Pedantic question

fivemack · 2016-02-02, 16:35

Code:

(with Prime95 v28.7 on a Xeon)
	80000363	C - Unverified	2016-02-02 14:21	36.2	6BF3F11F2438E2__	249.6400
(with CUDALucas on a GTX580)
Manual testing	80000273	C - Unverified	2016-01-15 20:08	18.4	09bf0c4864c3c1__	239.6050

Why are the GHz-days figures significantly different for two adjacent exponents?

LaurV · 2016-02-02, 16:49

Different FFT size. P95 is more stingy when choosing the FFT size. For cudaLucas the faster one is chosen, even if it is bigger than the optimal one. Bigger one means (in the mind of primenet) longer time to test, therefore bigger credit.

edit: oh, actually you got it viceversa... hm... if this is at the limit, and a FFT increase is needed for the bigger expo, this would explain...

Madpoo · 2016-02-02, 21:05

Quote:

Originally Posted by fivemack

Code:

(with Prime95 v28.7 on a Xeon)
	80000363	C - Unverified	2016-02-02 14:21	36.2	6BF3F11F2438E2__	249.6400
(with CUDALucas on a GTX580)
Manual testing	80000273	C - Unverified	2016-01-15 20:08	18.4	09bf0c4864c3c1__	239.6050

Why are the GHz-days figures significantly different for two adjacent exponents?

If it matters, the CUDALucas test used a 4320K FFT size. Unfortunately the server doesn't log the FFT size used by Prime95, as far as I can tell. I've seen it included in the communication between client and server, it's just not saved anywhere I'm aware of.

2016-02-02, 16:49	#2
LaurV Romulan Interpreter Jun 2011 Thailand 25BF₁₆ Posts	Different FFT size. P95 is more stingy when choosing the FFT size. For cudaLucas the faster one is chosen, even if it is bigger than the optimal one. Bigger one means (in the mind of primenet) longer time to test, therefore bigger credit. edit: oh, actually you got it viceversa... hm... if this is at the limit, and a FFT increase is needed for the bigger expo, this would explain... Last fiddled with by LaurV on 2016-02-02 at 16:54