Comparison of AS5 vs. MX-4 on Haswell i7-4790K

[graysky]

The FFT thing really doesn't make much of a difference so long as the test is conducted the same way/length of time each time as it spins through the FFT lengths just the same.

That's my point: selecting from sizes that are different and aborting after an arbitarty time is not good. Now I am repeating with a fixed length FFT and fixed duration.

 

[EarthDog]

Unless I misunderstand how it works, if you run P95 for 60 minutes, you should be in a strikingly similar spot in FFT's as they run a specific length and move on to the next. As always there are slight core speed differences so they will switch lenghts at different times, but in the end they should all be in a VERY similar place at the end of a fixed amount of time. THe longer you run it, the bigger the difference is. Also remember, you loop the lengths at some point anyway...

You don't have to spin your wheels on one length at one time for this testing to be valid. Eliminate the mounting variable and you have what you are looking for so long as your P95 test is the same length of time each run.

It does if it's not constant. The size of the FFT could impact thermal output because more or less circuits inside the CPU might be required to do each computation. It's best to keep that constant so it can be ruled out as a source of error. Anything done to test between coolers or TIMs really need to be done the same way for everything. It's differences in the process that enable abnormalities in numbers to appear.

So from a purely empirical standpoint, everything that can be made constant, should be. It will make the results more accurate across the board and as you suggested, multiple mounts also suggests multiple tests to validate those numbers. It may be tedious, but that's science™!

See response above as that should be applicable there too.

 

[graysky]

As I understand it from looking at the mprime logs, it varies the size of the FFTs from 128k-1024k but does so in a rolling fashion. In other words, not a complete sampling of all sizes. If I break it 45 min through, and it's only 60 % through it's cycle (I made that number up for the purpose of my point), then when it starts up again, it resumes on the FFT size thereby making the sets asymmetrical. Setting a fixed size just minimizes everything.

 

[EarthDog]

Just restart the program? Starts from scratch (figured that was happening anyway for this reason)!

I'd think testing through the entire range would be more beneficial than a single fft length.

 

[Aquinus]

I'd think testing through the entire range would be more beneficial than a single fft length.

I personally would be more interested in keeping thermal output consistent. By changing FFT you really don't know how much it changes thermal output which is a completely different experiment IMHO in which you want to keep the cooler and TIMs constant to gauge thermal output. At least if you keep FFT the same, you can treat that as a constant between all of your tests. I would rather see more runs at the same FFT as opposed to less spread out over several.

 

[EarthDog]

At that point get a hot plate and do it by wattage.

I just think running it like every single other person in the world does would be beneficial. A lot of sustained workloads change what they are doing right? I mean even f@h and such does. And p95 runs the gamut.

On another note, I would imagine he'd have to find the fft that stresses things out the most too, no? (Sounds like a PITA!)

 

[Aquinus]

On another note, I would imagine he'd have to find the fft that stresses things out the most too, no? (Sounds like a PITA!)

He just needs to find one and stick with it. Higher differences in temperature will reflect results better but keeping heat output constant would be important in my book. Issues with things like FAH and WCG/BOINC is the that work units can vary in load depending on the project or what's being calculated. That adds variation but you know, I would go so far to agree with you that a hot plate isn't a bad idea if you really care about figuring out which does its job best, the thing is that a hot plate isn't going to get mounted the same way as a CPU/motherboard would. So trying to keep everything the same as a real platform rules that out, at least in my book. Very nifty idea though; I like the idea of a hot plate as simulated constant load.

 

[graysky]

Correct, consistency is key. I have two sources of stress I'm evaluating to answer the question of breakin time.

1) cpuburnP6
2) mprime running a 440k FFT

I will post when I have results. I plan to go out the full 200 h recommended by the manufacture but if the data say not to, I will stop early. I have the box up 24/7 running full loads for 15 min of every hour to allow for thermal cycles. I am pulling data points every 8 h using the two stresses mentioned above.