Faster GPUs with Nvidia 310.70?


log in

Advanced search

Message boards : Number crunching : Faster GPUs with Nvidia 310.70?

Previous · 1 · 2 · 3 · 4 · 5 · Next
Author Message
Profile Paul D Harris
Volunteer tester
Send message
Joined: 1 Dec 99
Posts: 1123
Credit: 33,595,938
RAC: 626
United States
Message 1315395 - Posted: 15 Dec 2012, 2:25:48 UTC - in response to Message 1315391.

I am confused is Jason multibeam x41zb Cuda, supposed to replace the lunatics unified installer as if being faster or better than the lunitics?

There is no installer for it, but everything Jason does WAS Lunatics.....
Since his split with them over 'artistic issues', shall we say, he has been working as an independent.
So, you have to install it by hand, there is no automatic way to do so for now.

So I should run this now since lunatics no longer has a download for their versions. This is the first I have heard of multibeam x41zb from Jason. I realize it is a hand install. So this is what crunchers are using now?
____________

msattler
Volunteer tester
Avatar
Send message
Joined: 9 Jul 00
Posts: 37287
Credit: 497,983,504
RAC: 492,488
United States
Message 1315398 - Posted: 15 Dec 2012, 2:31:53 UTC - in response to Message 1315395.
Last modified: 15 Dec 2012, 2:32:13 UTC

I am confused is Jason multibeam x41zb Cuda, supposed to replace the lunatics unified installer as if being faster or better than the lunitics?

There is no installer for it, but everything Jason does WAS Lunatics.....
Since his split with them over 'artistic issues', shall we say, he has been working as an independent.
So, you have to install it by hand, there is no automatic way to do so for now.

So I should run this now since lunatics no longer has a download for their versions. This is the first I have heard of multibeam x41zb from Jason. I realize it is a hand install. So this is what crunchers are using now?

I have been running x41z for some time, but that was not public beta.
The zb version is newer, and has been released for public beta by Jason.
I shall probably cut over to zb myself when I have the time to modify 9 rigs...probably over Christmas vacation in a couple of weeks.
____________
******************
Crunching Seti, loving all of God's kitties.

I have met a few friends in my life.
Most were cats.

Manuel Palacios
Send message
Joined: 2 Nov 99
Posts: 64
Credit: 15,629,404
RAC: 10,093
Venezuela
Message 1315608 - Posted: 15 Dec 2012, 16:41:05 UTC - in response to Message 1315398.


I shall probably cut over to zb myself when I have the time to modify 9 rigs...probably over Christmas vacation in a couple of weeks.


I thought it was going to be a tedious task, but it wasn't time consuming. you just have to make sure you change the application markers in the app_info: (e.g for my GTX460 and GTX650 machines this is what changed in my app_info)
<file_name>Lunatics_x41zb_win32_cuda42.exe</file_name>
<file_name>cudart32_42_9.dll</file_name>
<file_name>cufft32_42_9.dll</file_name>

and of course make sure you copy and paste the appropriate files from Jason's x41zb folders.

a quick find and replace all did the trick for each one
____________

msattler
Volunteer tester
Avatar
Send message
Joined: 9 Jul 00
Posts: 37287
Credit: 497,983,504
RAC: 492,488
United States
Message 1315612 - Posted: 15 Dec 2012, 16:49:08 UTC - in response to Message 1315608.
Last modified: 15 Dec 2012, 16:49:49 UTC


I shall probably cut over to zb myself when I have the time to modify 9 rigs...probably over Christmas vacation in a couple of weeks.


I thought it was going to be a tedious task, but it wasn't time consuming. you just have to make sure you change the application markers in the app_info: (e.g for my GTX460 and GTX650 machines this is what changed in my app_info)
<file_name>Lunatics_x41zb_win32_cuda42.exe</file_name>
<file_name>cudart32_42_9.dll</file_name>
<file_name>cufft32_42_9.dll</file_name>

and of course make sure you copy and paste the appropriate files from Jason's x41zb folders.

a quick find and replace all did the trick for each one

Yes, I have done it many times.
Just takes a bit when you have 9 rigs to do. I just download the apps on my daily driver and put them on a thumb drive to transfer the correct app to each one. Some are 64 bit, some are 32 bit, some are Fermi, some are pre-Fermi. So I have to get the right app on each one, and then make sure I do the edits correctly...
And then do the same with the new drivers....which also take a bit of time on each one to update.
____________
******************
Crunching Seti, loving all of God's kitties.

