Originally Posted by ecksodia
Can anyone else weigh in on this?
WOW. Your responses have been incredibly comprehensive and I deeply appreciate it, along with everyone who has contributed to this thread. I realize this question probably gets asked many times and I apologize in advance if I continue to ask questions, as I'm not only genuinely curious about what really constitutes stability (and to what extent it matters if you're only gaming), but reaaaaally paranoid about getting ****ed up in the future by some random error.
What is your opinion of this post by user shad0wfax in an old thread I was just reading through?"You've been a very valuable source of information munaim1 and I'm glad that you've kept this thread going for us all. I'm truly grateful for all of the information in this thread from you and other users.
So please, don't take my following thoughts negatively; they are meant to be constructive criticism and encouragement for everyone here to pursue the most stability possible.
My machine won't fail at 12, 13, or even 14 hours when I am close to stability; my machine hates the 2688K FFT length and will fail there, at roughly the 15 hour mark, if I leave the time to test each FFT at the default of 15 minutes. In this case, it's not my opinion that 12 hours is not enough, it's a fact for my machine that 12 hours (at the default 15 minute time frame) is not enough. It doesn't matter if I do a default Blend or a Custom Blend to use > 90% memory; I still have the most trouble with the 2688K FFT length that comes up as test #60.
Although meeting the requirements to get the "Stable Club" badge in this thread is a default 15 minute per FFT length and 12 hour overall test, there are machines out there, like mine, that will consistently pass 12 hours with these settings and consistently fail on later FFT lengths. Thus, it is my opinion that it is far more important to test every FFT length than it is to test 48 of 60 FFT lengths.
I urge anyone who is seriously concerned about stability and wants to make only one pass to either test for 17.5 hours (or slightly more) in order to hit all 70 of the FFT lengths and qualify for the Stable Club, rather than meeting the minimum 12 hours.
If 18 hours is not an option, but two separate 12 hour tests is an option, for some reason, then I urge anyone seriously concerned about stability to make a second pass with a Custom Blend at 10 minute time to test each FFT length for 12 hours, which won't qualify for this club.
Here is a list of all of the FFT lengths Prime95 Blend will test on an i5 CPU. (Click to show)
I am in no way suggesting that the requirements for the "Stable Club" or "Super Stable Club" be changed; I simply want to let everyone know that 12 hours won't hit every possible FFT and your machine may have a weakness for a FFT length in the #49 to #70 range that you'll never find if you stop at 12 hours.
In my personal experience, 2688K is every but as much of a SB killer as 1344K and 1792K are. (Hmmm, 2688K is also 1344K x 2 ... coincidence?)"
In addition the method proposed by Boitz and corroborated by merlinx76 seems like a promising compromise - your thoughts on this as well?
Prime95 v27.7 build 2's Blend test has 82 FFTs. However, even using the older version of Prime95 that has 70 FFTs (I don't know which version that is), it will take longer than 17.5 hours to go through all 70 FFTs at 15 minutes per FFT because there is a delay in between each FFT. This means the math is not simply 70x15 equals 1,050, 1,050 divided by 60 equals 17.5.
For example, in order to find out how many FFTs Prime95 v27.7 build 2 has in its Blend test, I ran the Custom Blend test at 1 minute per FFT and waited until I saw it starting up the first FFT a 2nd time. It took exactly 105 minutes to get through all 82 FFTs instead of 82 minutes. I expected 82 minutes, but 105 minutes is what it took. Now, 82x15 equals 1,230 minutes. 1,230 minutes divided by 60 is 20.5, which means 20.5 hours if there's no delay in between each FFT. However, there happens to be a delay in between each FFT. So how long would 82 FFTs actually take at 15 minutes per FFT? I am guessing it would take approximately 24 hours and 49 minutes if the delay is always the same in between each FFT test.
Here's why I think that: 82 divided by 105 is about 0.79. This means that there's a 21% increase over my expected 82 minutes. So 20.5 hours multiplied by .21 for 21% equals 24.805 hours which means 24 hours, 48 minutes, and 18 seconds.
If it's always a 21% increase above the expected (that is, the expected as found by multiplying the number of FFTs by the number of minutes for each FFT and then dividing by 60), then 10 minutes per FFT for 82 FFTs would be 16 hours, 32 minutes, and 27 seconds. I mean, 13.67 x .21, then add the result to 13.67.
I doubt that it's always a 21% increase, but this is all I have to go by right now.
Edit: Prime95 v27.7 build 2 is the only version right now that supports the AVX instruction set. However, Windows 7 SP1 is required because SP1 provides access to the AVX instruction set. Or, it provides support - or however it should be described.
Edited by TwoCables - 8/16/12 at 6:35am