I'd say there are 3 ways to go about it.
1) High quantity of CPU's. They really need to be i5/i7 Sandy or Ivy Bridge and overclocked. Here you can get a mid range motherboard, only real important part is the VRM circuit on the board.
2) Server style system to get bigadv work units. Basically at least a board with dual CPU sockets, and I would be targeting 24 core/threads for headroom. AMD is way more cost effective. Only older Intel Xeon CPU's can be overclocked(SR2 motherboard or such as well). Heatsinks are tricky, For AMD there is a supermicro board that can be overclocked through linux.
3) GPU system. Save some space by putting 3 or 4 GPU's in one system. Supplying power and cooling(cards are close together/low air flow) is the main issue. Don't need to have the latest hardware either(460GTX still producing good and can be bought used pretty cheap, especially the lower RAM version card, not so good for gaming but good for folding). Nvidia is the only choice. Keep an eye on the news of the 600 series, for when they get some work units for it.
-Some of the other tricks are running Linux right now as it is currently quicker then windows.
-I also recomplied the kernel added the BFS scheduler, seemed just a tad quicker on frame times.
-If you do the bigadv route someday and use a newer linux version that uses ext4 filesystem by deafult it's probably best to at least have a partition of ext3 for the folding directory.
-For CPU's keep a close eye on running processes. For example in windows if you have a tray full of icons you are often hurting your frame times. To move on to the next calculation all the threads need to be completed. So if you have a process using a thread it is slowing down all of them. An example would be a 4 thread CPU, think of it as 4 of us changing 4 wheels on a car. If one of us takes off to change the spark plugs, the rest of us are done before that one guy and we can't move the car out of the garage and on to the next one until all of us are done.
-Of course 24/7 is a key, it's not practical for everyone, nor is not using any other process but it does help folding which is where the dedicated rig comes in.
-GPU and CPU in same system. This goes back to the running process thing. The GPU will steal cycles from the CPU and slow the CPU folding down. Optimally you would want the GPU or GPU's in a separate system. Not economically viable but it helps maximize PPD.
-Bigadv is a lot of points. As of a week ago it requires 16 thread/cores. There is a way to trick linux into thinking its got more cores/threads then you really have. Stanford is aware of this and has said it's not cheating, but is frowned upon. I don't know if it's worth it now with a 6 core 12 thread Intel CPU. You could easily do the 6905/6 WU's, but 8000 ones would be awfully close and you'd have to ask if you are helping the science running so close to deadline. Would need a minimum of around 4.2Ghz overclock on a 6 core/12 thread Intel to give yourself ~6+ hours to make deadline on the 8000's
I'd say with all the tweaks I squeaked out an extra 10,000PPD the last couple months.
Now that I rambled on and on an on...I was using the CPU quantity method
3770K at 4.8GHz
2600K at 4.5Ghz
1090T at 3.6Ghz
Q6600 at 3.2Ghz
5 various i5's/Core2's at stock.
460GTX at 900Mhz
570GTX at 850Mhz
240GT at stock
One other thing I just felt like adding. I like the Asrock motherboards. They have 775 mounting holes so you can reuse any older style heatsinks you have lying around that are better then Intel stock.
