[arch]

those machines are not silver bullet - that is, there are build for specific task. that task may eat a lot of cpu but less memory and little need to speak to other nodes over the network. it can be memory hungry and average use of cpu.. and so on and so on. to build/bуy such thing as super computer first you had to have problem to solve and you have to be fluent on that problem on paper. that is complete understanding of the problem to be solve and later build calculating unit to solve this problem. you can use nvidia;s cuda for massive parallel calculations - of course it has it's own limitations. you can build cluster from of the shelf hardware, with 100/1000mbit network. or go wit blade server and inifiband switch.so.. get to now math - it's sound pathetic but you need it to solve any problem you have. then make yourself competent enough in computer hardware - there is more than one architecture - i386 or AMD64 and take right decision about your super computer architecture. get to know programming - c/c++, fortran .. etc.. languages.
aaaaaaaaaaaand in the end - formulate your problem, and take appropriate approach in solving it, at lower as possible cost for you as money and time.