running the RAMs on a divider that is "NOT 1:1", taking a hit on the bandwidth. And under those circumstances, you might be able to tighten the latencies at the rated speed, and even run them at a higher speed.

 

[James01]

Well, at least two of the sticks run 100% stable at 876Mhz with the stock timings. They ran both Orthos and Memtest for 10 hours without errors. I haven't put the other two sticks through such extensive testing, but they ran Memtest for 3 hours OK.

One last question, could using a 9x multiplier (up from 8x) increase CPU heat? The CPU is running 10Mhz faster with the same voltage, but temps are at least 5C hotter than they were last time I tested. Or is this more likely to be caused by a higher ambient temp?