Over a year ago I posted a few comments about the pricing and power consumption of supercomputers and later I compared that to the power of and power use of the brain.
Now Wired News reports on a new chip providing 25 gflops at 3W power consumption. That translates into an ultralow 0.12MW consumption per petaflop. That figure is misleading since the power consumption of memory systems etc. is not included, but it is still impressive, since e.g. the IBM Blue Gene machine is expected to consume 2MW (which is a comparatively low rate of power consumption as supercomputers go). I don't really know how the power consumption of the usual supercomputer is distributed between the computational core and the memory system.
In terms of number of processors it also represents about a tenfold increace in the power per CPU compared to a regular PC of about a year ago, as that expressed itself in the theoretical computational speed of clusters like The Horseshoe.
We're still a while away from brain level efficiency.