2. Radeon HD 7970:
- Radeon cores: 2048
- Core/GPU frequency: 1GHz
- Effective Memory frequency: 8000MHz
- Type of Memory: XDR2
- Texture Units (TU): 128
- Raster Operation Units (ROP): 64
- Bus width: 256-bit
- Memory bandwidth: 256GB/s
- TDP: 190W
Source
And yes, this is a confirmed leak. AMD wants better compute capability, but doesn't want to give up efficiency for it. Judging from the specs, Tahiti will probably be smaller than Cayman.
1. 2048 x 1000mhz vs. 1536 x 880mhz = 51.5% greater pixel fill-rate
2. 128 TMUs x 1000 mhz vs. 96 TMUs x 880mhz = 51.5% greater texture fill-rate
3. 256 GB/sec vs. 176 GB/sec = 45% more bandwidth.
Either these specs are fake, each SP/TMU/ROP is way more powerful in GCN, or HD7970 will get clobbered by Kepler. A card that has 45-51% faster specs on paper on average will at best be 50% faster than HD6970, which will put it only 30% faster than a GTX580 (100% = HD6970, 115% = GTX580, 150% = HD7970).
I am going to say these specs are too conservative or AMD has another 20% or so increase in efficiency hiding in redesigned TMUs and SPs or HD6970 was ROP starved?