I've wanted to make my own beowulf cluster for a very long time. But on a personal note, I've never had an application for it - the hak5 guys do obviously with the video encoding, so thats not a problem. For me folding@home was never enough incentive to forke out 1500 or so Euro on such a project.
I dont know what anyone else is thinking regarding the processors etc, but my usual thoughts were:
* calculate the best value per mhz (since FLOPS isnt an available metric, at least nowhere i could find) available, taking account multiple cores.
* weigh the value-per-mhz value against the amount of power the processor consumes; afaik the highest cost of ownership of supercomputers/clusters and the like is the electricity bill, not the initial outlay.
I think its important to find some real hard info on what the advantages are of multicore over single core architectures. They're not linear for sure (4 cores vs 1, is definitely not 4x the power, maybe 30pc if your lucky). Im basically proposing taking a very scientific approach to this...
I think the end result is finding the cheapest, best value processor with the highest core count... which at the moment is the Intel i7 or possibly Core 2 Quad if memory serves me.
I think although the mini-itx idea seems sound, your cost-to-performace ratio would be taking a massive hit in exchange for size. But it should take alot less power, so... I'm not sure. Its a complex issue that needs alot of discussion i think :P