Melraidin

Was just playing last night with my A64 3400 DTR and noticed something strange when testing with Prime 95. When I run the small FFT torture test at around 2700 MHz it takes 30+ seconds per test. But when I reset my BIOS to defaults (4x CPU multiplier, so 800 MHz CPU) the tests blaze past in under 2 seconds apiece.

Anyone have any ideas what may be causing this?

Xavier1421

Try running super pi at default, and at 2700.

I'll come up with a theory after that.

__________________
Retirement isn't all it's cracked up to be

Melraidin

Tried that last night. At 2700 MHz I got roughly 33-34 seconds for 1 M; at default (800 MHz) I got 1 minute 43 (going from memory here, may have been longer).

Xavier1421

I should have seen that coming...lol

nice times btw! (for 2700)

My guess is, the FFT iterations do not converge, so it takes longer to run the test than it does at stock speeds.

See, overclocked causes some instability, and if the instability is too great, there is no convergence in the solutions. If there is minor instability, the convergence takes longer to find, whence longer FFT times over a slower clock speed.


Raise your hands if you're confused.

__________________
Retirement isn't all it's cracked up to be

AtoMicMoNkEy15

i have the same problem! When i OC here my comp does less FFT's but doesnt error. Does that still mean that its stable? And because it takes longer to find, will that actuallly make my OC slower than it would be if i reduced it?

__________________
i play drumset and marching drums. If you do too drop me a PM.

Melraidin

But if there were no convergence then of course Prime95 would (should) find that it has an incorrect solution and crap out with an error message (Execution halted). Instead it doesn't fail at all, it just keeps going, seemingly finding correct solutions, just much slower.

The only explanation I could possibly think of is if somehow dropping to the system's default speed (800 MHz) is re-enabling some form of cache or some such that gets disabled when I OC.

ETA:

BTW, the speed difference also exists at stock speed. Ie: Running at 2200 MHz Prime95 takes over 30 seconds for a single test, while running at BIOS defaults (800 MHz) it takes only a couple seconds.

Xavier1421

the numbers of iterations could have some correlation to clock speed.

The higher the clocks, the more iterations...lots more numbers to check for convergence.

Just a theory...

__________________
Retirement isn't all it's cracked up to be

muffin

Are we talking about Prime95 or SuperPI? I'm lost already

__________________
Nikon D90 + D40x / 18-55mm DX / 18-105 DX VR / 55-200mm DX VR / 50mm AF f1.8 / 70-210mm AF f4-5.6

Melraidin

Prime95, but ran SuperPi just to ensure that my machine wasn't running faster at BIOS defaults than at OCed speeds.

Melraidin

The only way I could possibly explain this is either Prime 95 itself is cutting down on the amount of work done or something about the nature of the work is changing.

Prime 95 could be deciding that the slower CPU needs less iterations per test to properly test itself. This would seem strange since as far as I can tell the same algorithm is being used for testing the same number, so it should require the same number of iterations no matter what the speed.

The other option is the nature of the work changing. If perhaps the CPU is disabling a cache when I run it OCed (or even at stock speeds), then that could possibly explain the problem. Or perhaps something about the lower CPU speed is causing the CPU to have more cache hits or have to flush its pipelines less often... Just a shot in the dark I know, but can't think of anything else it could be.