I have manually installed CUDA without JetPack, which works. I am thinking about two possibilities for why it won’t work. First, see if the nVidia-specific files are correct:
sha1sum -c /etc/nv_tegra_release
Second, the .deb file used for installing CUDA is actually a repository file, and then the apt package system does the actual CUDA install. If you got the wrong repo .deb file, then this would account for the problem. I can’t use JetPack because my host is Fedora, but JetPack does offer some options for what to install where, I’m wondering if there was any kind of selection where somehow it might have thought you intended to install CUDA for something else which wasn’t intended for a Jetson? E.G., JetPack is able to install CUDA to the x86_64 host, and JetPack runs on an x86_64 host, but options give JetPack access to multiple CUDA repo .deb files other than x86_64. The repo .deb file which goes with R23.2 is this one:
…this is because R23.2 has mixed 32-bit and 64-bit support. If you were installing to R24.1, then this is the .deb file used:
Double check to see if there are options for 32-bit armhf CUDA 7.0 configured, or if maybe it got the 64-bit arm64 CUDA 7.0 instead.