Unable to install cuda support on jetson nano

TO BEGIN WITH JETSON NANO JETPACK ONLY VERSION4.6.X IS SUPPORTED. THAT IS STRANGE CONSIDERING JETPACK IS VERSION 5. THEN WHEN I INSTALL THE STUFF USING THIS: Quick Start — Jetson Linux<br/>Developer Guide 34.1 documentation
IT DOES NOT INSTALL CUDA SUPPORT. I could not find instructions there that would get me to install cuda support. I will appreciate if someone can let me know some CLEAR INSTRUCTIONS STEP BY STEP FOR INSTALLING CUDA SUPPORT IN JETSON NAN O 2GB.
REQUEST: PLZ DO NOT RESPOND IF U DO NOT KNOW. IN THE PAST I HAVE POSTED MESSSAGES AND HAVE RECIEVED RESPONSES THAT DONT WORK OR ARE MISLEADING. PLZ DO NOT DO THAT.
tHANK YOU
PS: I just need the CUDA runtime installation

Hi,

I believe the Jetson Nano is coming to its EOL cycle in favor of newer Jetson platform flavors, so the latest JetPack that supports the Jetson Nano is version 4.6.3. For the most straightforward way to have a system image with CUDA support you have two options:

  1. Flash the SD card image that can be downloaded here: https://developer.nvidia.com/jetpack-sdk-463

  2. Use the sdkmanager. Instructions here: Installing Jetpack 4.6 | NVIDIA SDK Manager | RidgeRun Developer - RidgeRun Developer Connection

Jafet Chaves,
Embedded SW Engineer at RidgeRun
Contact us: support@ridgerun.com
Developers wiki: https://developer.ridgerun.com/
Website: www.ridgerun.com

1 Like

@jchaves is correct.

I would be surprised if JP 5 allowed any kind of flash or install to a Nano. You would need at least Xavier (which won’t be supported beyond JP 5, so Orin is even better). The integrated GPU of Jetsons (iGPU) uses different drivers and libraries than do the discrete GPUs (dGPU) on a PCIe slot. The Nano itself has a TX1 SoC, which is the very first SoC with 64-bit ARM (and started with compatibility 32-bit user space software). New features stopped being added quite some time back, and it went into maintenance mode (e.g., security updates). You won’t find anything newer than from the most recent JP 4.x release. The listing of releases here will indicate JP 5 is not compatible:

1 Like

Thank you @jchaves and @linuxdev . I did install it via the sdk but I get this error:
W tensorflow/stream_executor/platform/default/dso_loader.cc:65] Could not load dynamic library ‘libcudart.so.10.2’; dlerror: libcudart.so.10.2: cannot open shared object file: No such file or directory
BTW I hv only installed tensorflow not tensorflow-gpu. Is that the reason why I see that error which says: Could not load dynamic library ‘libcudart.so.10.2’; dlerror: libcudart.so.10.2. …
Also Strange: There is no cuda directory under /etc/local
Also: Inferencing works but I suppose it is using the CPU of the jetson nano and not the GPU since it is unable to use CUDA?
Any suggestion? I think I will make a separate question on this

FYI, looks like you have a duplicate thread which I already answered in here:
https://forums.developer.nvidia.com/t/jetson-nano-cuda-error-could-not-load-dynamic-library-libcudart-so-10-2/244323

Note that if you install CUDA via JetPack/SDKM, then it is possible it sets up “/etc” differently than if you just install CUDA from an apt-get command.

Yes, some software will switch to CPU-only if there is no CUDA detected. First make sure CUDA is in your linker path as described in the other thread.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.