Passes MemTest Handily, Fails Prime95 Miserably??

[duhvoodooman]

I'm building a PC for a friend, using some parts that were "freed up" by upgrades I did on a couple of my own systems. The core of this new machine is:

 

1. Mobo: DFI NFII Ultra Infinity Rev.A
   
2. CPU: Athlon XP 2500+ Barton (unlocked multiplier)
   
3. Memory: 2x256MB OCZ PC3200 Enhanced Latency
   

The DFI mobo was previously run problem-free for about a year at 213MHz FSB with a mobile XP 2500+ @ 2450MHz, plus 2x512MB Mushkin Black Level One PC3200 memory. The CPU and memory ran trouble-free for about the same length of time at 200MHz FSB (CPU @ 2200MHz) on an Asus A7N8X mobo. Both of these original systems were Prime95-stable for 24+ hrs. under the conditions described.

 

The problem: Since I put this combination together, I'm seeing severe Prime95 stability issues. I initially set up the system at 200MHz FSB, CPU clock multiplier of 10, stock voltages and memory timings, and a 1:1 memory ratio. The memory modules are in slots 1 & 3, dual channel mode. This setup ran 40 passes of Memtest86 v.3.2 with no errors. However, when I ran the Prime95 "Blend" test, I got failures very quickly, often in a matter of a few seconds. Some were "big number" failures and others fractional (e.g. expected 0.5, got 0.4). Lowering the CPU multiplier to 9 (i.e. CPU clockspeed = 1800MHz, below default) didn't help, so the processor doesn't appear to be the problem. So I backed off the FSB to 192MHZ, and still saw Prime95 failures within a few minutes. Next, I turned it down to 183MHz, but still can't run Prime95 for even an hour w/o failures. It's running at 166MHz as I type this, 2+ hrs. and counting.

 

What I can't understand is the disparity between the MemTest86 and Prime95 results. I could understand it if I had to back off a few MHZ to get Prime95 stability after MemTest ran well at 200MHz. But still failing at 183MHz?? The fact that CPU, mobo and memory have all been stable at 200MHz FSB in their previous systems puzzles me further, and seems to suggest some kind of incombatibility with this combination.

 

Questions:

 

• Ideas on why I see this large disparity in results between the two tests?
  
• Should I be running other stress tests to see if this is just a Prime95 test anomaly? If so, which do you recommend?
  
• Suggestions on things I should try to further isolate the problem? Swap the position of the two sticks of memory? Try different memory? Other??
  

Thanks in advance for any help!

[demonR6]

Which tests in memtest did you run? I also like to use hexus_pifast for quick and dirties to see if there are any immediate problems. My second test is SuperPi @ 32m. That normally sends up a red flag as well if there are any problems. Usually if I cannot get past those two, I won't even bother memtest or Prime because it would be wasting time for me.

[duhvoodooman]

I ran the full gamut of tests, i.e. the default sequence that loads when you boot from the MemTest floppy.

 

I'll try Super Pi at 32M and PiFast, as you suggest.

[demonR6]

With SuperPi you will get not exact in round errors anywhere within the first few seconds or even worse all the way at the end. (I hate it when I think I have a good pass and I fail on 23 or 24). Hexus you can get your results pretty quick and it will tell you if you blew it or not.

[duhvoodooman]

It appears that this instability was due to some kind of interaction between the OCZ PC3200 EL memory and the DFI mobo. While this memory had run flawlessly at CAS 2 in the Asus A7N8X board it came out of, it just isn't stable at that setting in the Ultra Infinity. I backed off to 2.5, and it's much improved. Prime 95 wouldn't run at 200MHz FSB for more than a few seconds without an error at CAS 2, but now goes about 2 - 3 hrs. at 2.5. While that's not great, I don't want to put too much stock in this one test. The machine looped 3DMark2001SE overnight without a problem, and gets through the 32M Super Pi calculation with no errors, so I think it's good to go. My friend is neither a gamer nor a videographer, and he isn't interested in discovering the next Mersenne prime number, seaching for extraterrestrial intelligence, or unlocking the secrets of protein folding phenomena. So I'm quite certain this degree of stability is more than adequate for him.

 

BTW, increasing either the core or memory voltages (or both together) doesn't help any. I think this is further evidence that the problem wasn't with a particular component itself, but rather an interaction/incompatibility between them. And the CPU clearly wasn't a factor, since the behavior was unaffected by running the CPU clock anywhere between 1800MHz (slightly below default) and 2200MHz, where it is now. Another thing that pointed to the memory as the key factor was a Prime95 test I ran with the FSB at 200MHz but with the memory ratio set to 5:6, so that it was running at only 166MHz. The test ran without errors for 8 hrs before I turned it off.