Several things to point out, yes the cut down version is alost 2/3's of everything isn't it, infact its odd that L1 cache says its 3/4 of the 380, but l2 is 2/3rds of the 280. Considering each cluster should have l1/l2 together, 2/3 the shader clusters should have 2/3rds the l1 cache, not 3/4.
Either way, those are the specs, bandwidth/texel wise IF they hit those clocks.
Considering they don't have A3 silicon yet and can not possibly have it for another 2-3 weeks I believe, maybe 4 weeks those aren't in any way possibly final cards, its literally impossible.
Those could be a couple A2 silicon versions, in which case they are not running those clocks unless heavily overvolted/cooled/heated/etc and silly power consumption.
If the final clocks don't hit their targets, those specs go out the window, though the one thing I think those numbers could be is giving people a shout about the limited ability of the lower end version and how it will be significantly slower than the 380 version. I think they are lowering expectations because they will probably have terrible yields and require a pretty heavily cut down version to be able to salvage as many cores as possible. The fact the bar seems set sooooo low for the salvaged parts makes me think the 380 will be rare as hell and priced through the roof.
This is all ignoring the underlying architecture and how it does things, is tesselation done in software only, will it be ridiculously slow in games that offer teselation and will the majority of games out in the next year use it, who the heck knows to be honest. There are other DX11 things it might only be doing in software going by the rumours, and realistically even if it rapes benchies of older games, we might not know really how good it is till midway through next year and we have 3-4 more DX11 titles.