They're not 100% comparable because of VLIW4 vs. VLIW5, but you get the point. If the die is any significant amount bigger, power consumption will go up.
The real world power consumption will probably go up, but so will the performance.
HD7870 peaks at
115W = 212mm^2
GTX680 peaks at 186W = 294mm^2
http://www.techpowerup.com/reviews/ASUS/GeForce_GTX_660_Direct_Cu_II/25.html
HD8870 with real world power consumption of 160W with performance ~ 1Ghz HD7970/GTX680 is not unreasonable a 12 months more mature 28nm node with a 280mm^2 die size.
Compared to the unrealistic claims of a 2880 SP 1Ghz GTX780 under 250W TDP, this rumor is actually viable.
Right...
The HD 5870 consumes on average ~145W in games. The HD 7970 consumes on average ~190W in games.
You cannot compare HD7970 to HD8870 though (and this is why you cannot compare Pitcairn XT to Tahiti XT directly).
HD7870 = 1280 SP, 32 ROPs, 80 TMUs, 256-bit bus = 212mm^2 die (13,200 transistors / 1 mm^2 density)
HD7970 = 2048 SP, 32 ROPs, 128 TMUs, 348-bit bus = 365mm^2 die (11,800 transistors / 1 mm^2 density)
Notice, TMUs/SPs increased 60%, and bus width just 50%, # of ROPs stayed the same, but the chip increased 72%!
Clearly Tahiti XT is less dense (or there are some types of less dense transistors within Tahiti XT chip that take up disproportional volume). My guess is the "less" efficient per/mm^2 transistors are what's needed to give you over 1 Tflop of double precision compute (we have seen AMD being very strict on this aspect and chopping off double precision compute or neutering it for mid-range parts --> HD6850/6870). That means double precision compute cannot be cheap transistor space.
Get rid of extra transistors that are needed for double precision compute for Tahiti XT (and all the other 'fat') and you end up with a 270-280mm^2 die 1792 SP, 32 ROP, 112 TMU 256-bit bus chip using less power than an HD7970.
Also, your logic contradicts any other viable possibilities for improving performance right now without growing the die size of HD8800 series. AMD can't add 20-30% increase in IPC since that's impossible in 12 months. That leaves you with 2 major options:
1) Increase clock speeds (this requires higher voltages to maintain 1200-1300mhz) - not going to happen on 28nm from the factory.
2) Make the chip larger by adding more functional units.
In other words, there is no other viable option but to make HD8800 series larger in size if you want to gain performance by 25-30%.
Right now AMD HD7970 can already be found for
$379.99, while HD7950 goes for as low as
$279.99. Both of those are 384-bit 365mm^2 die chips, with 3GB of VRAM.
AMD would make a lot more $ if they discontinued HD7950/7970 and sold an HD8850/8870 2GB 256-bit at $299-349 on a 280mm^2 die with performance ~ HD7950 V2 / HD7970 (or higher). Without the need for a 384-bit bus and double precision compute of Tahiti, AMD would realize substantial die size savings.
Since Tahiti XT is Pixel fill-rate limited, HD8870 @ 1100mhz with 32 ROPs would have 19% higher pixel fill-rate than the original HD7970. That will be huge for gaming performance (we have seen that HD7950 is barely faster than HD7870 at stock speeds despite a huge memory bandwidth advantage. The minute you start overclocking the 7950, the performance increases significantly - not so with the HD7870).
The specs are possible but it will make a much bigger die than the current HD7870 (Pitcairn). If the specs are legitimate, it could make the Tahitis replacement be more than 400mm2.
Ya, last rumor I read was @ 400-410mm^2 for 8970.
