I think it's likely that only part of a WU can be performed on a GPU, with the remainder requiring the CPU. I also think that it's very important that using a GPU does not compromise the accuracy of the calculations, eg by having to use 32-bit floating-point where 64-bit would be better, or by limiting the choice of algorithms.

Be careful that your GPU isn't a solution looking for a problem,

The major problem in going to gpu processing is the cost involved to hire someone to make the app and then stay around to tweak it so every brand and model of gpu can be used. The more gpu's that can be used the better, no sense writing an app that can only use gpu's that have at least 12gb of onboard memory if less than .01% of crunchers have them. GOOD programmers are not cheap and with most projects being run on a shoe string budget, very few actually get grants, meaning few projects can afford them. Also programmers do not usually go from project to project to offer their services so each project has to find a new one, meaning alot of on the job training. The gpu makers mostly stopped helping Boinc projects as we are not their primary market, gamers are followed by super computers usually built by governments, again meaning alot of on the job training for someone offering their services which can mean a long lag time to get it right. That lag time means it's not a 'I'll pay you 10K to build me a gpu app' because it can't be built and tweaked in a couple of weeks again making the costs prohibitive.