optimized clients

Optimized BOINC clients are nice for the ones searching adventure, but are not what the doctor ordered to speed up the actual calculations. E@H could use massive amounts of computation power, so any substantial speedup should be welcome. Hand optimized clients for different CPUs will get really messy to work with, but a *simple* recompile should be possible as soon as it's obvious, that the client will stay the same for some time. Some effort has to be put into an oberservation, wether the new version is really faster AND produces a correct result.

What I suggest is the following: after the dust of going public has settled, when everything is running smooth and the programmers are back from their holidays they should try out a recompile using SSE1/2/3 optimizations for windows, with SSE2 for win being the most important. (I assume that until now E@H doesn't use these optimizations) The intel compiler should be fine for this, or maybe one specialized on scientific applications. If this works out well (ne bugs, noticeably speed-up) other platforms like linux and Macs could be considered. While doing that may annoy users of more exotic systems it would certainly increase the overall computation power available to the project.

Another "optimization" which comes to my mind is using a GeForce 6x00 series or better. OK, this is where it gets really complicated and specific, but again the potential speed-up should be enormous. That's assuming E@H uses FFTs or other algorithms that are well suited for programmable graphics cards. Programming this should involve prepairing data by the cpu, than giving data packages to the GFX card, returning processed data to the cpu, doing post processing with the cpu and keeping up a pipelined flow between the two. I certainly don't want to do this, but again, if the core routines are constant the work has to be done only once. That may be well worth the effort.

Am I missing any important points / problems here?
Regards, Stephan