The Google FaceNet and other similar system typically perform an approximate nearest neighbor search in a ~128 dimensional embedding of faces.

Here is the paper for the reference: http://arxiv.org/pdf/1503.03832v3.pdf

Also 70% accuracy is like really really bad.

Here are results from MegaFace Challenge, which involved identifying face in presence of Million distractors:

http://megaface.cs.washington.edu/results/fgnetresults.html

That was with the FGNet dataset... Ntech labs (incidentally the algorithm behind FindFace) beat Googlr on the FaceScrub dataset.

http://megaface.cs.washington.edu/results/facescrubresults.h...

The accuracy is not 70%. The percentage of times it is able to match a face to a social media profile, given the person has a profile, is 70%. You can't compare apples-to-apples with the results in Google's FaceNet.

It's likely using a nearest neighbor or tree search; it might take a billion instructions to process a single photo and another billion to do the nearest neighbor search.

"from" a normal computer. meaning a terminal, connected to the internet.