GPU LL Testing FAQ - Page 11

LaurV · 2012-07-18, 05:16

Quote:

Originally Posted by TObject

I started a test with v1.4.2. Are the save files compatible?

Certainly not! All below 1.6 are using different checkpoint files structure.
(happened to me when switched from 1.49 to 1.61) If you are far into the test, like 5M-8M iterations for a 26M DC, then let it run, it will take less time to finish than a fresh restart with 2.03. If you are lower iteration count, then stop it, delete saved files, and restart with 2.03 (or 2.04), which will use a better FFT size and are much faster and you can rapidly recover the distance.

TObject · 2012-07-18, 18:15

I am over two-fifth into the test. I will let it finish before upgraiding. Thank you for your help.

Brain · 2012-07-18, 18:49

That's what I really tried to avoid: old PDFs hanging out there.
Two and a half options:
a) a reliable moderator regularly deletes old versions
b) gimme admin rights
The latter I don't deserve too much.
(c) stop the GPU cheat sheet)

TObject · 2012-07-18, 20:20

Here is an idea: store the PDF off-site and include a link to the PDF in the first post of this thread. Then, when a new version of the PDF is ready, put it in the same place where the old PDF was, with the same name, so that the old link always points to the latest version of the PDF.

Also, include the links posted by LaurV and Dubslow into the first post.

Dubslow · 2012-07-19, 04:32

I could potentially offer such a "live-update" (live to when I see a new version) service, though many others can as well.

As for links in the first post, I'd recommend the following three links:
http://sourceforge.net/projects/cudalucas/
http://mersenneforum.org/mfaktc
http://mersenneforum.org/mfakto (and perhaps also https://github.com/Bdot42/mfakto)

RichD · 2012-07-26, 14:22

How about a new sticky thread titled ":Links to GPU Programs" similar to what was done in the Factoring sub-forum?

Not sure if a "Links to GPU Projects" is needed since they all seem to be self-contained within this forum.

Aramis Wyler · 2013-01-27, 01:59

First I want to apologize for asking a question that might have a generally known answer. I swear I searched many threads looking for an answer.

I downloaded cudalucas today and was trying to work out an appropriate FFT length. I ran the bench and it seems like lower numbers have lower times - no surprise - and so I can pick out one of them that is the lowest compared to it's peers. What I don't know is the advantage of having a higher or lower number. Please advise if there are good fft ranges for exponents in whatever millions range.

swl551 · 2013-01-27, 02:51

Quote:

Originally Posted by Aramis Wyler

First I want to apologize for asking a question that might have a generally known answer. I swear I searched many threads looking for an answer.

I downloaded cudalucas today and was trying to work out an appropriate FFT length. I ran the bench and it seems like lower numbers have lower times - no surprise - and so I can pick out one of them that is the lowest compared to it's peers. What I don't know is the advantage of having a higher or lower number. Please advise if there are good fft ranges for exponents in whatever millions range.

If you using the recent beta version Dubslow says let it auto select the FFT size. That worked much better than trying to find the best FFT by trial and error. (I am no expert on the product though).

Dubslow · 2013-01-27, 03:52

Quote:

Originally Posted by Aramis Wyler

First I want to apologize for asking a question that might have a generally known answer. I swear I searched many threads looking for an answer.

I downloaded cudalucas today and was trying to work out an appropriate FFT length. I ran the bench and it seems like lower numbers have lower times - no surprise - and so I can pick out one of them that is the lowest compared to it's peers. What I don't know is the advantage of having a higher or lower number. Please advise if there are good fft ranges for exponents in whatever millions range.

Quote:

Originally Posted by swl551

If you using the recent beta version Dubslow says let it auto select the FFT size. That worked much better than trying to find the best FFT by trial and error. (I am no expert on the product though).

Letting the auto select pick is a good way to get into the right range.

The smaller the FFT, the faster the test goes. But, the smaller the FFT, the larger the round off error (the same type of roundoff as in Prime95). So choose the smallest FFT whose roundoff is low enough (the program should print some reasonable bounds on screen).

