Fedora 29: no GPU although CUDA is installed

Message boards : Questions and problems : Fedora 29: no GPU although CUDA is installed
Message board moderation

To post messages, you must log in.

AuthorMessage
Oxy

Send message
Joined: 26 Dec 18
Posts: 2
Germany
Message 89339 - Posted: 26 Dec 2018, 16:09:54 UTC
Last modified: 26 Dec 2018, 16:10:13 UTC

I have a fresh new Fedora 29 installation but can't run any GPU-task. Boinc client shows the following messages on start-up:

26-Dec-2018 17:02:07 [---] cc_config.xml not found - using defaults
26-Dec-2018 17:02:07 [---] Starting BOINC client version 7.14.2 for x86_64-pc-linux-gnu
26-Dec-2018 17:02:07 [---] log flags: file_xfer, sched_ops, task
26-Dec-2018 17:02:07 [---] Libraries: libcurl/7.61.1 OpenSSL/1.1.1 zlib/1.2.11 brotli/1.0.5 libidn2/2.0.5 libpsl/0.20.2 (+libidn2/2.0.5) libssh/0.8.5/openssl/zlib nghttp2/1.34.0
26-Dec-2018 17:02:07 [---] Data directory: /var/lib/boinc
26-Dec-2018 17:02:09 [---] OpenCL: NVIDIA GPU 0: NVE7 (driver version 18.2.6, device version OpenCL 1.1 Mesa 18.2.6, 1048576MB, 1048576MB available, 98 GFLOPS peak)
26-Dec-2018 17:02:09 [---] OpenCL: NVIDIA GPU 1: NVE7 (driver version 18.2.6, device version OpenCL 1.1 Mesa 18.2.6, 1048576MB, 1048576MB available, 98 GFLOPS peak)
26-Dec-2018 17:02:09 [---] OpenCL: AMD/ATI GPU 0: AMD CEDAR (DRM 2.50.0 / 4.19.10-300.fc29.x86_64, LLVM 7.0.0) (driver version 18.2.6, device version OpenCL 1.1 Mesa 18.2.6, 2048MB, 2048MB available, 104 GFLOPS peak)
26-Dec-2018 17:02:09 [---] OpenCL CPU: pthread-AMD FX(tm)-8350 Eight-Core Processor (OpenCL driver vendor: The pocl project, driver version 1.2, device version OpenCL 1.2 pocl HSTR: pthread-x86_64-unknown-linux-gnu-bdver2)
26-Dec-2018 17:02:09 [---] App version needs CUDA but GPU doesn't support it


When I add Einstein@home it tries to use the ATI-GPU but fails with "calculation error" immediately. The already existing SETI@home-tasks are all rejected due to missing CUDA (buth which is there, all Nvidia/CUDA packages are installed).

So...any idea what could be wrong/what to check?

Thanks!
ID: 89339 · Report as offensive
Profile Keith Myers
Volunteer tester
Help desk expert
Avatar

Send message
Joined: 17 Nov 16
Posts: 863
United States
Message 89367 - Posted: 27 Dec 2018, 18:28:38 UTC - in response to Message 89339.  

You don't have any CUDA drivers loaded. You have the Mesa drivers loaded which is only OpenCL. But the real problem is the Mesa OpenCL driver is not a 'real' OpenCL driver. It's version of OpenCL is not compatible with the majority of projects that have OpenCL science applications like Seti.

If you say that you have installed the proprietary Nvidia drivers then the Mesa drivers are the only ones loading. If you don't see a statement in the beginning of the Event Log proclaiming the Nvidia CUDA loading like this:
26-Dec-2018 07:22:29 [---] CUDA: NVIDIA GPU 0: GeForce GTX 1080 (driver version 410.78, CUDA version 10.0, compute capability 6.1, 4096MB, 3980MB available, 9523 GFLOPS peak)

then you aren't using your Nvidia drivers. Normally when you install the Nvidia drivers, they blacklist the nouveau drivers which is what reports the Mesa drivers. So the blacklist didn't happen. You can always create your own blacklist.

But I don't have a clue what your Mesa driver is reporting as a NVE7 device. I don't know of any previous Nvidia card identified as such.
ID: 89367 · Report as offensive
Oxy

Send message
Joined: 26 Dec 18
Posts: 2
Germany
Message 89376 - Posted: 28 Dec 2018, 13:14:15 UTC

Finally: blacklisting the nouveau-drivers and disabling OpenCL via cc_confog.xml did do the trick - thanks!!! :-)

Now crunching again with 8 CPUs and 2 GPUs :-)
ID: 89376 · Report as offensive
Profile Keith Myers
Volunteer tester
Help desk expert
Avatar

Send message
Joined: 17 Nov 16
Posts: 863
United States
Message 89379 - Posted: 28 Dec 2018, 21:11:44 UTC - in response to Message 89376.  

The manufacturer provided OpenCL drivers, i.e. Nvidia or AMD are proper OpenCL drivers. And you should be able to use science apps that require OpenCL. There are quite a few. MilkyWay and Einstein are some as well as the stock Seti gpu app which is also OpenCL based.

It is only the pseudo OpenCL support provided by the nouveau and Mesa drivers that cause problems for compute. If you have installed the Nvidia proprietary drivers you should have OpenCL support.

One quick way to check is to install clinfo
 sudo apt install clinfo
and run clinfo in Terminal. It will report the CUDA version of the drivers and the OpenCL version of the drivers. If for some reason the Nvidia metapackage of the drivers missed the OpenCL installation you can always load them separately.
sudo apt-get install ocl-icd-libopencl1
ID: 89379 · Report as offensive
mmonnin

Send message
Joined: 1 Jul 16
Posts: 146
United States
Message 89401 - Posted: 31 Dec 2018, 18:29:18 UTC

Have you restarted the BOINC Client after bootup? On one PC of mine the client loads prior to the AMD drivers so BOINC never saw the GPU even though clinfo showed the drivers. Restarting the client after PC boot allowed the client to use the GPU.

If not then I've had to install 390 and 396 a few times in Linux to pick up CUDA correctly.
ID: 89401 · Report as offensive
Richard Haselgrove
Volunteer tester
Help desk expert

Send message
Joined: 5 Oct 06
Posts: 5077
United Kingdom
Message 89696 - Posted: 19 Jan 2019, 22:51:16 UTC - in response to Message 89695.  
Last modified: 19 Jan 2019, 22:51:28 UTC

It depends what tool you use for your 'remote session'.

If you use Microsoft's Remote Desktop, it loads a cut-down GPU driver which can't be used by BOINC.

If you use {any, if not all} of the third-party remote access tools, they use the installed standard driver, and BOINC can continue to use the GPU.
ID: 89696 · Report as offensive
mitchb13

Send message
Joined: 19 Jan 19
Posts: 2
United States
Message 89707 - Posted: 21 Jan 2019, 0:12:15 UTC - in response to Message 89696.  

It was Remote Desktop. I kept installing drivers and never noticed that the problem only happened when I used Remote Desktop.
ID: 89707 · Report as offensive

Message boards : Questions and problems : Fedora 29: no GPU although CUDA is installed

Copyright © 2024 University of California.
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.