I found the ultimate CPU test

[ShawnD1]

I figured out some Prime95 settings that will make it crash even faster than Linpack crashes.
Use 4096kb sizes and run it in place. That's 4x bigger than the FFT used for Large FFT.

It didn't really hit me how much better this test was until I read another thread where I had posted what my stable voltage was for Linpack:

That's remarkably good. I ran OCCT's small Linpack test and it failed after a few minutes today, so I had to increase it to 1.475V at 3.71GHz. It's a 1055 phenom.

Doing this 4096kb test, the processor isn't stable below 1.50V. Dropping it down just one notch to 1.4875V causes it to fail.

 

[jvroig]

Interesting. How does that work? If you use 4MB FFT size, and in-place is supposed to fit in L2, the 1055T only has 3MB of L2... does that mean L3 gets used also?

EDIT: Ah nevermind, it will always spill a little into RAM anyway, since with the multiple cores there's no way everything actually fits into available cache.

 

[aphorism]

well, aside from implementation FFT in prime95 and MMM in linpack are the same thing. you can tweak the same variables in linx/linpack too.

also 4096MB is ideal because there are two matrices. this would use 8MB of cache w/o touching main mem if tiled properly.

 

[ShawnD1]

also 4096MB is ideal because there are two matrices. this would use 8MB of cache w/o touching main mem if tiled properly.

If you turn up the size by factors of 2, it starts to act more and more like Linpack. For example, Linpack usually jumps between 100% CPU and sometimes as low as 25% CPU, right? That's what it does on my system anyway. In Prime95, 4096 is always at 100%. 8192kb in-place starts to jump a bit. 32768kb in-place (maximum allowed by prime) never actually goes to 100% CPU; right now I'm doing the 32768 test and it keeps the CPU at a constant 84%. That's interesting!


GRRR!!!
32768 just failed at 1.50V. God damn it. Ok the new ultimate test is using Prime95 32768 in-place. Maybe the motherboard is failing.

 

[jvroig]

Maybe you are exposing it to scenarios that it would never encounter in real life?

Perhaps you can try that 32768 in-place FFT at stock and see if it still fails. If it doesn't fail when you are not overclocked, then indeed that is an ultimate CPU test.

 

[jvroig]

Another thing, maybe at that size of FFT, the strain is on the IMC/L3 (after all, you mentioned utilization is at 84% constant). If this hypothesis is true, then it is not vcore you need to adjust but CPU-NB voltage to stabilize the IMC.

 

[ElFenix]

If you turn up the size by factors of 2, it starts to act more and more like Linpack. For example, Linpack usually jumps between 100% CPU and sometimes as low as 25% CPU, right? That's what it does on my system anyway. In Prime95, 4096 is always at 100%. 8192kb in-place starts to jump a bit. 32768kb in-place (maximum allowed by prime) never actually goes to 100% CPU; right now I'm doing the 32768 test and it keeps the CPU at a constant 84%. That's interesting!


GRRR!!!
32768 just failed at 1.50V. God damn it. Ok the new ultimate test is using Prime95 32768 in-place. Maybe the motherboard is failing.

are you sure your ram is stable? this may not be a processor thing, entirely.

 

[SHAQ]

It looks like a cross between blend and large FFT so it could be the RAM also.

 

[richierich1212]

yup I your memory could be going bad

 

[ShawnD1]

I dont think its the ram because i always underclock the ram.

 

[DrMrLordX]

Could be the Northbridge. Have you tried running Memtest HCI lately?

 

[jvroig]

His RAM is underclocked, so if anything else, it's probably uncore/CPU-NB.