CUDA on Linux

log in

Advanced search

Message boards : Cruncher's Corner : CUDA on Linux

Author Message
Profile microchip
Send message
Joined: 10 Jun 06
Posts: 48
Credit: 6,947,269
RAC: 0
Message 111676 - Posted: 15 Apr 2011, 16:57:25 UTC

Hi,

Got a question here. I'm fairly new to CUDA and would like to add a CUDA capable GPU in my crunch system. I'm on Linux and my question is if I have to install anything special to make it work with BOINC and E@H? I mean do I need the proprietary driver from nVIDIA and do I need some specific CUDA kit for it to work?

Thanks :)
____________

Team Belgium

Michael Karlinsky
Avatar
Send message
Joined: 22 Jan 05
Posts: 896
Credit: 18,976,107
RAC: 14,358
Message 111677 - Posted: 15 Apr 2011, 17:18:30 UTC - in response to Message 111676.

Hi,

you need to install the NVIDIA driver.

download

and

instructions

excerpt:

1) Kernel sources must be installed and configured. Usually this means installing the 'kernel-source', 'make' and 'gcc' packages with YaST2. 2) Run the following commands echo "blacklist nouveau" > /etc/modprobe.d/nvidia.conf # recreate initrd without KMS, if the use of KMS is enabled in initrd if grep -q NO_KMS_IN_INITRD=\"no\" /etc/sysconfig/kernel; then sed -i 's/NO_KMS_IN_INITRD.*/NO_KMS_IN_INITRD="yes"/g' /etc/sysconfig/kernel mkinitrd fi 3) Reboot your machine. 4) Use the nvidia installer for 260.19.36. sh NVIDIA-Linux-x86_64-260.19.36.run -q IMPORTANT: You need to recompile and install the nvidia kernel module after each kernel update. sh NVIDIA-Linux-x86_64-260.19.36.run -K


Michael

PS: instructions are for SUSE, but I guess they should work everywhere else.
____________
Team Linux Users Everywhere
Profile microchip
Send message
Joined: 10 Jun 06
Posts: 48
Credit: 6,947,269
RAC: 0
Message 111678 - Posted: 15 Apr 2011, 17:39:57 UTC

Thanks for the info. Yes, I know how to manually install the NV driver, but thanks anyways... So I don't need anything else for CUDA on linux? I saw on the NV site a CUDA kit but it's an SDK, not sure if I need that
____________

Team Belgium

Michael Karlinsky
Avatar
Send message
Joined: 22 Jan 05
Posts: 896
Credit: 18,976,107
RAC: 14,358
Message 111679 - Posted: 15 Apr 2011, 18:01:16 UTC - in response to Message 111678.

I saw on the NV site a CUDA kit but it's an SDK, not sure if I need that


This is for development only. All BOINC-project (should) provide the rest (there are some libs, which are necessary).

Michael

____________
Team Linux Users Everywhere
Jeroen
Send message
Joined: 25 Nov 05
Posts: 361
Credit: 547,874,400
RAC: 453,733
Message 111687 - Posted: 15 Apr 2011, 22:51:01 UTC - in response to Message 111676.
Last modified: 15 Apr 2011, 22:51:27 UTC

Depending on which GPU you add to your system, you may be able to run multiple work units at once per GPU using an app_info.xml file. Specifically the 4xx and 5xx Fermi GPUs can run two or more work units at a time optimally provided there is enough GPU memory. Each work units requires around 300-350 MB of GPU memory.

Profile microchip
Send message
Joined: 10 Jun 06
Posts: 48
Credit: 6,947,269
RAC: 0
Message 113827 - Posted: 2 Sep 2011, 12:23:43 UTC - in response to Message 111687.

Depending on which GPU you add to your system, you may be able to run multiple work units at once per GPU using an app_info.xml file. Specifically the 4xx and 5xx Fermi GPUs can run two or more work units at a time optimally provided there is enough GPU memory. Each work units requires around 300-350 MB of GPU memory.


Can you be more specific about the app_info.xml file? Where do I find it and what do I have to put in it?
____________

Team Belgium
Richard Haselgrove
Send message
Joined: 10 Dec 05
Posts: 1660
Credit: 54,984,704
RAC: 44,088
Message 113829 - Posted: 2 Sep 2011, 13:33:13 UTC - in response to Message 113827.

I'd recommend that you start with the driver stuff on it's own - you don't need an app_info.xml file for that.

Get it running, check some tasks through and make sure that they validate. The time for moving up to multiple copies is after you've got the plain vanilla version running OK.

Profile microchip
Send message
Joined: 10 Jun 06
Posts: 48
Credit: 6,947,269
RAC: 0
Message 113940 - Posted: 8 Sep 2011, 14:14:26 UTC

Guys, a final question...

Do I need a specific driver version or can I grab the 280.13 one?

I'm asking because on http://developer.nvidia.com/cuda-toolkit-40 I see there's an older driver version (270.41.19) while the normal driver download page of nVIDIA offers version 280.13... so which one should I get ?

Thanks
____________

Team Belgium

Richard Haselgrove
Send message
Joined: 10 Dec 05
Posts: 1660
Credit: 54,984,704
RAC: 44,088
Message 113941 - Posted: 8 Sep 2011, 14:32:30 UTC - in response to Message 113940.

You certainly don't need a developer toolkit. Just the ordinary 280.13 should do it, though we have been having some problems with downclocking with the newer *windows* drivers. If you seem to have any problems like that, look for a menu item for "Beta and Archived Drivers" on the general - not developer - download page, and you should be able to find what you need there.

Profile microchip
Send message
Joined: 10 Jun 06
Posts: 48
Credit: 6,947,269
RAC: 0
Message 113943 - Posted: 8 Sep 2011, 15:35:41 UTC

OK... thanks Richard for your help :)
____________

Team Belgium

Message boards : Cruncher's Corner : CUDA on Linux


Home · Your account · Message boards

This material is based upon work supported by the National Science Foundation (NSF) under Grants PHY-1104902, PHY-1104617 and PHY-1105572 and by the Max Planck Gesellschaft (MPG). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the investigators and do not necessarily reflect the views of the NSF or the MPG.

Copyright © 2016 Bruce Allen