Installing CUDA 9 without Flashing

Is it possible to install CUDA 9 on the TX2 without flashing the OS? Flashing is not currently an option for me due to extensive setup that would be difficult to reproduce.

I was able to use the repository .deb files that Jetpack 3.2 downloaded to manually install the CUDA 9.0 toolkit without connecting directly to the device, using the instructions found in ~/jetson/_installer/cuda-l4t.sh. I can build the samples, however when running them, I get an error message stating that “CUDA driver version is insufficient for CUDA runtime version”. Setting LD_LIBRARY_PATH appropriately does not fix the problem.

Is there any way to update the drivers on the device without wiping the root filesystem? I can blow away /usr and /var if necessary, but not /etc. I am comfortable with the configuration, compilation, and installation of the linux kernel as well as kernel modules. I can gain access to the machine to program via USB, though it is much easier to upgrade remotely via an ssh connection. Thanks!

Hi,

As you mentioned, CUDA 9.0 requires the GPU driver contained in rel-28.2.
So it won’t work with JetPack3.1, which is rel-28.1.

You can check our cloning document here: https://elinux.org/Jetson/TX2_Cloning
But suppose this approach is only available fore restoring the image of same version.

It’s still recommended to reflash your device with JetPack3.2 to well-set all the environment.

Thanks.

Alright, this is what I was afraid of. I think we’re going to have to pass on this version, as the amount of time it would take to set everything back up would be prohibitive. Also, we don’t even want CUDA 9.0, CUDA 9.1 would be much preferred as we are actually targeting the DRIVE PX 2, which, for whatever reason, does not have CUDA 9.0 support.

Is there a chance that we will see CUDA 9.1 on the TX2 later this year?

Hi,

Sorry that we can’t disclosure our future plan : (

You can pay attention to our JetPack page and TX2 forum.
All the release update can be found on our forum once available.

Thanks.