I'm looking to build a low-end computational server (my jargon in this field is especially limited so if someone can state that better please change that to meet jargon). I'm basically running computational fluid dynamics programs, large matrix computations and bioinformatics code. What would be the best way to approach cost/benefit analysis on what to put in the system? Perhaps even more general: How does one approach cost/benefit analysis on hardware theoretically (doing the analysis before building the machine)?
For your kind of work-load, you may want to consider GPGPU computing instead. NVIDIA certainly seems to have the best supported libraries for such things - CUDA. Unfortunately, Sony recently neutered the PS3, otherwise it would have been a cost-effective beast for such things.