Cross compilation: Host vs target Cuda version

I have a PX2 with OS version, thus CUDA 9, cuDNN 7.0.4.
I wanted to give it a try with cross compilation, but I’ve noticed that by installing the nvidia software on the host machine by sdkmanager application, a more recent version of cuda (9.2) is installed.
Will it still be possible to crosscompile application for the PX2 board or shall I upgrade the DRIVE OS by flashing the PX2 ?

Thank you

Dear dario.turchi,
Please follow if you have any CUDA version issues

Ok, but that doesn’t entirely answer my question.
My situation is the following:
CUDA 9.2 on the host machine installed by sdkmanager, CUDA 9 on the target installed with the drive OS.
What are you suggesting is to manually install CUDA 9 on the host and then follow the procedure explained in your link to select the specific CUDA version to use for crosscompilation ?

How do I manually install CUDA 9 for cross compilation ? Simply installing the cuda-toolkit from will be enough ?

Dear dario.turchi,
It seems the PDK versions on host and Drive Px2 are different. So DW cross compilation results in error. It is advisable to upgrade to latest PDK. Could you flash the PX2 with latest PDK(Drive 5.0.10)