What I don't understand about the NV30 and its' "finesse" is that it only integrates four FP ALU's (which have to be shared with the texture units). The R3XX on the other hand has 24 (8 for texture coords). Even if NV30 had 100 percent computational efficency (ie every cycle an instruction is issued on every ALU) it still couldn't approach the real world performance of ATi in PS2.0+.