The Institute of Neuroinformatics at the Eidgenössische Technische Hochschule is developing computational models for hybrid analog/digital semiconductors that implement neuromorphic cognitive systems. (Source: ETH)
Don't have any numbers for silicon retinas, but the long-term goal IBM's cognitive computers is to simulate corelets on a supercomputer that consumes mega-Watts, then execute them on a cognitive computer that consumes kilo-Watts--1000-to-1 less power.
Regarding power dissipation it should be drastically less, because artificial neurons dissipate very little power except when firing voltage spikes, which typically only occurs every few hundred milliseconds.
Regarding using smart silicon retinas in humans, that would probably be a decade or more away, but in the shorter term the fact that cognitive functions can be built-in should make human-like perception possible for robots.