I have met a few friends in my life.
Most were cats.

SupeRNovA
Volunteer tester
Avatar
Send message
Joined: 25 Oct 04
Posts: 131
Credit: 9,238,042
RAC: 0
Bulgaria
Message 1315743 - Posted: 15 Dec 2012, 21:20:48 UTC
Last modified: 15 Dec 2012, 21:23:00 UTC

msattler you can try directly x41zc_Winx64_cuda50 for my Quad GTX 295 cuda 50 is faster then 42 i have installed 310.70 driver too. I can say that with SetiPerformance64 stock work unit the difference is 8sec almost in favor of cuda50 and 310.70

i have also try the older version x41zb i can say that the x41zbcuda42 version is almost faster then the cuda50 of x41zc

I haven't try to to compare z41zbcuda50 and z41zccuda50, because i do not have z41zbcuda50.

and i can say that z41zbcuda32 is slower then others too and needs i do not know why more cpu %

I hope i have help't for your host of gtx's 295 :)
____________

Jeroen
Volunteer tester
Send message
Joined: 28 Mar 01
Posts: 13
Credit: 6,971,888
RAC: 3,281
United States
Message 1315805 - Posted: 15 Dec 2012, 23:42:05 UTC - in response to Message 1314696.

I have been running driver 310.54 for sometime without any issues. I just installed 310.70 and x41zc CUDA 5.0 for my system with a single 680 clocked at ~ 1300 MHz. I am so far seeing processing time ranging from 59 - 331 seconds running 1 task at a time. GPU load is from 94-96%.
____________

Profile CLYDE
Volunteer tester
Avatar
Send message
Joined: 9 Aug 99
Posts: 759
Credit: 17,397,716
RAC: 33,150
United States
Message 1315815 - Posted: 15 Dec 2012, 23:58:41 UTC - in response to Message 1314696.
Last modified: 16 Dec 2012, 0:01:37 UTC

Has anyone noticed that the Cuda _x41g tasks are processing faster under Nvidia beta driver 310.70 or is it just me? I have even increased my tasks per GPU for both the 1 x GTX660SC @ 2Gb and 1 x GTX460SE @ 1Gb from 2 to 3. Processing went from an average of 19-22 min to 15- 20 min. Temps for the GTX660SC remain below 67c & the GTX460SE below 57c.

SAME HERE for beta 310.70 - Just running Seti on Card

Running GTX550Ti (Know it's not the best, but was affordable)

Will monitor for any problems.
____________

Profile Bill G
Avatar
Send message
Joined: 1 Jun 01
Posts: 340
Credit: 29,804,363
RAC: 65,639
United States
Message 1315834 - Posted: 16 Dec 2012, 0:44:50 UTC - in response to Message 1315392.

None of my video cards are really great but I have switched to the new drivers and every one of my computers seems to be doing more......take a look at me and my computers.

What apps are you running when you see this improvement?

Lunitics installer.....that is what I have at present. (the one no longer publicly available)
____________

zoom314
Avatar
Send message
Joined: 30 Nov 03
Posts: 44517
Credit: 35,395,714
RAC: 9,287
Message 1315876 - Posted: 16 Dec 2012, 4:43:24 UTC - in response to Message 1315743.

msattler you can try directly x41zc_Winx64_cuda50 for my Quad GTX 295 cuda 50 is faster then 42 i have installed 310.70 driver too. I can say that with SetiPerformance64 stock work unit the difference is 8sec almost in favor of cuda50 and 310.70

i have also try the older version x41zb i can say that the x41zbcuda42 version is almost faster then the cuda50 of x41zc

I haven't try to to compare z41zbcuda50 and z41zccuda50, because i do not have z41zbcuda50.

and i can say that z41zbcuda32 is slower then others too and needs i do not know why more cpu %

I hope i have help't for your host of gtx's 295 :)

I'm using x41zc, so far so good on the ol 590 and 306.97 x64 under Win7 Pro x64.
____________

Maximus13
Send message
Joined: 16 May 99
Posts: 5
Credit: 1,729,694
RAC: 0
United States
Message 1316538 - Posted: 17 Dec 2012, 20:30:38 UTC

Could someone post the manual install instructions for someone who is doing a fresh install? I just returned to S@H in the last few weeks and would like to run an optimized client.

