Nope, they actually scale perfectly - as long as the bandwidth is sufficient. The 680 is a double 650 Ti in every regard and clocked about 10% higher. And it has 220-230% of the 650 Ti's performance. Go figure
GTX 650 Ti has 768 CUDA cores with 4 SMX , 2GPC, 128 bit memory , 16 ROPs. GTX 680 has 1536 CUDA cores with 8 SMX ,4 GPC, 256 bit memory and 32 ROPs. almost every aspect is doubled and in fact clocks on GTX 680 are higher. GTX 780 has 2688 stream processors, 14 SMX, 5 GPC, 384 bit memory, 48 ROPs. clocks are expected in the 850 - 900 mhz on gtx 780 compared to 1058 mhz on GTX 680.
With 75% more CUDA cores and SMX and just 25% more GPC the GPC/ SMX ratio has decreased from 1:2 to 1:3 on 4 of the 5 GPCs. only the last GPC with one SMX disabled will have 1:2 ratio. with only a 50% increase in bandwidth and ROPs there too its not a doubling of resources. with so many changes even at same clocks getting a perfectly linear scaling is not going to be possible.
anyway the launch is only a month away so we can get a clue as to how efficient GK110 design is and how well it scales performance with respect to GTX 680.