Build using CUDA 7.5 for CUDA 6.5 runtime?

I’m trying to cross-compile code for the TegraK1 Meerkat. It has L4T release 21 revision 3, which includes CUDA 6.5. I’ve been informed that updating the CUDA runtime requires updating the kernel… so a reasonable alternative seemed to be to try a cross-compile.

For the cross-compile I’m using CUDA 7.5 on Ubuntu 14.04.
I’m trying to build the the deviceQuery executable from the sample code.

The build command I’m using is:
make TARGET_ARCH=armv7l TARGET_OS=linux SMS=32
(Specifying SMS appends “-gencode arch=compute_32,code=sm_32” to the build arguments)

However, when I try executing on the Tegra I get the following error:
cudaGetDeviceCount returned 35
→ CUDA driver version is insufficient for CUDA runtime version

Can anyone point me to documentation about how to build for previous runtime versions?

Discussion about CUDA drivers:

L4T Release Notes:

Compute versions:

To build for a previous version you have to install a previous version. It’s possible to have both CUDA 6.5 and CUDA 7.5 installed on your development machine. But if you want to build a binary suitable for deployment in a CUDA 6.5 (driver) setup, you will need the CUDA 6.5 libraries and compiler on your development machine.

Thank you @txbob !

That completely answered my question.