As the Givens' rotations are applied to the tridiagonal matrix, they are also applied to a matrix in which eigenvectors are accumulated. While one Implicit Francis QR Step requires $O( n )$ computation for chasing the bulge, this accumulation of the eigenvectors requires $O( n^2 )$ computation with $O( n^2 )$ data per step. This inherently means the cost of accessing data dominates on current architectures.