mersenneforum.org  

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

Reply
 
Thread Tools
Old 2015-01-04, 19:31   #67
Xyzzy
 
Xyzzy's Avatar
 
"Mike"
Aug 2002

3·2,689 Posts
Default

Wouldn't RFI, both generated and received, be an issue with an open frame?
Xyzzy is offline   Reply With Quote
Old 2015-01-04, 20:50   #68
kladner
 
kladner's Avatar
 
"Kieren"
Jul 2011
In My Own Galaxy!

2×3×1,693 Posts
Default

Faraday cage perhaps? That should not impede airflow too much.
kladner is offline   Reply With Quote
Old 2015-01-07, 21:27   #69
stars10250
 
stars10250's Avatar
 
Jul 2008
San Francisco, CA

3·67 Posts
Default

This is weird. My GTX780Ti working on a 75M exponent at 3.5 ms/iter uses 250 watts. According to the CUDALucas performance page, a GTX980 working on a 75M exponent at 7.4 ms/iter uses 165 watts. So my older card is more power efficient than the new GTX980 when running CUDALucas?
stars10250 is offline   Reply With Quote
Old 2015-01-07, 21:46   #70
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

293110 Posts
Default

Quote:
Originally Posted by stars10250 View Post
This is weird. My GTX780Ti working on a 75M exponent at 3.5 ms/iter uses 250 watts. According to the CUDALucas performance page, a GTX980 working on a 75M exponent at 7.4 ms/iter uses 165 watts. So my older card is more power efficient than the new GTX980 when running CUDALucas?
Yep!

Perhaps there are improvements to be made with CUDALucas for the new Maxwell GPUs.
Mark Rose is offline   Reply With Quote
Old 2015-01-07, 23:25   #71
Roy_Sirl
 
Roy_Sirl's Avatar
 
Sep 2002
Cornwall, UK

2×5 Posts
Default

I have noticed that with my 980, when running cudalucas the power consumption is about 71 or 72%. So that is perhaps 120W.
With trial factoring the figure is 99%, so the full 165W.
(I've just run some iterations on 75000047 and I get timings of 7.75ms - I've not tried optimising the ini file settings)

The difference in power consumption is even more striking with my 750ti with cudalucas only using 20 to 30 Watts, ie 50% or less of full power consumption. Trial factoring on the 750ti uses 90%.
(Just for comparison, 75000047 gives an iteration time of 25.5ms)
Roy_Sirl is offline   Reply With Quote
Old 2015-01-07, 23:32   #72
kladner
 
kladner's Avatar
 
"Kieren"
Jul 2011
In My Own Galaxy!

2·3·1,693 Posts
Default

My experience with a 460 and a 570 suggests that CuLu generates more heat than mfaktc. Your later model cards may be throttling on that account.
kladner is offline   Reply With Quote
Old 2015-01-08, 05:25   #73
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
Jun 2011
Thailand

939010 Posts
Default

Quote:
Originally Posted by kladner View Post
My experience with a 460 and a 570 suggests that CuLu generates more heat than mfaktc. Your later model cards may be throttling on that account.
That is normal, cudaLucas uses the memory, additional, compared with mfaktc which uses the same cores (well...), but uses no memory (see the GPU-Z, for example, for a quick view/comparison). In this respect, 580's are still in the best cards for LL (price/power/speed).

Last fiddled with by LaurV on 2015-01-08 at 05:27
LaurV is offline   Reply With Quote
Old 2015-01-09, 21:06   #74
stars10250
 
stars10250's Avatar
 
Jul 2008
San Francisco, CA

3×67 Posts
Default

Roy_Sirl sees more gpu use for TF, LaurV says LL uses more resources. My experience has been both GTX780 and 780Ti use 99% (presumably 100%) gpu on either LL or TF, so maybe I'm optimized.

I still don't understand why the GTX980 efficiency is less though. Even if we consider 120W for 7.75 ms/iter...that's equivalent to 3.72 ms/iter for 250 watts which is slightly worse than the GTX780Ti at 3.5 ms/iter. The whole point of Maxwell was power efficiency, right? It must be a code issue.

Last fiddled with by stars10250 on 2015-01-09 at 21:08
stars10250 is offline   Reply With Quote
Old 2015-01-09, 21:11   #75
VBCurtis
 
VBCurtis's Avatar
 
"Curtis"
Feb 2005
Riverside, CA

22·1,187 Posts
Default

Quote:
Originally Posted by stars10250 View Post
Roy_Sirl sees more gpu use for TF, LaurV says LL uses more resources. My experience has been both GTX780 and 780Ti use 99% (presumably 100%) gpu on either LL or TF, so maybe I'm optimized.

I still don't understand why the GTX980 efficiency is less though. Even if we consider 120W for 7.75 ms/iter...that's equivalent to 3.72 ms/iter for 250 watts which is slightly worse than the GTX780Ti at 3.5 ms/iter. The whole point of Maxwell was power efficiency, right? It must be a code issue.
No, it could be a hardware-design decision, specifically the relative efficiency of double-precision calculation to single-precision. The new generation may be vastly more power-efficient at single-precision, but no better at double. LL uses more double-precision calculations.
VBCurtis is online now   Reply With Quote
Old 2015-01-09, 21:40   #76
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

3·977 Posts
Default

Quote:
Originally Posted by VBCurtis View Post
No, it could be a hardware-design decision, specifically the relative efficiency of double-precision calculation to single-precision. The new generation may be vastly more power-efficient at single-precision, but no better at double. LL uses more double-precision calculations.
That makes sense. Maxwell has 1/32 FP64/FP32 performance versus Kepler's 1/16.

Perhaps code changes could be made to reduce CUDALucas's FP64 usage.
Mark Rose is offline   Reply With Quote
Old 2015-01-10, 02:34   #77
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
Jun 2011
Thailand

2·3·5·313 Posts
Default

Quote:
Originally Posted by stars10250 View Post
Roy_Sirl sees more gpu use for TF, LaurV says LL uses more resources. My experience has been both GTX780 and 780Ti use 99% (presumably 100%) gpu on either LL or TF, so maybe I'm optimized.
How 'bout you install GPU-Z, (or your favorite tool), switch to the right tab and have a _good_ look to the bars (ALL bars, not only the core occupancy, but also memory use, memory transfers, etc), for both cudaLucas and mfaktc, and we talk after that?
LaurV is offline   Reply With Quote
Reply

Thread Tools


All times are UTC. The time now is 15:18.

Fri Apr 23 15:18:07 UTC 2021 up 15 days, 9:58, 0 users, load averages: 2.88, 2.94, 2.57

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, 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.