ShadesOfMarble does have a point though, IMO. Asics (and FPGAs) are sort of in between hardware and software. The GPU (the core itself) is basically an ASIC. With lots of functionality that's not required for mining.
I think this misses the reason that mining specific ASIC are efficient. The GPU core is a parallel vector compute engine, in effect a specialised CPU. I'm not sure that there is a big overhead on a graphics card (above a few $$ driver chips) which could be saved. The power benefit of a SHA or SCRYPT ASIC comes from the fact that they are not programmable. The single function that they compute is hard-wired (hence it is impossible to re-target them effectively). The same reasoning is behind ASIC being more efficient than FPGA (although the mechanism is different). Silicon resource that is not used 100% every cycle is a cost - but is needed if you need to support configurable algorithms.
An algorithm which pushes the energy cost almost completely to pushing data through RAM will further reduce the benefit to be made by doing the compute side efficiently.