Not just GPUs, most labs don't embrace parallelism in general. I managed to get a 55x speedup between serial and parallel versions using the same program (SPM) by using several mid-grade computers and caching (Sorry I don't have anything written on this, I've been putting it off).

It's hard to get researchers to switch using the algorithms that they know and are comfortable with, which is why I think stuff like BROCCOLI, or PSOM haven't taken off.