What's your source for saying that the 7870 will be CGN? I haven't seen anything other than it will be a 28nm die shrink of the VLIW4 architecture with 1536 shaders, same as a 6970, which gets about 390mh/s. You usually get a small bump in speed with the die shrink, so you should get at least 390 Mh/s, probably a bit more.
There are a lot of them, some in this thread. The early reports seemed to indicate a die shrink. The later ones now indicate only the low end OEM garbage will be a die shrink rebadge. I agree with you (and had stated it in the past) IF the 7850 is a VLIW4 on 28nm it will be a monster card. Sadly that doesn't seem to be the case.
As for the 5970, I have quite a lot of them running in my data center. Looking at the console output right now, my slightly overclocked (750MHz/300MHz) 5970's, using phatk2 and the usual phoenix miner optimizations are getting 340 mh/s per core or 640 mh/s total. I haven't been able to get 375 per core without increasing voltage (and good cooling). Thus, I doubt you're getting 750mh/s speed under 300W. I can hook up one of my rigs to 120v with a kill-a-watt and check the next time I'm at the data center.
I have to break it to you but those number are pure crap for a 5970. Not just against what I have gotten but against what lots of other people have gotten. The limit on each card varies but my slowest card is running 820Mhz/160Mhz. 3x5970 in open air rig pulls 870W at the wall (measured over 3 days using kill-a-watt). Backing out 100W for the MB, power supply inefficiency, CPU, RAM, etc that puts each card @ roughly 250W a piece. That is w/ cgminer 2.1.0 running on Linux w/ 11.6 drivers & 2.4 SDK.
If you aren't getting AT LEAST 700MH/s you are doing something horribly wrong. Note: 700MH/s isn't "good" it is just barely acceptable.
