Prime95: Passing Small FFT's, Failing Blends (Mild OC-3.6 Ghz-Phenom II 955 BE)

[Wolfrequiem]

Dear OC,

I've been scouring both the intrawebz and OC forums and cannot seem to help myself with the following problem. I'm working with an older MB, specs and BIOS settings below. (DISCLAIMER: I did find Dolk's extremely helpful guide here, but am apparently inept at implementing it.)

System:

MB: Gigabyte GA-MA785GM-US2H
BIOS VER: F12f
CPU: Phenom II x4 955 BE
GPU: GTX 670 FTW
RAM: Generic DDR2 800 1.8v
FANS: 90 mm IN, front and rear; 120 mm top PSU out; 90 mm top out (2 in, 3 out, counting the reference blower on the GPU)
COOLER: Arctic Pro

DISCLAIMER: Heating has not been an issue, monitoring test runs with HWMonitor and CoreTemp, even with the cheaper cooler. I pull the side off during testing just to make sure that isn't a factor.

Problem: When I run Prime95 Small FFT's test, I can run it all day long, no issues (12 hours, literally). However, the MINUTE I switch to Blend, to test everything together, I get "FATAL ERROR: Rounding was 0.5, expected 0.4", no matter what minor tweaks I make (I'm beyond the BSOD phase where I'm tweaking, usually). And I'm like a dog chasing his tail trying to resolve the issue, as 3.6 Ghz with this CPU, and on this board, is MILD, especially when others are pushing 3.8 with it. I do understand my memory is an issue, being only 800 and cheap-o, but I did run Memtest and it checks out okay, and, given that I'm not pushing it, it seems like I ought to be okay there, as well. Maybe not, though.

Anyway, no matter what I do, no combination of anything seems to work me through Blends, at least, not beyond the third photo, which proved stable. Seems like I ought to be able to do a little better rather easily from all I've read though, especially if I'm not pushing temps (and they are the same temps at the stable plateau as they are at the error plateaus).

Photos 1 & 2: The simple baseline I was trying to clock it at 3.6 with. Photo 3: What has proven stable. Again, none of the various combinations seem to get me beyond the 218 at 16.5, no matter which route I go, less multiplier, more FSB, or more multiplier, less FSB. I'm thinking I'm just not hitting voltage or something correctly.







- - - Updated - - -

Correction/Clarification: I do BSOD it sometimes, I just mean, for instance, if I switch to x17 on the multiplier, even with no FSB meddling, I get error messages and workers shutting down. That's kind of where I am, minor things that don't necessarily lead to crashes, though, again, I do crash it, too, some.

 

[Johan45]

Sounds to me like a ram/nb problem.
Could we see some CPUz main, mem and spd tabs. One thing I have noticed before with DDR2 is it can be very sensitive to voltage/timing/speed. More so than DDR3. Trying to run it loose can cause problems too.

 

[Wolfrequiem]

Sure thing (and, sorry, as many threads as I've looked at, I should have thought to add it). Here it is at the baseline x18 I was trying:







The bad thing is, even keeping notes, at this point, it's hard to keep track of all the various little things I've done. Another weird thing is, that such innocuous tweaks on some point--which may go along with what you're saying--have the worst adverse effects sometimes.

 

[Johan45]

OK so you have 4 sticks of memory are they all identical? Have you tried with just 2 sticks in?
I'll be away for a bit but will have a look here in an hour or so.

 

[Wolfrequiem]

OK so you have 4 sticks of memory are they all identical? Have you tried with just 2 sticks in?
I'll be away for a bit but will have a look here in an hour or so.

I have something for you, when you get back, or for whomever. It may prove germane. (And that depends on what you mean by identical. They are all the same, but A-Tech made two and Memory Master made two. The SPD's on them are the same. I am thinking they were the same, and that maybe A-tech bought MemoryMaster out, is all. I'm not sure, though. And I have not tried your suggestion; I didn't think to.) Anyway, a perhaps salient point: Auto BIOS defaulted the memory to 667 MHz originally, and when I make JUST that change, i.e., bumping the RAM up to 800 MHz (x4.00), changing no other settings, while I don't get 3 fatal errors and stopped workers on Blend, I do get 1 on those runs. So, I'm not sure how that all plays into what you all might advise, but there's that. Here is CPU-Z for stock settings/Auto-pilot on everything, BUT with bumping the RAM up to the 800, instead of the 667 (x3.33). I left this fact out, but this is the fact that led me to believe whatever is giving my OC hiccups is totally RAM related, though, again, it all tested without errors on Memtest. I'll readd the SPD image, though this time with the other brand, as there are a few differences in what they list. I do have matching sticks in matching channels.

 

[Johan45]

The timings will list the same since they are standard timings. JEDEC is an International memory standard that all memory must conform too. My first suggestion would be to try only one set of 2 sticks in the master channel. usually slot 2 and 4 for AMD but not always but they will be the same colours. Mixing ram is not always a good idea since they may not work well together. One set reads elpida the other could be nanya it doesn't say. Try that out and see if it fixes the problem.

 

[Wolfrequiem]

The timings will list the same since they are standard timings. JEDEC is an International memory standard that all memory must conform too. My first suggestion would be to try only one set of 2 sticks in the master channel. usually slot 2 and 4 for AMD but not always but they will be the same colours. Mixing ram is not always a good idea since they may not work well together. One set reads elpida the other could be nanya it doesn't say. Try that out and see if it fixes the problem.

I've actually got a little time and am about to try this right now. So, I'll be back in a bit to see what we come up with (and, if one set doesn't work well, I'll change them out and see what the other set does, so that we have a reference for both). Thanks for the help.

 

[Johan45]

No problem. Looking at all the CPUz shots there isn't anything that stands out to me. One thing we could try, if we determine that the sets will run on their own is to add a bit of voltage to the CPU_NB and see if that helps.

 

[Wolfrequiem]

You know, it's funny how sometimes the most common sense thing is the thing you overlook the most. I've been killing myself on this, when things were pointing to the RAM, but didn't have sense enough to isolate them. On that note, first, THANK YOU. Second, though Memtest still pulls nothing up on it, it's apparently a bad stick of one pair. In fact, I've been on Amazon trying to RMA that pair this afternoon. And that's weird, because when it comes to RAM--not counting the high dollar OC stuff, etc.--I've always heard almost everyone say, "It's RAM, there are only two choices, it works or it's dead," as if there was no middle, glitchy ground in there. Apparently, there is. Thanks, again, if matching--if it means 4 completely new sticks--don't pan out, I may visit this thread again.

 

[DNRDustin]

Like Johan said, try upping your CPU_NB VID a bit too. I bumped it 1/10th a volt, It was a night and day difference on my fx6200//GA-970a-d3