[tmunn]

Have you ever heard something like "The NB is stressed the most during 512K FFTs. This happens about 2-3 hours into Blend"? Or, have you ever had Prime95 consistently fail on the same FFT size? Well, that's when the custom test comes in handy.

Pre-set Modes
We all know about the three pre-set modes available for Prime95: Small FFTs, In-Place FFTs, and Blend. Small and In-Place modes test the FPU and caches. Blend mode tests everything, including RAM.

Custom Mode
By selecting Custom, you have total control over which FFT size/s are run. In general, the smaller FFTs stress the CPU, while the larger FFTs stress the memory and chipset - with one exception, the "in place" FFT.

What is an "in-place" FFT?
An "in place" FFT is simply an FFT that is calculated entirely inside its original sample memory. In other words, calculating an "in place" FFT does not require additional buffer memory. - Source(#2.3)

Ok, so between the definition above and the fact that Prime95 labels the "In-place large FFTs" test as max heat, power, and some RAM tested. We know that when the FFTs are "in place", the RAM and chipset aren't being tested all that hard, regardless of the FFT size.

Conclusion
Below are a few screen shots of different custom tests.

CPU Test

RAM & Chipset Test

Note: when not using the "in place" option, be careful not to set the "memory to use" too high. If it's set too high, the system will start to use the paging file which will not stress the RAM.

Oh noes...it's always failing at 512 FFT size and I don't want to wait two hours for blend to reach that test

or you could do something like this

Maybe this is common knowledge to most here? If not, this should give you a general idea of what to do.

 

[TwoCables]

Whoa, very nice. I like it so much that I reported this thread saying, "I feel that this needs to be stickied" in the report.

In other words, I didn't know this stuff, and now I'm extremely glad I do.

 

[ITaoI]

Should 3gb be entered in amount of memory to use for 4gb system? Or should i just copy and paste linx amount memory use during all test

since it usually leaves like 400-500mb left and memory usage at like 90%.

 

[TwoCables]

Quote:

Originally Posted by ITaoI Should 3gb be entered in amount of memory to use for 4gb system? Or should i just copy and paste linx amount memory use during all test since it usually leaves like 400-500mb left and memory usage at like 90%.

If you're careful enough, then you can probably get it to use 3300, or maybe even 3350. This would be roughly 95-96%.

 

[radaja]

tmunn,very nice thread

 

[tmunn]

Quote:

Originally Posted by ITaoI Should 3gb be entered in amount of memory to use for 4gb system? Or should i just copy and paste linx amount memory use during all test since it usually leaves like 400-500mb left and memory usage at like 90%.

You could use linx or the performance tab in windows task manager to calculate how much memory to use. Whatever you do, be careful not to go over, or the test may be run in vain.

Quote:

Originally Posted by radaja tmunn,very nice thread

Thank you.

 

[onnetz]

Great info. I've always just used blend for chipset and memory and small ffts for cpu.
Going by what you've posted I was able to determine that my nb needed more voltage. In turn was able to lower vcore.
Forgot that I had started running my ddr2 800 at 1066.

 

[Shrimpykins]

I am totally messing around with my overclock when I get home!

Great post! +rep!

 

[saiyanzzrage]

wow, excellent, i always just hit blend and crossed my fingers..but now i can actually see what is not stable and single it out, like either ram/nb or cpu!

thanks so much!

4.1 - 4.2 here i come! (hopefully)

 

[saiyanzzrage]

wait im confused...if nb is stressed the most at 512k FFT then why did you specify 2048 - 4096k fft to test ram/chipset?

if i want to test my nb stability, should i put 512k fft or the 2048 - 4096k??

 

[tmunn]

Quote:

Originally Posted by saiyanzzrage wait im confused...if nb is stressed the most at 512k FFT then why did you specify 2048 - 4096k fft to test ram/chipset? if i want to test my nb stability, should i put 512k fft or the 2048 - 4096k??

The 512K thing is just a quote from a Phenom II guide. That quote used to be slung around the AMD forums quite a bit. I included it only because I thought most AMD overclockers would have read that line once before.

2048 - 4096k(not "in-place") is indeed the ram/chipset test.

The 512K thing is, I'm guessing, something exclusive to the Phenom II IMC. It was in a lot of guides, but never fully explained.

 

[saiyanzzrage]

Quote:

Originally Posted by tmunn The 512K thing is just a quote from a Phenom II guide. That quote used to be slung around the AMD forums quite a bit. I included it only because I thought most AMD overclockers would have read that line once before. 2048 - 4096k(not "in-place") is indeed the ram/chipset test. The 512K thing is, I'm guessing, something exclusive to the Phenom II IMC. It was in a lot of guides, but never fully explained.

any AMD oc'ers can answer/clarify this for us?

to stress cpu-nb freq stability only and not the cpu, what fft size should we use on a custom test?

tmunn, thanks for the guide!

 

[tmunn]

Personally, I'd run a round of 448 - 640k(not "in-place"), just to be in accordance with the Phenom II guides. Then do a 2048 - 4096k(not "in-place") run.

Further, any clarification from anybody on the 512K thing would be much appreciated.

 

[saiyanzzrage]

i would assume the 8k fft not in place would work for either amd or intel to test cpu stability, right?

Basically, what im trying to get out of this is to see what settings to run prime95 at so i can single out what is causing the instability in my oc, whether i need to add volts to the cpu, or play with my cpu-nb voltage and freq

so it can save me from running blend for hours and single it out that way

 

[radaja]

2hrs of 20k seems to be the best at checking cpu vcore according to leeghoofd at madshrimps

 

[saiyanzzrage]

Quote:

Originally Posted by radaja 2hrs of 20k seems to be the best at checking cpu vcore according to leeghoofd at madshrimps

cool, thanks for the info..according to tmunn's guide, they should be run as in place right?

 

[tmunn]

With 20k FFT, I don't think it matters much. However, if you want to be doubly sure the test is CPU centric as possible, then run it in-place.

 

[Nawamin]

Wow, thank you very much for this. I have been pulling my hair out for days now why my computer rebooting after hours in to the test. But now I am armed with the info you posted I can now narrow down the problem quickly.

Turned out that my VCORE was low by a few hundredth of a volt

 

[PyroTechNiK]

Very useful guide.

+1.

 

[reaper~]

Yup, very nice. Thanks OP for posting.... now I can see some light show on my AFP. lol