So either go with FPGA-S multiple of them or cell processors, or combination of them. Maybe there could be some other kind of processors, with large enough amount of registers or easily handleable cache. maybe would cost somewhat less money to build such then buy equivalent? one can calculate a lot of stuff on such, fpga-s should be higher in number, then other speical processors.

beowulf clusters are by far and away the cheapest way to make a supercomputer. They're likely easier too, as I don't know where you can easily get a bunch of FPGAs and cell processors for cheap. Perhaps they'll be more efficient, but this is for hobbyists.

I think you're way above our grade level though, so don't expect much discussion here. There are real tech forums that would be more suited, but I understand that might be difficult since you aren't very good with english.