mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware > GPU Computing

Reply
 
Thread Tools
Old 2015-10-23, 16:02   #1
saeres
 
Oct 2015

22·5 Posts
Default CUDALucas not fast on my slow hardware

I have CUDALucas set up on my desktop running a 100 million digit LL with FFT 19208 (auto selected). I know GPU's are faster the CPU's but my gtx 550 ti is only getting 125 ms/it. Is there a setting i need to implement to get the full benefit of the GPU?

GPU gtx 550 TI
900 mhz clock speed (increasing speed even minimally crashes CUDALucas)
runnign CUDALucas 6.5
saeres is offline   Reply With Quote
Old 2015-10-23, 16:45   #2
kladner
 
kladner's Avatar
 
"Kieren"
Jul 2011
In My Own Galaxy!

990510 Posts
Default

Quote:
Originally Posted by saeres View Post
I have CUDALucas set up on my desktop running a 100 million digit LL with FFT 19208 (auto selected). I know GPU's are faster the CPU's but my gtx 550 ti is only getting 125 ms/it. Is there a setting i need to implement to get the full benefit of the GPU?

GPU gtx 550 TI
900 mhz clock speed (increasing speed even minimally crashes CUDALucas)
runnign CUDALucas 6.5
Before addressing speed, may I ask if you have run the CUDALucas self-tests, (-r 0), (-r 1), successfully? You are embarking on a very long task. If you have not already, it would be a very good idea to run a few Double Check LL tests to make sure your hardware is giving good results.

Even then, as I am sure you will hear from others, the chances for errors are substantial on such a run. From personal experience with nVidia cards, I would strongly suggest underclocking your GPU memory, if not the chip itself. The self-tests are good guides in this regard, but successful DC tests will the real proof of basic stability.

EDIT: Oh! Welcome to the forum, too!

Last fiddled with by kladner on 2015-10-23 at 16:47
kladner is offline   Reply With Quote
Old 2015-10-23, 16:48   #3
fivemack
(loop (#_fork))
 
fivemack's Avatar
 
Feb 2006
Cambridge, England

2·52·127 Posts
Default

Quote:
Originally Posted by saeres View Post
I have CUDALucas set up on my desktop running a 100 million digit LL with FFT 19208 (auto selected). I know GPU's are faster the CPU's but my gtx 550 ti is only getting 125 ms/it. Is there a setting i need to implement to get the full benefit of the GPU?

GPU gtx 550 TI
900 mhz clock speed (increasing speed even minimally crashes CUDALucas)
runnign CUDALucas 6.5
125ms per iteration isn't actually all that bad; hundred-million-digit LL jobs are very large, and the GPU you have is not very fast as GPUs go: one core of i7/4770K gets about 12 GHz-days per day, the GTX550Ti in http://www.mersenne.ca/cudalucas.php?model=18 gets about 11.2.

On a hundred-megadigit number, one core of a 2.6GHz Ivy Bridge does about 200ms/it , so you're a bit faster than that. I think you are getting about the right speed.

Last fiddled with by fivemack on 2015-10-23 at 16:55
fivemack is offline   Reply With Quote
Old 2015-10-23, 17:58   #4
saeres
 
Oct 2015

22×5 Posts
Default

I haven't ran a self-test, haven't found anything on that. However, I have been comparing the CUDALucas reports to the Prime95 reports and as stated i'm getting 125 ms/it in CUDALucas whereas i'm getting 8 ms/it in Prime95. (both LL 100m...)

Last fiddled with by saeres on 2015-10-23 at 18:04
saeres is offline   Reply With Quote
Old 2015-10-23, 18:03   #5
VBCurtis
 
VBCurtis's Avatar
 
"Curtis"
Feb 2005
Riverside, CA

3×5×281 Posts
Default

Quote:
Originally Posted by saeres View Post
I haven't ran a self-test, haven't found anything on that. However, I have been comparing the CUDALucas reports to the Prime95 reports and as stated i'm getting 125 ms/it in CUDALucas whereas i'm getting 8 ms/it in Prime95.
You're getting 8ms/it in Prime95.... for what exponent? For what FFT size? Compare apples to apples, sir.

Or, do a test on your GPU of the same-size exponent you're doing with Prime95, compare those speeds.
VBCurtis is online now   Reply With Quote
Old 2015-10-23, 19:41   #6
saeres
 
Oct 2015

101002 Posts
Default

Sorry not at my desktop currently so can't get the specific number. However, the number I'm running I had running in prime95. That's where I'm getting my basis from. Same 100m number and prime is over 15x faster. This is based on stage 1 but I don't know if that has any influence.

I believe ftt for both is 19208 but I will validate when I get home.

Last fiddled with by saeres on 2015-10-23 at 19:42
saeres is offline   Reply With Quote
Old 2015-10-23, 19:46   #7
blip
 
blip's Avatar
 
Jan 2014

2×73 Posts
Default

Stage 1 is P-1 testing, not LL. So you cannot compare that, as they are completely different beasts.

Last fiddled with by blip on 2015-10-23 at 19:46
blip is offline   Reply With Quote
Old 2015-10-23, 20:37   #8
saeres
 
Oct 2015

22·5 Posts
Default

Okay, so the number i'm working on is M332213083. In GIMPS assignments it states that it is LL type and acquired through the selection of 100M digit. In Prime and CUDAL it uses 19208 FFT. As I am new to this and from what i'm gathering from your previous message is that initializing the work in Prime forces it into p-1 whereas CUDAL immediate starts LL testing? is this correct or am i missing something?
saeres is offline   Reply With Quote
Old 2015-10-23, 21:20   #9
blip
 
blip's Avatar
 
Jan 2014

2·73 Posts
Default

As you can see here, no P-1 has been done on that exponent yet, so I assume your worktodo.txt reads
Code:
Test=blabla,332213083,77,0
.
The "0" at the end instructs mprime to do P-1 before LL. Change to "1" to let it run LL, and then look at the timings again. Stop after a while and change back to "0" (P-1 should be done before LL).

Last fiddled with by blip on 2015-10-23 at 21:20
blip is offline   Reply With Quote
Old 2015-10-23, 21:22   #10
saeres
 
Oct 2015

22×5 Posts
Default

Awesome, Thanks for all the help!

Is there any way to start p-1 on CUDALucas or do i need to run it through prime first?

Last fiddled with by saeres on 2015-10-23 at 21:31
saeres is offline   Reply With Quote
Old 2015-10-23, 21:24   #11
blip
 
blip's Avatar
 
Jan 2014

9216 Posts
Default

You could also run P-1 on the GPU with CUDAPm1to compare its timings with mprime
blip is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
How fast/slow is your internet connection? VictordeHolland Hardware 19 2018-04-30 14:25
Don't DC/LL them with CudaLucas LaurV Data 131 2017-05-02 18:41
Do normal adults give themselves an allowance? (...to fast or not to fast - there is no question!) jasong jasong 35 2016-12-11 00:57
Trying to run CUDALucas on Windows 8 CP Rodrigo GPU Computing 12 2012-03-07 23:20
custom GIMPS hardware: how fast and how much? ixfd64 Hardware 17 2011-03-12 20:32

All times are UTC. The time now is 18:36.

Sat Jul 11 18:36:23 UTC 2020 up 108 days, 16:09, 0 users, load averages: 1.82, 1.64, 1.55

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.