From a MH/W basis it is hard to beat 5970.

My rigs pull about 250W per card (memory underclocked to 160Mhz).  
I run 3x5970 per rig.  
Using usb key for OS, underclocked sempron, and MB w/ everything unecessary turned off the entire system pulls ~860W at the wall.

Rigs get ~2.2GH to 2.3GH on 860W or ~2.5 to 2.7 MH/W.

I don't think it is possible to do much better than that unless you go FPGA or wait for 7800 series cards.