FS#77800 - CUDA update to 12.1 breaks CUDA tools
Attached to Project:
Community Packages
Opened by Michael (ZeroBeat) - Thursday, 09 March 2023, 22:21 GMT
Last edited by Toolybird (Toolybird) - Thursday, 23 March 2023, 20:29 GMT
Opened by Michael (ZeroBeat) - Thursday, 09 March 2023, 22:21 GMT
Last edited by Toolybird (Toolybird) - Thursday, 23 March 2023, 20:29 GMT
|
Details
Description:
Latest CUDA update to version 12.1 $ pacman -Q | grep cuda cuda 12.1.0-1 doesn't match to the driver $ pacman -Q | grep nvidia nvidia 525.89.02-6 nvidia-settings 525.89.02-1 nvidia-utils 525.89.02-2 opencl-nvidia 525.89.02-2 according to https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html this driver (currently beta) is mandatory: CUDA 12.1 GA >= 530.30.02 Trying to compile a CUDA application on current driver 525 will end up in an ERROR: cuLinkAddData(): the provided PTX was compiled with an unsupported toolchain. * Device #1: Kernel /usr/share/hashcat/OpenCL/shared.cl link failed. Error Log: ptxas application ptx input, line 9; fatal : Unsupported .version 8.1; current version is '8.0' * Device #1: Kernel /usr/share/hashcat/OpenCL/shared.cl build failed. Solution: Downgrade to 12.0 until driver 530.30.02 is in the repository. I suggest to add a notice to "latest news" that users should not update to CUDA 12.1 unless 530.30 has been published. |
This task depends upon
Closed by Toolybird (Toolybird)
Thursday, 23 March 2023, 20:29 GMT
Reason for closing: Fixed
Additional comments about closing: nvidia 530.41.03-1
Thursday, 23 March 2023, 20:29 GMT
Reason for closing: Fixed
Additional comments about closing: nvidia 530.41.03-1
https://bugs.archlinux.org/task/69005
The steps to reproduce it are the same as mentioned in FSX6905:
$ pacman -Q | grep cuda
cuda 12.0.1-1
$ git clone https://github.com/hashcat/hashcat
$ cd hashcat
$ make
everything is compiled fine.
Now do the update to CUDA 12.1 and compile it again:
cuLinkAddData(): the provided PTX was compiled with an unsupported toolchain.
* Device #1: Kernel /usr/share/hashcat/OpenCL/shared.cl link failed. Error Log:
ptxas application ptx input, line 9; fatal : Unsupported .version 8.1; current version is '8.0'
* Device #1: Kernel /usr/share/hashcat/OpenCL/shared.cl build failed.
API of the 525 dirver is 8.0
API of CUDA is 8.1
API of next driver 530 will be 8.1
https://bugs.archlinux.org/task/69005
"Comment by Michael (ZeroBeat) - Thursday, 17 December 2020, 15:36 GMT
I wonder why NVIDIA released CUDA 11.2 before release of driver 460.27.4 (with regard to the API change 7.1 -> 7.2)."
And I still wonder (again) why NVIDIA released CUDA 12.1 before release of driver 530 (with regard to the API change 8.0 -> 8.1).
We can strongly assume that this will happen again on change of API 8.1 -> 8.2 and later on, on 9.0 -> 9.1
Regarding Linux, NVIDA always publish CUDA update before the driver has left beta status.
GeForce Game Ready Driver
Version: 531.18 WHQL
Release Date: 2023.2.28
Operating System: Windows 10 64-bit, Windows 11
Language: English (US)
File Size: 855.68 MB
but never ever together with the Linux driver update:
Linux x64 (AMD64/EM64T) Display Driver
Version: 525.89.02
Release Date: 2023.2.8
Operating System: Linux 64-bit
Language: English (US)
File Size: 394.93 MB
Current output of nvidia-smi (regardless whether CUDA 12.0 or CUDA 12.1 is installed):
NVIDIA-SMI 525.89.02 Driver Version: 525.89.02 CUDA Version: 12.0
Is there any reason not to retract the cuda 12.1 update until the compatible nvidia 530 update is available?
Now a short message in latest news should be ok.
Usually it will not take a long time until the final version of the driver will be released.
The next upcoming CUDA/driver updates will be ok - as long as ptx doesn't change.
Luckily this will be announced in the compatibility table.
https://www.nvidia.com/download/driverResults.aspx/200481/en-us/