Quite some time back CUDA was available as a separate install, but eventually required JetPack under arm/arm64. I think desktop is still available separately. JetPack should be able to install CUDA on the Jetson, it’s just a case of selecting the TK1 and then CUDA (you can deselect all else, e.g., deselect flash).
I install CUDA manually (and video driver) on both my host and Jetsons…I can’t use JetPack because I use a Fedora host. Manual install works for both host and Jetson. Here is the URL for desktop CUDA downloads:
https://developer.nvidia.com/cuda-downloads
I personally just use the “runfile” on my x86_64 host…this puts CUDA in “/usr/local/”, then I add required locations to my PATH environment variable.
Note that the last release for 32-bit is CUDA version 6.5 (there won’t be any newer versions for 32-bit). Assuming you are using the most recent L4T version on your JTK1 (R21.5), this is the URL for downloads, including JetPack for that version:
https://developer.nvidia.com/linux-tegra-r215
Normally what JetPack will do for CUDA is to install a “.deb” package which makes CUDA and other packages visible via a local repository in “/var” on the JTK1. This does not in itself install CUDA, but it makes it so apt (and apt-get) can search for and find CUDA. Once you install that “.deb” file it is an ordinary Ubuntu package install, e.g., “apt search cuda”, followed by the application “sudo apt-get install cuda…”.
If you can use JetPack, just let it install CUDA and ignore the details. If you cannot use JetPack, then you have to find out what the download URL is from JetPack and manually download and install the “.deb”, followed by search and install of CUDA via apt. To get JetPack to unpack some of its files on a non-Ubuntu 14.04LTS system (e.g., from Fedora) you would run this command:
bash ./JetPack-L4T-2.3.1-linux-x64.run --noexec
./Chooser
# You can exit Chooser now
After doing the above you will get file “repository.json”. This file contains the URLs for downloading various packages manually (be careful to note architecture…a JTK1 is arm/armhf…you can wget your desired package). For CUDA you’d search that file for the URL of cuda (6.5), download that cuda repo file, copy it to the Jetson, and install via dpkg. After that “apt update”, then any apt search will show CUDA packages which apt can install by standard Ubuntu apt-get methods.