Sorry if my comment was misunderstanding. I used probably because I’ve learnt to be cautious when posting.
I don’t know how to do that.
Considering that NVIDIA didn’t port Nano GPU driver to kernel-5 although having full details and sources, I suspect it is not trivial.
I think this would probably require a huge work of retro engineering for an unsure result.
And even if I knew that (which is not the case), I wouldn’t be allowed to disclose such proprietary information.
Incidentally, in your earlier post, you quoted some compiler commands and their results. In that command you had several architectures via content such as:
...
-gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70
...
The above is an excerpt of your compile. Note that you can specify as many architectures as you want, and then the different GPUs will work if the CPU architecture is correct. You are missing the architecture which @Honey_Patouceul mentioned, “-gencode arch=compute_53,code=sm_53”. You would need to add this. If it runs only on that Jetson, then you could remove all of the other architectures. This applies regardless of the CUDA version used.
However, it still won’t work. CUDA 11 cannot work on a Nano.
I want to believe you,but there is this tutorial :
that keeps my hopes alive. The title speak clear. And since it has been written 3 years ago,it cannot be referred to the jetson orin. I tried what he suggested,but unfortunately I haven’t been able to complete it because it needs a specific file that’s so old that it does not exist anymore on the repositories.
I have doubts CUDA 11 can be made to work on a Nano, at least not via GPU. If it can, the implication is that the X server was also updated (the X server is not just for graphics; it is a uniform API for GPU access, and the NVIDIA driver is a binary plug-in to the X server; not all CUDA uses X, but much of it does).
what’s the problem if the X server needs to be updated ? the newer release what will break ?
Each X server has a specific ABI (Application Binary Interface) for dynamically loading compiled plug-ins. The NVIDIA content is provided only as a binary file. At least it all used to be, I think maybe there was a recent release of some part of it as source code (I hope someone from NVIDIA can mention what part of this is now available as source code). I’m not certain, but so far as I know, it isn’t possible for the end user to recompile that content such that it would work with a different ABI (and the newer server has a newer ABI).
There are actually multiple plug-ins for the X server, e.g., the input devices (mouse, keyboard) also work as a plug-in. The part which lacks the correct ABI will fail to load and that part of the server will either fail or revert to some other version. See:
grep 'ABI' /var/log/Xorg.0.log
If this is available as source code, then it is very new. Could someone from NVIDIA comment on whether or not all plug-ins to the X server are now available for recompile against another ABI?
The link you shared may be used for a discrete NVIDIA Ampere GPU (connected to a PCIe connector/adapter). This doesn’t apply to Nano’s Maxwell integrated GPU.
Just to be sure… Are you using Jetson Nano or Jetson Orin Nano ? For the latter only, JP5 would be available and you would be able to use CUDA11.
nano.