I believe the GPU machine type is legacy code that probably should have been removed.
Our lab has made good use of domdec_gpu, but it is a hybrid method and requires a good balance of CPU, GPU, and a fast node interconnect such as Infiniband for optimal use.
OpenMM works well on GPUs, although I've come to prefer the Python interface. It can read CHARMM files, notably topology, parameter, psf, and coord files; it produces CHARMM compatible trajectory files, a.k.a. DCD format. I'm not sure the CHARMM implementation supports some newer features available with the OpenMM Python interface, such as Drude force fields, long range VDW via LJ-PME, or the Nose-Hoover thermostat.
OpenMM is good for producing ensembles, but needs further development to support computing time dependent properties such as relaxation phenomena, diffusion, and viscosity.