Depending on which App you intend to run, you would probably want to install the latest ToolKit from nVidia version 7.5 here, https://developer.nvidia.com/cuda-downloads?sid=480386. It's only available for Ubuntu 14.04 & 15.04 and installs driver 352. If you plan on running the current stock MBv7 Linux app you will probably need driver 350 or above. It might work OK with lower drivers but testing at Beta showed the best results with drivers 350 and above. If you plan on running the CUDA App and the AP App, the lower drivers would be fine. I'm running the CUDA & AP Apps with driver 346.59 and it's working great, http://setiathome.berkeley.edu/results.php?hostid=7258715&offset=100&appid=20. I've installed driver 346.59 from nVidia a couple of times and both times it gave me OpenCL without any fuss, I did have to make a link to get CUDA though. The next time I run out of APs I plan on testing the OpenCL MBv7 App with driver 346.59, but I keep getting sent APs ;-)

Seems it's not picking up my Nvidia 760GTX - using the latest 355.11 drivers..

Baiteh,

You might want to consider taking a look at sgfxi. I use the sgfxi script for installing my Nvidia drivers. It automates the entire installation and upgrade process. I highly recommend it.

http://smxi.org/docs/sgfxi-manual.htm

X needs to be stopped prior to installation.

To install, I run (as root):
cd /usr/local/bin && wget -Nc smxi.org/sgfxi && chmod +x sgfxi && sgfxi

For upgrades, I simply run:
sgfxi

It has worked flawlessly for me on several machines.

just install the cuda toolkit with the driver leave the sample crap out of the install and done everything gets created for you and actually survives a kernel update......also creates all the links

been working smooth for me install took 3 minutes

---

I came down with a bad case of i don't give a crap

I just had an interesting experience. I wanted to switch from CUDA to the current stock MBv7_7.08 app but didn't want to lose any tasks on Main. So I tried running a task at Beta to have the app downloaded so I could run it under Anonymous platform. Apparently there is still a driver restriction on Beta. Strange, the same app on Main apparently doesn't have a driver restriction. Anyway, I tried to update Ubuntu 15.10 to the newest version of 352.55 and it trashed the system. great. I went back to the Ubuntu 14.04.3 partition and tried it there. No problems this time, and I didn't have to do a thing to have BOINC see CUDA and OpenCL. Nice.

So, if you are still having problems try the newest release from nVidia, http://www.nvidia.com/download/driverResults.aspx/92826/en-us
I didn't do any purge or anything else, just installed it over 346.59 and everything works.

Weirdly boinc sees my card - now a 970GTX :D - and even requests work. It just isn't getting any units.

Weirdly boinc sees my card - now a 970GTX :D - and even requests work. It just isn't getting any units.

The host still isn't being seen as having OpenCL. Same as back here;
http://setiathome.berkeley.edu/forum_thread.php?id=78288&postid=1736723

It appears BOINC still doesn't see OpenCL. You need OpenCL to receive the current stock SETI Apps.

This is Your host; NVIDIA GeForce GTX 970 (4095MB) driver: 355.11
One with OpenCL would say; NVIDIA GeForce GTX 750 Ti (2047MB) OpenCL: 1.1
If you check your Event Log it is probably just listing CUDA.
There aren't any Stock Linux CUDA Apps, so, your requests will go unfilled.
If you want to run CUDA under Linux you will have to install the App yourself and run under Anonymous platform.
This CUDA App should work with your system, http://www.arkayn.us/forum/index.php?action=tpmod;dl=item132
You will not receive/run any OpenCL work until BOINC sees OpenCL on your system.

What I wonder is what is the consequence of
installing the nVidia cuda toolkit on gpu tasks
(seti or others) that are currently running (or stopped)?

Do they just get errored off (ok I suppose) get odd error?
Unpredictable?

For the toolkit itself, the installation should have no effect on running tasks (either Cuda or OpenCL) though the driver component could be problematic in some situations (for starters it should be building and installing a new kernel module, which I'd imagine resets the device(s) )

For Cuda multibeam, Since the error handling in odd situations *should* mostly go to a temporary exit, that should respond in some moderately sane way (though some incorrect error exit codes are being reviewed in preparation for the next release builds).

I'd appreciate any feedback on odd behaviour or weird reactions by the Linux Cuda multibeam builds floating around, since they are due for update fairly soon (work commitments notwithstanding)

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

I'd appreciate any feedback on odd behaviour or weird reactions by the Linux Cuda multibeam builds floating around, since they are due for update fairly soon (work commitments notwithstanding)

Well some might think it odd, but I have been using "setiathome_x41zc_x86_64-pc-linux-gnu_cuda60" for months and it has never even given me one problem that I can remember. It just keeps crunching along. I would have to think long and hard before upgrading to a newer version without some compelling reason to do so.


I wish I could say the same for the AP app but to be fair, I cannot say if it is the app, Nividia's driver, or the parameters I use.

