CUDA 55 tasks fail on 384.90 (Debian) driver


Message boards : Problems and bug reports : CUDA 55 tasks fail on 384.90 (Debian) driver

Message board moderation

To post messages, you must log in.
AuthorMessage
MarkJ
Avatar

Send message
Joined: 27 Jun 12
Posts: 129
Credit: 62,716,918
RAC: 78
Message 5515 - Posted: 15 Oct 2017, 9:38:39 UTC
Tried running some tasks after upgrading nvidia-kernel-dkms to the 384.90 driver. All the CUDA55 tasks failed. Seti CUDA80 tasks work fine on it.

Link to one of the failing tasks: result 173016915

Stderr from task
<core_client_version>7.8.3</core_client_version>
<![CDATA[
<message>
process exited with code 2 (0x2, -254)</message>
<stderr_txt>
CUDA RC12!!!!!!!!!!
CUDA Device number: 0
CUDA Device: GeForce GTX 1060 3GB
Compute capability: 6.1
Multiprocessors: 9
Grid dim: 144 = 9*16
Block dim: 128

</stderr_txt>
]]>

BOINC startup
15-10-2017 08:18 PM CUDA: NVIDIA GPU 0: GeForce GTX 1060 3GB (driver version 384.90, CUDA version 9.0, compute capability 6.1, 3014MB, 2918MB available, 3936 GFLOPS peak)
15-10-2017 08:18 PM OpenCL: NVIDIA GPU 0: GeForce GTX 1060 3GB (driver version 384.90, device version OpenCL 1.2 CUDA, 3014MB, 2918MB available, 3936 GFLOPS peak)
BOINC blog
ID: 5515 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
MarkJ
Avatar

Send message
Joined: 27 Jun 12
Posts: 129
Credit: 62,716,918
RAC: 78
Message 5700 - Posted: 20 Feb 2018, 5:48:53 UTC
The 384.111 driver seems to work, so I think we can safely say it’s something in the older driver.

Due to a bug in the BOINC client it truncates the CUDA minor version number to two digits so it appears as 384.11. I reported it on the BOINC mailing lists when the driver came out.
BOINC blog
ID: 5700 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote

Message boards : Problems and bug reports : CUDA 55 tasks fail on 384.90 (Debian) driver