Right now running Bionic version 7.0.28(x64).
I have a Kepler card with 310.70 drivers.

I have downloaded the x41zc files linked earlier in this thread.

Since I did not have a prior Lunitics version, I do not have the referenced files to copy and paste to.

Any help is appreciated.

Sp@ceNv@der
Avatar
Send message
Joined: 10 Jul 05
Posts: 41
Credit: 67,895,202
RAC: 107,344
Belgium
Message 1316797 - Posted: 18 Dec 2012, 12:22:17 UTC - in response to Message 1314753.

Hello Jason, I got those .7z packages from the mentioned link:

I read the info files first:

x41zc_Winx64_cuda50.7z holds the same README_x41zc.txt as within the x41zc_Winx64_cuda42.7z : It doesn't mention anything specific on CUDA50. Could you please elaborate a bit on CUDA50? What generation of cards would benefit best from the CUDA50 files? I have GTX560,GTX580 & GTX680 cards at my disposal now.

This moment 310.70 x64 Gforce drivers are installed, the rigs run rock solid using the files from x41zc_Winx64_cuda42.7z, and indeed, they give a significant speedup over the lunatics version 3.2 that was included with their last release.

What interest me most is to know in what way CUDA50 differs from CUDA42 and if current (beta)builds are capable of using the improvements yet. At least if you can tell us a little about it in rather simple words that is.

In the meantime, keep up the splendid work, kind regards
____________
To boldly crunch ...

Profile jason_gee
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 24 Nov 06
Posts: 4808
Credit: 71,576,909
RAC: 9,266
Australia
Message 1316863 - Posted: 18 Dec 2012, 17:28:25 UTC - in response to Message 1316797.
Last modified: 18 Dec 2012, 17:30:33 UTC

Hello Jason, I got those .7z packages from the mentioned link:

I read the info files first:

x41zc_Winx64_cuda50.7z holds the same README_x41zc.txt as within the x41zc_Winx64_cuda42.7z : It doesn't mention anything specific on CUDA50. Could you please elaborate a bit on CUDA50? What generation of cards would benefit best from the CUDA50 files? I have GTX560,GTX580 & GTX680 cards at my disposal now.

This moment 310.70 x64 Gforce drivers are installed, the rigs run rock solid using the files from x41zc_Winx64_cuda42.7z, and indeed, they give a significant speedup over the lunatics version 3.2 that was included with their last release.

What interest me most is to know in what way CUDA50 differs from CUDA42 and if current (beta)builds are capable of using the improvements yet. At least if you can tell us a little about it in rather simple words that is.

In the meantime, keep up the splendid work, kind regards


Hi,
Cuda 5 embeds native Kepler & Kepler 2 class code. I'm not using the added dynamic parallelism & other K2 features yet as such, but the native compiler & libraries benefit the Keplers by some small amount.

For the time being there will be some small performance benefit mostly on Kepler class ( e.g. GT 680 ) if you use the Cuda 5.0 build. That's mostly related to the build having embedded Compute capability 3 binaries, and use of Cuda 5's Kepler specific libraries & compiler... IOW natively it'd be a slightly better match for these cards. On similar lines the Cuda 4.2 build will probably remain preferable for Fermi Class cards at least for the time being.

That Cuda version dependance with GPU generation could be anywhere from splitting hairs to noticeable, and mostly ties to underlying driver, compiler and library technology changes (that are ongoing). At some point in the future I intend to replace the FFT library used, but for the short to medium term that will be the dominant source of version performance variation under Cuda.

