RLS (recursive least squares) converges much faster than a gradient descent filter. Gradient descent algorithms converge to the minimum MSE at a rate proportional to that of the eigenvalue spread of the channel autocorrelation matrix (high eigenvalue spread means the ratio of the maximum eigenvalue to minimum eigenvalue is high, and the gradient descent algorithms like LMS will converge slowly). RLS adjusts the taps of the adaptive filter such that convergence is much faster, but requires more computational complexity. Also being recursive, you may run into problems with divergence and fixed-point error propagation. So you would use gradient descent when you don't care about convergence speed (i.e. stationary channel or slowly-varying channel distortion) and when computational simplicity is required.