I avoid making such comparisons. Each GPU based MD out there right now (HOOMD, NAMD, Ascalaph, and OpenMM) optimizes for a different scenario and implements a different set of interactions. In principle, one could take them all to the least common denominator and just benchmark a Lennard-Jones liquid, but even that would be unfair to some degree. For instance, the current version of OpenMM is O(N^2) and optimized for systems of less than 5,000 particles. And the NAMD code carries so much extra baggage along with their particles for the all-atom force field that HOOMD would blow it away running just a simple LJ liquid.
Basically, you need to identify your particular MD model and see which of the GPU MD codes out there is best optimized for that case.
HOOMD: General-purpose, but primarily targeted at coarse-grained models
OpenMM: Tuned for very small molecules in an implicit solvent (currently)
NAMD: Tuned for the only thing that NAMD does: gigantic biomolecules, typically in explicit solvent
Ascalaph: I’m not sure what they optimize for… the only info they give is a benchmark of a SPC water model
I’ll drop a shameless plug here and add that of all of them, HOOMD is by far the most general purpose of the bunch and optimized for just about any case you can throw at it (small systems, large systems, dense systems, dilute systems, …). Perhaps the biggest drawback currently (especially if you want an all-atom force-field) is the lack of electrostatic forces. I say “currently” because HOOMD is under very active development right now and electrostatics will likely be implemented in a couple months.
I saw a presentation by one of the OpenMM developers last week. They now have a solution for O(N) MD which will be in the next preview release. That should open OpenMM up to work efficiently on a much wider variety of problems.