'cuda-memcheck' works on one machine, does not work on some others

I am trying to use cuda-memcheck to diagnose some of my code, but I am having trouble getting it running on some of my machines. My hardware is:

Machine 1: Intel Core i7-6770HQ, GTX 1080Ti (connected through Thunderbolt as an eGPU)
Machine 2: Intel Core i7-6600K, GTX 1070
Machine 3: Intel Core i5-6300HQ, GTX 960M.

All three machines are running CUDA 9.0 on Ubuntu 16.04, and I am attempting to run cuda-memcheck on the same code. It only works on machine 1; machines 2 and 3 report:

======= Internal error (7)
======= No CUDA-MEMCHECK results found

My command usually is
cuda-memcheck --leak-check full BINARY

  • I have made sure that I have write access to whatever directory I am using cuda-memcheck in.
  • The CUDA include files and libraries are already in PATH and LD_LIBRARY_PATH respectively.
  • Also, there is no version conflict between the CUDA SDK and the CUDA-MEMCHECK program.

On the non-working machines, I tried doing a clean uninstall of the CUDA SDK, reinstalled the graphics driver and reinstalled CUDA, but with no luck when it comes to cuda-memcheck.

How can I debug this problem?