The take away from my earlier post should be a really top quality motherboard and Power Supply is always a great investment; the rest of the gear should be selected to meet your needs.
However realise that coding queue based, multi-threaded software isn't all that hard once you understand the basics (my thesis in the early 80s was on coding frameworks for parallelism and language design...) - so 35 years on hopefully this practice will one day catch on once all the frameworks are set up so even semi skilled idiots can do it
... So maybe in a few years it will be really, really common - as multi cored machine parallelism isn't going away.
To really leverage GPUs you need problems that 1) don't need masive amounts of data transferred to a dedicated hardware unit for just a short period of time and 2) benefit from mass parallelism by being able to be broken down easily into a set of much smaller, individual problems to solve.
DirectCompute (Or OpenCL) is taking off slowly - as the frameworks do exist but are a lot harder to get the hang of. Once this is done - mass parallelism is in everyones hands. You are correct that a CUDA or STEAM core isn't as generally powerful as a RISC core - but its access to shared, high speed memory and ability to do specific transforms really fast, in a massively parallel way is stunning once it kicks in. If our astronomy programs were translated to shift load across a modern CPU / GPU platform - you would simply have sub-second responses for all operations - one day maybe.
One day a bright Uni student may come across a simple way to re-program on the fly from procedural languages to GPU languages - then over night your programs could be re-compiled to run on this hardware (not optimally maybe - that's harder) but it would run a lot faster. I saw this in the late 90s when one student wrote a translater from Fortran to C and recompiled Cayley (a 10,000 page Fortran program designed to do "infinite" precision maths on groups of over 10 ^ 60 elements - think ~ number of atoms in the visible universe - and got it all working over one weekend). Imagine if someone did that for general compute algorithms - game changer... (
http://research.ijcaonline.org/volum...pxc3892109.pdf or
http://au.mathworks.com/company/news...mathworks.com#) - one day!