Trial division with CUDA (mmff) -- used, but runs like new!

Prime95 · 2012-09-08, 01:33

Today I'm releasing a beta version of a GPU factoring program for the double-Mersennes (MM61, MM89, MM107, MM127) and small Fermat numbers (F26 to F157). Sources attached. Nvidia cards with 2.0 compute capability is required.

Luigi Morelli will be managing reservations and results (Thanks!). Luigi has calculated that a single GTX 580 can reproduce in 6-8 months all the work that has been done by fermatsearch.org in the last 12 years!! This is an excellent opportunity to push trial factoring limits several bits higher. (My pet project, factoring MM127, has an ~4% chance of finding a factor if we push the trial factoring limit to 2^184 -- admittedly a good deal of work).

See the included readme.txt file in the sources for more details (excerpts below)

This program uses GPU sieving so is an excellent choice for computers with decent nvidia GPUs but underpowered CPUs.

See post #288, #236 for the updated source, readme.txt and linux executable.

Prime95 · 2012-09-08, 01:40

I've reserved testing MM127 from 2^177 to 2^178. I've already completed testing from 2^169 to 2^177 which repeats much of Ernst Mayer's earlier work). BTW, I find that a GPUSievePrimes setting of 810000 gives me the best throughput.

Rcv has reserved testing MM127 from 2^178 to 2^179.

Batalov · 2012-09-08, 01:43

Runs great! Kudos!

Code:

F31 has a factor: 469316...3377 [TF:75:76:mmff 0.20mmff mfaktc_barrett89_F30_61gs]
= test passed.

Uncwilly · 2012-09-08, 01:51

I would love for this to be the end of the "MM127 is prime" speculation!

Great work GW!!

firejuggler · 2012-09-08, 01:53

win 64 build please?

Prime95 · 2012-09-08, 01:54

Until Luigi gets his reservation system fully working, these ranges would take about 10 days each on a GTX 460:

MMFactor=127,179,2.5e15
MMFactor=127,2.5e15,2.8e15
MMFactor=127,2.8e15,3.1e15
MMFactor=127,3.1e15,3.4e15
MMFactor=127,3.4e15,3.7e15
MMFactor=127,3.7e15,4.0e15
MMFactor=127,4.0e15,4.3e15

etc.

c10ck3r · 2012-09-08, 02:48

Can haz PM on how to makefile? I am a noob...:(

Prime95 · 2012-09-08, 02:56

Linux 64-bit executable: http://www.sendspace.com/file/oh7mfi

I've not built any Windows executables, perhaps someone can help with that.

flashjh · 2012-09-08, 06:24

Windows 64 executable: http://www.sendspace.com/file/zy4t27

Compiled with CUDA 4.2, sm_20 and sm_30

Let me know how it works.

@Prime95: I had to change line 243 of tf_common_gs.cu to:
#ifdefined (TF_BARRETT_76BIT_GS) || defined (TF_BARRETT_77BIT_GS) || defined (TF_BARRETT_79BIT_GS)

firejuggler · 2012-09-08, 08:11

thanks flashjh, it work

Code:

F31 has a factor: 46931635677864055013377 [TF:75:76:mmff 0.20mmff mfaktc_barrett89_F30_61gs]
found 1 factor for F31 from 2^75 to 2^76 [mmff 0.20mmff mfaktc_barrett89_F30_61gs]

Prime95 didn't you mess with

Code:

MMFactor=127,179,2.5e15
MMFactor=127,2.5e15,2.8e15
MMFactor=127,2.8e15,3.1e15
MMFactor=127,3.1e15,3.4e15
MMFactor=127,3.4e15,3.7e15
MMFactor=127,3.7e15,4.0e15
MMFactor=127,4.0e15,4.3e15

shoudn't it be
MMFactor=127,179,180
and
FermatFactor=127,2.5e15,2.8e15 etc?

rcv · 2012-09-08, 11:55

Quote:

Originally Posted by Prime95

rcv has reserved testing MM127 from 2^178 to 2^179.

I've had access to several versions over the last couple of weeks. Getting close to completion of my first assigned MM127 range:

Code:

 4488/4620 |    243.70G | 1563.9s | 11h43m | 155.83M/s |      810549 |   n.a.%

The only issue I've had was slow screen updates when running on the same card as my system video.

I've also tested George's new version (mmff) with excellent results. Even without George's hint, I found (in class #243):

Code:

F31 has a factor: 46931635677864055013377 [TF:75:76:mmff 0.20mmff mfaktc_barrett89_F30_61gs]

Don't be afraid to tweak the GPUSieveSize parameter. Too big, and you may experience very slow screen updates, if your NVIDIA card is also running your video. This parameter also controls how well mmff shares the GPU with other GPU software. Smaller GPUSieveSize is more polite to other GPU software and system video; larger GPUSieveSize is less polite to other GPU software and system video.

