Basically, you will never get a linear increase with parallel computing, especially on consumer type apps that are much more synchronous. So running multiple cores, with requisite context switches will get less computing per watt, even if it isn't as bad as the 2x number.
His definition of "superlinear" is a bit tricky. As he explains, you can only get superlinear by "Do disproportionately less work." and/or "Harness disproportionately more resources.". However, his example of the former is actually a subtle instance of the latter.
Superlinear speedups are always due to more effective caching. Sutter uses "disproportionatly more resources" when the added effectiveness comes from larger caches. However, his "do less work" examples are "just" better cache behavior.
Can you explain this part?
The first time I read this I didn't understand it, but the author has laid a trap here for the unwary reader in the word "doubling". If you can get 4x the performance using 4 cores the energy efficiency is the same or better, but if your program scales poorly and only gets 2x performance from 4 cores the energy efficiency is worse.