There's still some problems to be ironed out for output fidelity on GPU computing.
Last time I checked in my field (biophysics simulations), a basic NVE system of particles don't always come out the same on GPU-enabled molecular simulations on repetition with the same initial seed. Bad, but some people can afford to gloss-over those "minor violations" to the statistical ensemble as long as the overall results are agreeable.
The whole story changes for multiplayer since you're running two simulations which have to agree. The current implementation in, say, NAMD, is most reliable for single-node, shared memory builds - essentially single player. As soon as you start communicating across multiple nodes on a cluster, oh boy...
The point is, people get jobs just to do this kind of thing full-time.
