Hello,
I am attempting to flash my TX1 using the sdkmanager. I have found that after loading JetPack 4.4.1 on the machine that I do not have sufficient space to install PyTorch, Torchvision, or the networks that I would like to have.
Can anyone provide some help to me on how to solve the space issue on my TX1?
I have a 32gb SD card and I am storing things on it. However after flashing I am left with only a couple gigabytes of storage on the TX1 itself, therefore attempting to install libraries or things I need is not an option.
Can anyone suggest some options I have to address the space issue on the TX1?
https://elinux.org/Jetson/L4T/Boot_From_External_Device
please refer to this page to mount rootfs on external device.
The method you linked uses the commandline rather than the sdkmanager to flash. Flashing from the command line leaves me at 40% use of the TX1 space, but I am missing some dependencies in CUDA and other things I need for the tutorials on the TX1.
Is there a way to run SDKMANAGER and flash into the SD card? Also I am having some issues with figuring out how to run the commands in the link you gave.
Specifically I am having difficulty in the figuring out how to use gdisk to get the correct filesystem and partition sizes.
Can you help me with this please?
FYI, you can then run SDKM, but deselect flash. You can also deselect installing anything to the host PC. Just make sure your Jetson is fully booted and has an account you can log in to (first boot setup) so that ssh can work. Then install CUDA or the other dependencies.
Note that when SDKM runs it starts out in recovery mode, which is needed if and only if you are flashing. When it comes time for the optional extras (e.g., CUDA), the install is over ssh on a fully booted unit.
Starting in the R32.x series (not the very first release, but only a couple of releases in) many packages became available via the standard “sudo apt-get install ...” mechanism. You could also install using apt-get without the SDKM for that case. Example:
sudo apt search nvidia-
sudo apt-get install nvidia-l4t-cuda
Notice that I searched not for “nvidia”, but instead for “nvidia-”. The reason is that there are packages related to “nvidia”, but which are not part of the Jetson packages. All of the NVIDIA Jetson packages start with “nvidia-something”, and adding that hyphen in there gives a more exact list of the Jetson packages.
Question,
What is the difference between the nvidia-l4t-cuda and nvidia-cuda packages?
When I am running a cmake of a tutorial I am missing the cuda toolkit. I am wondering if that is found in the nvidia-cuda package.
The “-l4t-” version works on a Jetson’s operating system. The non-l4t CUDA is for a PC. The two are not interchangeable, but both can support the same API, e.g., CUDA 10 has both L4T releases and PC releases. You just could not use the L4T version on a PC, and a PC version would not run on a Jetson. Software compiled for the right architecture would work on either if it runs the same release, e.g., CUDA 10.