Here's a Google translated version of the specs:
As the Shader Multiprocessors (SMs) are built is unclear. Supposedly 512 ALUs are set, but not in the form of a full GF100, but by a restructuring of the cluster: GF100 offered per SM for the 32 ALUs and 4 TMUs (8:1), but GF108/106/104 48 ALUs and 8 TMUs (6 1.) GF110 but shall receive more ROPs more TMUs, more 128, and thus twice that of a (not available) GF100 with all 16 SMs. This should rebuild the Nvidia SMs again, so that ALU 32 / 8 or 16 TMUs ALUs / 4 TMUs work - so a 4:1 ratio would be given almost as GT200. The fact that Nvidia throws like the GF10x for players redundant parts of the chip pile decreases the die area, which is filled by the additional ROPs and TMUs again. Nevertheless, the current consumption as GF100 and are eventually rise to Taktfreudigkeit. The question is how strong the power only through more ROPs / TMUs up fast when the arithmetic-Power is hardly dressed. The comparison of a GeForce GTX 460/1G and a Geforce GTX 470 shows that the latter itself a highly overclocked GTX 460 is often ahead, although the GTX 460 ascends considerably more texturing power per arithmetic power. Nevertheless, GF100, through its lower part TMU count as a disadvantage. Quite apart from the question arises, how many Polymorphengines GPC and the GF110 ascends, Nvidia would hardly be the Tessellationleistung a Geforce GTX undercut 480th