The CHARMM/OpenMM interface has been undergoing continuous development, as has OpenMM itself.
I've found that CHARMM/OpenMM has dependencies on the CHARMM version and the OpenMM version, and that optimal pairings are (alas) not listed in the CHARMM documentation. OpenMM also has CUDA dependencies, but these are clearly listed on the OpenMM site.
With c39b1, I ended up building with OpenMM 5.2.0 and CUDA 5.5, and was able to run the test cases, but I did not pursue it much further (e.g. dual GPUs). I have instead been focusing on DOMDEC usage, which I regard as superior.
I had experienced and reported (see below) similar problems with dual-GPU usage about a year ago for c39a2, the development version which became c39b1.
Bug tracking is gradually moving to www.charmm.org/redmine