FS#74684 - Enable cuda sm_37 for package tensorflow-opt-cuda

Attached to Project: Community Packages
Opened by Feng Wang (wangfeng) - Monday, 09 May 2022, 06:51 GMT
Last edited by Konstantin Gizdov (kgizdov) - Sunday, 10 July 2022, 10:02 GMT
Task Type Feature Request
Category Packages
Status Closed
Assigned To Sven-Hendrik Haase (Svenstaro)
Konstantin Gizdov (kgizdov)
Architecture All
Severity Medium
Priority Normal
Reported Version
Due in Version Undecided
Due Date Undecided
Percent Complete 100%
Votes 0
Private No

Details

Description:

My old server equips a legacy Tesla K80 GPU. This GPU supports CUDA up to sm_37.
However, tensorflow-opt-cuda is built with CUDA compute capability starting from sm_52.


I would suggest updating the PKGBUILd this line
`export TF_CUDA_COMPUTE_CAPABILITIES=sm_52,sm_53,sm_60,sm_61,sm_62,sm_70,sm_72,sm_75,sm_80,sm_86,compute_86`
to
`export TF_CUDA_COMPUTE_CAPABILITIES=sm_35,sm_37,sm_50,sm_52,sm_53,sm_60,sm_61,sm_62,sm_70,sm_72,sm_75,sm_80,sm_86,compute_86`,
so that it will also support legacy Nvidia GPUs.

Additional info:
* package version(s):
python-tensorflow-opt-cuda 2.8.0-7


Steps to reproduce:
1. Find a workstation with a K80 GPU installed
2. Install python-tensorflow-opt-cuda package
3. Run a small deep learning application, such as <https://github.com/keras-team/keras-io/blob/master/examples/vision/mnist_convnet.py>
4. Tensorflow will complain that sm_37 is not enabled, and decide to recompile itself with sm_37 option on the fly.
This task depends upon

Closed by  Konstantin Gizdov (kgizdov)
Sunday, 10 July 2022, 10:02 GMT
Reason for closing:  Won't fix
Additional comments about closing:  reluctant to include upstream deprecated features
Comment by Toolybird (Toolybird) - Sunday, 10 July 2022, 05:25 GMT
Support for legacy capabilities was removed [1] in 2020. Assigning to PM's who are best equipped to decide on this request.

[1] https://github.com/archlinux/svntogit-community/commit/86eb65c3
Comment by Konstantin Gizdov (kgizdov) - Sunday, 10 July 2022, 10:01 GMT
This is not practically possible since CUDA 11 as far as I know. Support for these architectures has been deprecated upstream - https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#title-new-cuda-libraries

We've tried to compile packages with 3.5 and 3.7 before, but they don't seem to work. If you are able to compile the whole chain of packages dependent on CUDA with these architectures included and can share working PKGBUILDs, then we can talk again.

However, we are normally reluctant to include deprecated features as these can be removed upstream at any time and break things for everyone.

Loading...