Thanks for the feedback. Yeah in general x41zc has stuck around so long, in part because it's been pretty solid (despite several known issues) . Funnily I had been comparing it to Volkswagen for some time, until of course the Berlin police raided that place, lol, so looking for something else to compare it to now. [Edit:Saturn V maybe ? heavy solid and old.. feels right :-O]

Other than that, we're at a pretty big juncture where new technologies, contributors, and development methodologies are being mashed together. It's good to know some work can stand solid while the chaos sorts itself out

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Hmm. https://developer.nvidia.com/cuda-downloads?sid=480386
I downloaded the runfile version ( the .deb could
not be downloaded).
Thanks for the link, TBar

The nvidia download page offered the driver
and optionally the "GPU Deployment Kit"
as a separate download. What is that?
The NVIDIA CUDA INSTALLATION GUIDE
FOR LINUX (capitalization is theirs) does not mention
the Deployment Kit so I'm ignoring the Deployment Kit.

Not sure I want to try installation today.

The nvidia runfile for 7.5 refuses to run because it notices
nvidia_uvm is a loaded module, and quits.
This on Ubuntu 14.04.

I updated /etc/default/grub to boot in text mode
and then 'sudo update-grub'. So I get a text window
and no graphics. But no joy anyway.

I don't know how to get past this issue.
google queries have been no help. So far.

Take a look at my suggestion above. I'm running Mint 17.2, which I believe is based on Ubuntu 14.04. Installation took 5 minutes. Both CUDA and OpenCL work fine. No other downloads, no linking files. It has worked perfectly on all 5 of my Linux machines. Might be worth a shot.

The nvidia runfile for 7.5 refuses to run because it notices
nvidia_uvm is a loaded module, and quits.
This on Ubuntu 14.04.

I updated /etc/default/grub to boot in text mode
and then 'sudo update-grub'. So I get a text window
and no graphics. But no joy anyway.

I don't know how to get past this issue.
google queries have been no help. So far.

You could try Fawkesguy's advice, or, see if you can remove the offending package using the package manager. It sounds as though you didn't purge nvidia before installing the Toolkit and there is an old package still installed. Open the package manager, enter nvidia into the filter box, and look for an installed nvidia_uvm package. Hopefully you can find one that isn't needed by the nVidia 352.xx driver and just uninstall it.

I've been looking into the older drivers, around 346.xx, and found they give better AP times than the newer drivers. I have My AP times down to around 30 minutes on My 750Ti for the unblanked ones. Only problem is you get many inconclusive results if you use that driver with the Stock OpenCL MB App. So, it's either the newer driver and the Stock OpenCL MB App, or the older driver and the CUDA App. It would be Nice if the mbcuda.cfg settings worked in Linux...and OSX. That way I could raise the GPU load on the CUDA App, right now it runs around 86-89% on my 750Ti, which is just at the point where running another instance may help. But, I'd rather just run one CUDA instance...

It would be Nice if the mbcuda.cfg settings worked in Linux...and OSX.

Agreed. It would be very useful. :-)

I understand petri has his CUDA App adjusted to run near full load in Linux, so it must be possible ;-)

BTW David, is there something wrong with your other machine with the 2 760s? I don't see it running any GPU tasks.

Fawkesguy's advice seems to have worked.
Thanks!

Nothing untoward seems to have happened.
Running 355.11 now.

I'll wait a day or two to see what
happens and if nothing notable I will do it
on my other other
machine with modern nVidia.

I'm glad it appears to have worked. Fingers crossed. :-)

Fawkesguy's advice seems to have worked.
Thanks!

Nothing untoward seems to have happened.
Running 355.11 now.

I'll wait a day or two to see what
happens and if nothing notable I will do it
on my other other
machine with modern nVidia.

About the machine with 760's:
There has to be something wrong.
The latest request for Seti tasks says 0 tasks
available. Not even getting CPU tasks.
Puzzling.

einstein is working though.

Tomorrow I'll update that one per Fawkesguy.

It appears it has never received any GPU work, and it says it has OpenCL. Strange. Since you already have an app_info.xml you could just add the 3 CUDA files from CA, the app section to the app_info.xml, and see if it works. I have the mbcuda.cfg in my app_info (it doesn't work), but removed it would be;

   <app>
      <name>setiathome_v7</name>
   </app>
 	<file_info>
 	   <name>setiathome_x41zc_x86_64-pc-linux-gnu_cuda60</name>
	   <executable/>
 	</file_info>
	<file_info>
	   <name>libcudart.so.6.0</name>
	   <executable/>
	</file_info>
	<file_info>
	   <name>libcufft.so.6.0</name>
	   <executable/>
	</file_info>
       <app_version>
	  <app_name>setiathome_v7</app_name>
	    <version_num>704</version_num>
	    <plan_class>cuda60</plan_class>
	    <avg_ncpus>0.1</avg_ncpus>
	    <max_ncpus>0.1</max_ncpus>
	  <coproc>
	    <type>CUDA</type>
	    <count>1</count>
	   </coproc>
	  <file_ref>
	    <file_name>setiathome_x41zc_x86_64-pc-linux-gnu_cuda60</file_name>
	    <main_program/>
	  </file_ref>
	  <file_ref>
	    <file_name>libcudart.so.6.0</file_name>
	  </file_ref>
	  <file_ref>
	    <file_name>libcufft.so.6.0</file_name>
	  </file_ref>
   </app_version>

I'm not having any trouble receiving files right now, even have a few APs.