You can’t extrapolate size of model from speed that way. Architecture difference, load etc will screw up the approximation