With my GTX 480, I am happy with GPUSieveSize=4. I suspect a significantly smaller card may need a smaller GPUSieveSize to retain good video performance.

Thanks, George!

2012-09-08, 01:33	#1
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 73×113 Posts	Trial division with CUDA (mmff) -- used, but runs like new! Today I'm releasing a beta version of a GPU factoring program for the double-Mersennes (MM61, MM89, MM107, MM127) and small Fermat numbers (F26 to F157). Sources attached. Nvidia cards with 2.0 compute capability is required. Luigi Morelli will be managing reservations and results (Thanks!). Luigi has calculated that a single GTX 580 can reproduce in 6-8 months all the work that has been done by fermatsearch.org in the last 12 years!! This is an excellent opportunity to push trial factoring limits several bits higher. (My pet project, factoring MM127, has an ~4% chance of finding a factor if we push the trial factoring limit to 2^184 -- admittedly a good deal of work). See the included readme.txt file in the sources for more details (excerpts below) This program uses GPU sieving so is an excellent choice for computers with decent nvidia GPUs but underpowered CPUs. See post #288, #236 for the updated source, readme.txt and linux executable. Last fiddled with by Prime95 on 2012-12-13 at 00:27 Reason: Link

2012-09-08, 01:43	#3
Batalov "Serge" Mar 2008 Phi(4,2^7658614+1)/2 2·13·17·23 Posts	Runs great! Kudos! Code: F31 has a factor: 469316...3377 [TF:75:76:mmff 0.20mmff mfaktc_barrett89_F30_61gs] = test passed.

2012-09-08, 01:54	#6
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 73·113 Posts	Until Luigi gets his reservation system fully working, these ranges would take about 10 days each on a GTX 460: MMFactor=127,179,2.5e15 MMFactor=127,2.5e15,2.8e15 MMFactor=127,2.8e15,3.1e15 MMFactor=127,3.1e15,3.4e15 MMFactor=127,3.4e15,3.7e15 MMFactor=127,3.7e15,4.0e15 MMFactor=127,4.0e15,4.3e15 etc. Last fiddled with by Prime95 on 2012-09-08 at 02:58

2012-09-08, 06:24	#9
flashjh "Jerry" Nov 2011 Vancouver, WA 463₁₆ Posts	Windows 64 executable: http://www.sendspace.com/file/zy4t27 Compiled with CUDA 4.2, sm_20 and sm_30 Let me know how it works. @Prime95: I had to change line 243 of tf_common_gs.cu to: #ifdefined (TF_BARRETT_76BIT_GS) \|\| defined (TF_BARRETT_77BIT_GS) \|\| defined (TF_BARRETT_79BIT_GS) Last fiddled with by flashjh on 2012-09-08 at 07:24 Reason: Figured out why sm_30 wouldn't compile

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Mersenne trial division implementation	mathPuzzles	Math	8	2017-04-21 07:21
trial division over a factor base	Peter Hackman	Factoring	7	2009-10-26 18:27
P95 trial division strategy	SPWorley	Math	8	2009-08-24 23:26
Trial division software for Mersenne	SPWorley	Factoring	7	2009-08-16 00:23
Need GMP trial-division timings	ewmayer	Factoring	7	2008-12-11 22:12

2012-09-08, 01:40	#2
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 73×113 Posts	I've reserved testing MM127 from 2^177 to 2^178. I've already completed testing from 2^169 to 2^177 which repeats much of Ernst Mayer's earlier work). BTW, I find that a GPUSievePrimes setting of 810000 gives me the best throughput. Rcv has reserved testing MM127 from 2^178 to 2^179. Last fiddled with by Prime95 on 2012-09-08 at 01:43

2012-09-08, 01:51	#4
Uncwilly 6809 > 6502 """"""""""""""""""" Aug 2003 101×103 Posts 23×479 Posts	I would love for this to be the end of the "MM127 is prime" speculation! Great work GW!!

2012-09-08, 01:53	#5
firejuggler "Vincent" Apr 2010 Over the rainbow 2×3×5×97 Posts	win 64 build please?

2012-09-08, 02:48	#7
c10ck3r Aug 2010 Kansas 547 Posts	Can haz PM on how to makefile? I am a noob...:(

2012-09-08, 02:56	#8
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 73·113 Posts	Linux 64-bit executable: http://www.sendspace.com/file/oh7mfi I've not built any Windows executables, perhaps someone can help with that.