Originally Posted by ssateneth
if it's not stable, it's not stable. there is no "stable enough" in my book. if you're running dry ice/LN2, that's a different story, but you aren't running dice/ln2.
P95 with AVX for absolute CPU heat torture (sometimes linpack), 8 hours minimum, 16 hours pref. linpack will help catch floating point rounding errors. if the residuals are not all identical, it's not stable, even if the program says "PASS". they must all be identical or it's not stable. otherwise if you use other cpu tester's, you will possibly run into a situation where "i can't run this video encoder program because it crashes my cpu, but everything else runs fine"
For RAM, googlestressapp for 16 hours for each setting tweaked (and also confirms if a ram setting is any faster or not), HCI for 1000% or 10 hours (whichever comes first) for final verification/screenshots. most GSAT users recommend 1 or 2 hours, but i've had a single RAM error pop up at the 13 hour mark of GSAT. 1 error = not stable.
I'm sorry but this is completely utterly absurd.
I don't want to pour 115C into my CPU just to be prime95 small FFT AVX stable at 5 ghz.
Find me an application that spams the CPU cache without addressing main memory at all and then maybe I'll care.
NOTHING crashes my 5 ghz CPU except small FFT AVX prime95 because it either crashes, hardlocks (LinX 35000 sample size 0.9.5) or reaches 115C (with enough voltage).
More than half of the stable systems here would not be stable if they used small FFT FMA3/AVX prime95 for their stress testing! Heat is unmanageable!!
Do you even know or understand how Prime95 FFT range 4k-112K works?
You do NOT need to be prime small FFT stable (AVX) to claim stability on your computer!
Even Silicon Lottery doesn't do this--they use AVX offsets.
Why? I'll explain.
1) AVX only has equal number of memory access channels as your CPU has memory lanes! Consumer boards are DUAL CHANNEL, meaning that only *TWO* AVX threads can access main memory simultaneously. That is the reason why power draw drops off heavily after the 112K FFT range (FMA3 small range has the most power draw at 112K FFT, while AVX oddly enough has the most power draw at 36K--note I did not test 100K and 96K AVX).
The other stress test programs, real world tests like x264 stress test, Blender (BMW, Classroom, others), Cinebench R20 (3600/7200 second loops), and Realbench 2.56 have less power draw because the AVX threads must access main memory. These are why these are real world tests.
2) Small FFT with AVX disabled is very important as as CPU stress test. This test MUST be passed, because if it fails, it means simply your vcore is too low, and the AVX tests mentioned above will use even more power draw than this (more power draw=more voltage droop=even more instability).
3) you're seriously limiting your overclock potential if passing small FFT/FMA3 is a criteria for your stability. ONLY limited users who do advanced number crunching for a living would need to focus on this kind of stability.
If you want to use real world prime95 to claim stability (this test is going to be harder to pass than realbench 2.56), you need to mimic real world instructions by doing the following exactly as listed:
1) Prime95 29.8 build 6, smallest FFT, AVX/FMA3 disabled (the newest primes let you disable these options directly in the GUI now), set to 2 below the maximum # of threads on your processor or set it to the max # of threads minus memory channels on your processor (9900K=14 threads, 9700K=6 threads).
2) Second instance of prime95, 29.8 build 6, FMA3 disabled, AVX enabled, # of threads=equal to the # of memory channels on your CPU, so 2 for 9900K and 9700K, FFT range 512K-8192K (click blend then custom), time to test:0 minutes.
This is a far more realistic torture test than spamming 16 AVX threads that reside purely in the CPU caches.