Unfortunately you can't really go by shaders per dollar. For example the 390 is only a few Mh behind the Fury X even though theoretically, the FuryX should be around 50% faster. This is likely just an optimization problem but that's how it stands.
The 4GB 380 was the best value card up until last week but prices prices crept up enough where the 380x is probably now the ideal card if you want to get your feet wet. I still think if you're going to spend the money the 390's make the most sense to buy. Sure they'll take a little longer to get your ROI, but after that it'll earn at a 40% faster rate and hold better resale value. Fewer cards to manage as well.
It'll take six 380's to match the speed of four 390's. The added expense for connectors etc is not really worth it if you're trying to scale to higher speeds.
I'm at 225Mh now but stopping here unless I can figure out a way to get cheaper power. My cards should all provide a ROI well before Polaris hits. Luckily I had enough other equipment where I only had to buy a few connectors and cards. If you have to buy everything it gets expensive fast.
That's really interesting on the Fury's. I wonder why they perform so poorly? The other GCN chips (Pitcairn, Tahiti, Tonga and Hawaii) all seem to track pretty much right on with shader count. Perhaps there's something in the algorithm that just really doesn't like the 500Mhz memory bus.