Dumb question: does doubling all the hardware components (SPs, TMUs, ROPs, Bus width etc.) while keeping the same clock-speed (for both core and memory) result in a greater than double TDP?
If not, then 2x Bonnaire and 3x Bonnaire @ 1 GHz Core, 6 GHz VRAM speeds should result in no more than 170W and 265W TDPs for presumed mid-high and high-end/enthusiast class cards (1792/112/32/256 bit, 2688/168/48/384 bit), which should be competitive with the 760 and 780 if priced right.
Further sacrificing Power for higher core (1.1-1.2 GHz) and VRAM (7 GHz) clock speeds (190 W, 280 W) should then net a decent performance boost where such hypothetical cards would/should compete with 770/Titan on performance.
Or AMD could choose to put more transistors (3072/192/64/512 bit massive die, <1GHz and <7GHz core and memory speeds) in a bid to match both performance and TDP. In which case we may possibly see even Tahiti refreshed for lower TDP, higher clocks (1.1 GHz core, 7 GHz VRAM, 2048/128/32-48/384 bit) for high-end (greater than 770) performance. This is of course all hypothetical talk just for the sake of discussion.
However, if my premise is not false, logically, 2x and 3x Bonnaire would make a lot of sense. Consider a 7950 like GPU with a core speed of 1100-1200 MHz, VRAM @ 7 GHz but on 256-bit bus, should be great for 1080p (if priced around 250$). Should beat 760 quite well and give 770 a run for the money.
A 2688 shader part with 48 ROPs and 168 TMUs @ 1 GHz should have ~ 31-32% more shading and texturing throughput than the 7970 GE, and 50% more ROP throughput. From what I have read, Tahiti was ROP limited so this should on average net about a 35% increase over the 7970 GE, putting it close to Titan.
Adding in hardware frame metering for CF and dedicating some transistors for better Boost would also be nice supplementary features.
Again the assumption here is that increase in hardware constituents lead to no more than proportional increase in TDP at same clock speeds.