Yes, You are right, It is a major area of research about optimizing the number of cores as compared with the performance. Turning on and off the cores as an when required dynamically is the present day solution while using more cores. Lets see if a major power improvement comes in the mean time by ARM.
Peter, your point on multiple small cores versus a single larger one has power management implications as well. It gives the potion of shutting down cores to save power, as opposed to (or maybe in addition to) clock rate management. Do the new CPUs support both of these power management strategies? Did ARM discuss any improvements that they might have coming in terms of power management?
The analysis of power consumption is tricky and should take into account voltage-frequency scaling.
I believe the assertion "With a Cortex-A15, at five times the area and five times the power consumption, you can get two or three times the performance of the Cortex-A7" is a statement about what happens when the A15 and A7 are run at the same frequency and voltage. If we look at running the two processors at the same performance point then the frequency of the A15 can be one-half to one-third of the A7; that means that the A15 power will be 2.5x to 1.7x the A7 power at the same voltage. But the A15 design is being run slow - so the voltage can be dropped. Taking the figures shown on the ST-Ericssson blog* for 28FDSOI as guidance, dropping from 1.0V to 0.7V drops frequency by 2x. If we drop the voltage then the power consumption drop by a further 0.7^2 = 50% - so for applications where the A15 is 2x the performance of an A7 at the same frequency, the power consumption is 25% higher than A7, for those where A15 is 3x the performance of A7 the power consumption is at at least 15% better than A15.
For sure I've ignored leakage but its likely leakage is only really an issue at high-performance points which cannot be reached by an A7 anyway.
Finally, it is also the case that to first order, from a power perspective, you woul dbe better using two half-fequency processors to do a job than 1 full-frequency. It's just there is a little problem with software.