Jason
____________
"It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change."
Charles Darwin

Profile jason_gee
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 24 Nov 06
Posts: 4808
Credit: 71,576,909
RAC: 9,266
Australia
Message 1316868 - Posted: 18 Dec 2012, 17:38:08 UTC - in response to Message 1316538.

Could someone post the manual install instructions for someone who is doing a fresh install? I just returned to S@H in the last few weeks and would like to run an optimized client.

Right now running Bionic version 7.0.28(x64).
I have a Kepler card with 310.70 drivers.

I have downloaded the x41zc files linked earlier in this thread.

Since I did not have a prior Lunitics version, I do not have the referenced files to copy and paste to.

Any help is appreciated.


Hi,
Being beta/advanced user software at this time, I'd advise waiting for an installer/release/update, especially If you are unfamiliar with the app_info.xml file configuration. I'm not sure when Lunatics will provide one though, if you feel adventurous the anonymous platform mechanism is documented at http://boinc.berkeley.edu/wiki/Anonymous_platform. You would basically need to splice the provided appinfo stub content into app_info.xml, along with entries for any other applications you'd be using (such as CPU apps & Astropulse)

Jason

____________
"It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change."
Charles Darwin

andybutt
Volunteer tester
Avatar
Send message
Joined: 18 Mar 03
Posts: 251
Credit: 99,712,213
RAC: 95,397
United Kingdom
Message 1316875 - Posted: 18 Dec 2012, 18:09:03 UTC - in response to Message 1316868.

Jason
Which version would you recommend for a mixed GPU machine. 2 x 580's and a 690?

Thanks
Andy
____________

Profile jason_gee
Volunteer developer
Volunteer tester
Avatar
Send message
Joined: 24 Nov 06
Posts: 4808
Credit: 71,576,909
RAC: 9,266
Australia
Message 1316899 - Posted: 18 Dec 2012, 22:38:54 UTC - in response to Message 1316875.
Last modified: 18 Dec 2012, 22:39:41 UTC

Jason
Which version would you recommend for a mixed GPU machine. 2 x 580's and a 690?

Thanks
Andy


On the one hand for simplicity I'd say Cuda 4.2 will do fine, but probably in splitting hairs territory with gains in one type offset by marginal reduction in the other. IOW, in total throughput & other characteristics you'd have trouble telling the difference without looking quite deep.

On the other hand you could say Cuda 5.0, & associated drivers & libraries, are likely to still mature a bit further, and seem to be quite good under system contention etc already.., so it'll just depend on personal preference & whatever else you might be using the machine for.
____________
"It is not the strongest of the species that survives, nor the most intelligent that survives. It is the one that is the most adaptable to change."
Charles Darwin

Jeroen
Volunteer tester
Send message
Joined: 28 Mar 01
Posts: 13
Credit: 6,971,888
RAC: 3,281
United States
Message 1316927 - Posted: 19 Dec 2012, 0:57:33 UTC - in response to Message 1316538.
Last modified: 19 Dec 2012, 0:59:03 UTC


I have downloaded the x41zc files linked earlier in this thread.

Since I did not have a prior Lunitics version, I do not have the referenced files to copy and paste to.

Any help is appreciated.


Here is the app_info.xml that I made for running x41zc and the Astropulse GPU application that downloads by default from the Seti project servers if this project and GPU support are enabled. Both apps are running good for me via my GPU. The file does not have any CPU apps in it currently.
____________

Maximus13
Send message
Joined: 16 May 99
Posts: 5
Credit: 1,729,694
RAC: 0
United States
Message 1316946 - Posted: 19 Dec 2012, 2:04:36 UTC

Thanks Guys!

I got it to work using the information in the link Jason provided and I was able to confirm my app_info.xml settings with the example Jeroen gave.

I do have another question, however. I understand the <count> parameter, currently I have mine set to "0.5" for two processes, but I am a bit unclear as to the <avg_ncpus> and <max_ncpus> parameters.

I am currently using the value of "0.040000" for each, but have seen other references using a value of "0.200000". I am hoping someone can shed a bit more insight into these two parameters and what changing the values will do.

Thanks!

