Prime95 Blend vs. Small FFT Problem

[MoeStooge]

I was trying out the torture test with a new PC I'm building and for some reason Prime95 crashes within 2-4 hours if I perform the blend test, but will run indefinitely with the small FFT test. It never gives me any type of calculation error, it just crashes. Originally, I thought that I had a memory problem but after replacing it I am still getting the same results. Memtest86 runs perfectly and all my normal applications run flawlessly. Other than memory, what would be involved in the blend test that could give me a hint where the problem is?


btw, this is a P4 2.6C system running with default settings (no overclocking). CPU temps are normal.

P4 2.6C SL6WS Malaysia w/ stock cooling
2x512MB Mushkin LII PC3500
ASUS P4C800E
Geforce MX2 (until I swap in my Radeon 9800 Pro)
WF 36GB Raptor SATA

 

[NookieN]

If the memory isn't the problem and the CPU isn't the problem, that calls the chipset into question. It's pretty unlikely that the chipset would fail if you aren't overclocking, but I have seen Prime fail due to a bad chipset before, so it is possible.

It could still be memory though. Since the CPU is dumping more heat into the case when running Prime, the memory can get a little bit hotter than from running memtest alone.

A few things to try: Does it fail if you set your memory timings as high (slow) as possible? Does it fail when your case is open? Does it still fail when your case is open and you have a house fan blowing on the motherboard (rules out any heat related problems). Does it always fail on the same FFT size? If it does, does it fail more quickly if you do a custom test using just that FFT size?

 

[Lancelot]

I can strongly advise anybody with a HT system to download the latest version 23.7.1
It's heavily P4 optimized, and just running one instance of the blend-test will completely occupy/hog your system! I was using an older version before 22.x.x and it would give almost instant errors from 251FSB on, now with this P4 optimized version I'm running 255FSB for 20mins. now while I'm typing this. So the older versions might be causing misleading results!

 

[MoeStooge]

Does it always fail on the same FFT size? If it does, does it fail more quickly if you do a custom test using just that FFT size?

It always fails on M420217 with an FFT size of 20k. If I alter the memory size on the custom test (default = 926MB) then I can get it to pass without crashing.

I know it's not a heat issue just based on my MBM temps. I do have the latest version of the software for Windows XP. Very odd.

 

[NookieN]

It always fails on M420217 with an FFT size of 20k. If I alter the memory size on the custom test (default = 926MB) then I can get it to pass without crashing.

Hmm, that's a very specific failure. A chipset flaw is still a possibility, though I'm certainly not saying that _is_ your problem. If you run only 20k FFTs with "In Place" checked does it fail? In place will basically force that test to run in the cpu's cache (a 20k FFT should fit in the cache if you're not running another application).

 

[MoeStooge]

Actually, I've found two other people with the same problem (after posting my problem on a couple of different message boards). We all have Asus P4C800E motherboards with 2x512MB of ram (other hardware is different).

I downloaded the previous version of Prime95 and did not have this problem. Of course, they added a lot of P4 enhancements in the new version. So right now I'm thinking it's either a motherboard specific problem, chipset specific problem, or software bug. I just need to have someone with the new Abit 875 board try the latest version of Prime95 to help further narrow my choices.

When running with "In Place" checked it does not crash.

 

[NookieN]

Ok I just tried v23.7 in my P4 system (w/ and Albatron 865PE Pro board). It doesn't crash with 20k 927M, but it doesn't do much of anything really. v23.8 is the same. And I see similar behavior with most FFT sizes, although the larger ones are more active.

It looks like it might be a glitch in the code. The only fix in v23.7 has to do with avoiding virtual memory problems. So if v23.6 works for you there's no reason not to use that instead. There aren't any P4 specific optimizations in 23.7.

 

[deathman20]

Hmmm I think I have the same problem with running the blend test. Maybe I'll have to try what your doing and maybe I really can get more out of my cpu then I think. Cool

 

[mccoyn]

The blend test hasn't been working for me. I have 2x512MB Ram as well. I don't get a crash though, it just doesn't do anything. Just grabs about 1 GB of memory and then sits there.

 

[oc_newbee]

I can strongly advise anybody with a HT system to download the latest version 23.7.1

Lancelot, where do you go to download the 23.7.1?
All I can find was the first 23.7v of prime95.

Thanks,
OC_Newbee

 

[Lancelot]

www.mersenne.org

 

[oc_newbee]

www.mersenne.org

Yeah, I went to the download section of that site and the only version available to download is the 23.7v of prime95. So, I don't see where you can download the 23.7.1v like you say earlier.

 

[mtnbikerjerry]

I am not sure if this violates OC policies,

but this is where you can download the recent prime.

ftp://mersenne.org/gimps/p95v237.exe

IT is recommended that you run two instances of Prime 95, even with version 23.7.

I have ran it just fine both with blended and small fft's, and custom ones just fine using two instances.

You can always unlock advanced options by inputting the password of 9876, which will allow you to change the priority.

If you do a search on prime95, there is a whole nice read on configuration settings, and what people have tried.

I will post it below.

http://forum.oc-forums.com/vb/showthread.php?s=&threadid=248225&highlight=Prime95

 

[Lancelot]

After you install v23.7, in Prime goto help and about, it says v23.7.1

 

[MoeStooge]

Actually, if you read the following post:
http://www.mersenneforum.org/showthread.php?s=&postid=15801#post15801

There is a link to an updated executable (v23.8.1) that fixes the problem. I've been using it to stress test for several days now.

 

[deathman20]

OOooooh now I have to try this out.

 

[Kidd636]

Now with the latest v23.8.1, i suggest you guys try the "Large FFTs" test (the 2nd one i think). According to one thread I've read it's the ultimate stress tester, and I think they were right, I could do hours of Prime with Small FFTs and Blend, but my system chokes out at the very first test of Large FFTs (Darn it!)