I wanted to compare this to more sane options so first some flops:
1x RPi (Rasberry Pi) CPU: 175MFlops GPU: 26GFlops for 1.5W
64x RPi total floppage: (26+0.175)*64 = 1675 GFlops for 96Watts
giving an efficiency rating of 17.45 GFlops/W
in comparison an AMD 7970 (GHz edition) has slightly more than 4TFlops (4000GFlops) for 225 W
giving an efficiency rating of 17.78 GFlops/W
Remarkably similar. These numbers would shift with system power considerations and undervolting potential though so it is unclear which would win an efficiency war.
It would be more practical to build a GPU based supercomputer (4x 7970 at about $3k giving 16TFlops, so 9x cheaper per performance too!), which also has more mature software pipelines, but I applaud the different approach taken here and we might see some very novel applications of this super-shrinking