Edit: This also applies to Prime95, however its auto selector is built up from lots (lots) of manual testing by George to be an accurate jump table. (Among other things, the error for a given FFT in CUDALucas is much, much more varied per graphics card than the error in Prime95 for a given processor. George needs to create one table per instruction set [AVX, SSE, etc.] and the error is similar among all processors using that instruction set, whereas with CUDALucas, even a GTX 560 and GTX 580 respond quite differently to FFT lengths for the same exponent, so attempting to build a table would be rather silly. The auto selector just gets you "into the ballpark", and is "good enough", but with lots of fiddling, you might be able to get 5%, 10%, or maybe even more speed gains compared to the auto select. Most people [myself included] just don't bother though.)

Aramis Wyler · 2013-01-27, 04:04

Ah ha! That's exactly what I need, thanks.

mbn451 · 2013-04-18, 12:10

Hey are there any plans for a Mac version of the GPU computing programs? I would really like to put my graphics card to work on my new iMac, but not sure about compatibility.

Thanks!

2012-07-18, 18:49	#113
Brain Dec 2009 Peine, Germany 331₁₀ Posts	Grr That's what I really tried to avoid: old PDFs hanging out there. Two and a half options: a) a reliable moderator regularly deletes old versions b) gimme admin rights The latter I don't deserve too much. (c) stop the GPU cheat sheet)

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Anti-poverty drug testing vs "high" tax deduction testing	kladner	Soap Box	3	2016-10-14 18:43
What am I testing?	GARYP166	Information & Answers	9	2009-02-18 22:41
k=243 testing ??	gd_barnes	Riesel Prime Search	20	2007-11-08 21:13
Testing	grobie	Marin's Mersenne-aries	1	2006-05-15 12:26
Speed of P-1 testing vs. Trial Factoring testing	eepiccolo	Math	6	2006-03-28 20:53

2012-07-18, 18:15	#112
TObject Feb 2012 110010101₂ Posts	I am over two-fifth into the test. I will let it finish before upgraiding. Thank you for your help.

2012-07-18, 20:20	#114
TObject Feb 2012 3⁴·5 Posts	Here is an idea: store the PDF off-site and include a link to the PDF in the first post of this thread. Then, when a new version of the PDF is ready, put it in the same place where the old PDF was, with the same name, so that the old link always points to the latest version of the PDF. Also, include the links posted by LaurV and Dubslow into the first post.

2012-07-19, 04:32	#115
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 1110000110101₂ Posts	I could potentially offer such a "live-update" (live to when I see a new version) service, though many others can as well. As for links in the first post, I'd recommend the following three links: http://sourceforge.net/projects/cudalucas/ http://mersenneforum.org/mfaktc http://mersenneforum.org/mfakto (and perhaps also https://github.com/Bdot42/mfakto)

2012-07-26, 14:22	#116
RichD Sep 2008 Kansas 3×5×13×17 Posts	How about a new sticky thread titled ":Links to GPU Programs" similar to what was done in the Factoring sub-forum? Not sure if a "Links to GPU Projects" is needed since they all seem to be self-contained within this forum.

2013-01-27, 01:59	#117
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 3×137 Posts	First I want to apologize for asking a question that might have a generally known answer. I swear I searched many threads looking for an answer. I downloaded cudalucas today and was trying to work out an appropriate FFT length. I ran the bench and it seems like lower numbers have lower times - no surprise - and so I can pick out one of them that is the lowest compared to it's peers. What I don't know is the advantage of having a higher or lower number. Please advise if there are good fft ranges for exponents in whatever millions range.

2013-01-27, 04:04	#120
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 3·137 Posts	Ah ha! That's exactly what I need, thanks.

2013-04-18, 12:10	#121
mbn451 Mar 2013 2²×3 Posts	Hey are there any plans for a Mac version of the GPU computing programs? I would really like to put my graphics card to work on my new iMac, but not sure about compatibility. Thanks!