WinterKnight
Volunteer tester
Send message
Joined: 18 May 99
Posts: 8218
Credit: 21,771,852
RAC: 13,132
United Kingdom
Message 1316966 - Posted: 19 Dec 2012, 3:12:49 UTC - in response to Message 1316946.

Thanks Guys!

I got it to work using the information in the link Jason provided and I was able to confirm my app_info.xml settings with the example Jeroen gave.

I do have another question, however. I understand the <count> parameter, currently I have mine set to "0.5" for two processes, but I am a bit unclear as to the <avg_ncpus> and <max_ncpus> parameters.

I am currently using the value of "0.040000" for each, but have seen other references using a value of "0.200000". I am hoping someone can shed a bit more insight into these two parameters and what changing the values will do.

Thanks!

I'm pretty sure, but don't quote me unless it is confirmed, that it is to reserve part of the cpu so that it can do the management on data transfers et.

But due to the powers of the GPU's and the relative lower power of the CPU cores these days. I just restrict my CPU processing by one core. e.g. on Quad in BOINC preferences I set "On mltiprocessor systems, use at most 75.00 % of the processors"

Horacio
Send message
Joined: 14 Jan 00
Posts: 536
Credit: 59,858,458
RAC: 87,023
Argentina
Message 1316977 - Posted: 19 Dec 2012, 3:49:01 UTC - in response to Message 1316966.

I do have another question, however. I understand the <count> parameter, currently I have mine set to "0.5" for two processes, but I am a bit unclear as to the <avg_ncpus> and <max_ncpus> parameters.

I'm pretty sure, but don't quote me unless it is confirmed, that it is to reserve part of the cpu so that it can do the management on data transfers et.

But due to the powers of the GPU's and the relative lower power of the CPU cores these days. I just restrict my CPU processing by one core. e.g. on Quad in BOINC preferences I set "On mltiprocessor systems, use at most 75.00 % of the processors"

Those numbers are used by the scheduler as a reference of how much CPU is needed by the GPU app but it doesnt limit the CPU usage of the GPU apps.
The scheduller adds all the CPU fractions (of the running GPU apps) togheter and the result trimmed to the least integer is the number of cores that will be "reserved" for feeding GPUs... For these calcs it uses the avg_ncpus, Im not sure for what is used the max_ncpus, so I set both to the same value, just in case...

This could be used to reserve extra cores dinamycally based on the running apps, for example I have the AP apps set to 0.7 CPU and the MB to 0.25, as I run 3 concurrent WUs, if there is 3 APs running BOINC will leave 2 extra free CPU cores, if there is 3 MB running it wont be "reserving" any extra core, and in the other cases it will reserve 1 or 2 depending on which combination of apps are running... (the APs need more CPU support due to something related with the "high CPU usage" bug in the implementation of newer versions of OpenCL).
____________

Josef W. Segur
Volunteer developer
Volunteer tester
Send message
Joined: 30 Oct 99
Posts: 4134
Credit: 1,003,215
RAC: 231
United States
Message 1316978 - Posted: 19 Dec 2012, 4:05:40 UTC - in response to Message 1316946.

...but I am a bit unclear as to the <avg_ncpus> and <max_ncpus> parameters.

I am currently using the value of "0.040000" for each, but have seen other references using a value of "0.200000". I am hoping someone can shed a bit more insight into these two parameters and what changing the values will do.

Thanks!

BOINC uses the avg_ncpus value to decide whether to free CPU cores to support GPU processing or not. If avg_ncpus multiplied by the number of GPU tasks running is 1.0 or more it will start one less CPU task, etc. Obviously the max_ncpus should be equal or greater than avg_ncpus, but it isn't really used so that would just be protection against some future version of BOINC where the logic changes.

I imagine you'll be running multiple GPU tasks on your 660 Ti, you might consider setting avg_cpus to 0.5 so whenever there are 2 or more GPU tasks running one CPU core would be freed. But if your cache ever got down to one or less GPU tasks all CPU cores would also be doing tasks. With the current limits, that's a fairly likely scenario if the project has another server problem.
Joe

Previous · 1 · 2 · 3 · 4 · 5 · Next

Message boards : Number crunching : Faster GPUs with Nvidia 310.70?

Copyright © 2014 University of California