Uhh, that's not the article, the article is running a ml model, on phone and floating point opps for tensor multiplication seems to be off.