8

I tried uninstalling CUDA 10 and installing 9.2. Now nvcc -V returns 9.2, but nvidia-smi says CUDA 10.0.

Any idea why this may be happening or how to fix it? Can't find anything else related to 10.0 still left on my system.

Edit: added screenshot as reply to nvidia employee showing that nvidia-smi clearly states a CUDA version..

enter image description here

Austin
  • 6,921
  • 12
  • 73
  • 138
  • Check the top-right of the screenshot I just added... can I get that downvote back? – Austin Nov 13 '18 at 20:50
  • That is the CUDA version the driver supports, not the CUDA toolkit version – talonmies Nov 13 '18 at 21:04
  • So you are saying the "CUDA VERSION: 10.0" is not saying that CUDA is version 10.0? If so that's a pretty glaringly obvious point of confusion – Austin Nov 13 '18 at 21:05
  • 4
    nvidia-smi knows absolutely nothing about your cuda toolkit. It is part of the driver. If you run the CUDA runtime API version of devicequery, you will see two CUDA versions reported, the toolkit and the driver. If you run the driver API version, you will see one -- the driver version. That is what nvidia-smi reports. If you want to change that, you would need to downgrade the driver, although it is completely unnecessary to do so. – talonmies Nov 13 '18 at 21:09
  • I wonder whether it is a bug in a particular version of `nvidia-smi`, because on Windows 7, driver 411.63, no CUDA version is reported. – njuffa Nov 13 '18 at 21:10
  • how (from what source, using what method, on what kind of machine) did you install this 410.73 driver? – Robert Crovella Nov 13 '18 at 21:21
  • Amazon Ubuntu 16.04 EC2 P3.2 instance with a V100 GPU. I believe nvidia-smi was preinstalled. I ran the following directions for 10, and then for 9.2: https://www.python36.com/how-to-install-tensorflow-gpu-with-cuda-9-2-for-python-on-ubuntu/ – Austin Nov 13 '18 at 21:27
  • 5
    Yes, this is something new that was inserted in the driver (at least on linux) somewhere between 410.48 and 410.72. I was not previously aware of it. Anyway, @talonmies comments are applicable here. The reported version is the driver compatibility version, not the CUDA runtime (i.e. toolkit) version installed. This will be evident with some thought or testing, as you can witness this with only the GPU driver installed and no CUDA toolkit installed at all. You can also observe it even though you may have multiple CUDA toolkits installed. – Robert Crovella Nov 13 '18 at 21:30
  • 2
    Okay thanks for clearing that up. Makes sense, although changing it to specifically say "CUDA compatibility version" or something along those lines would probably save people like myself some confusion – Austin Nov 13 '18 at 21:34
  • @RobertCrovella wouldn't it make sense for the post marked with a later date to be designated as the duplicate instead? Thanks for that answer by the way. – Austin Nov 29 '18 at 21:00
  • I don't know that there is any such convention or expectation on SO. I'm pretty sure I can find examples of earlier questions marked as duplicate of later. Also, a question needs an upvoted (or accepted) answer before another question can be marked as a duplicate of it. This question has no answer. You can help with the curation of topics here by providing your own answer. If it were sufficient, I'm sure I would have marked the other a duplicate of this one. I'm not sure why it matters. – Robert Crovella Nov 29 '18 at 21:32
  • I just found it a bit odd that you were aware of this post but decided to instead post an answer on a newer version and mark this one as the duplicate. I guess it does not really matter. thanks again for the clarification. – Austin Nov 29 '18 at 21:37

0 Answers0