Designing multidimensional radiofrequency pulses for clinical application must take into account the local specific absorption rate (SAR) as controlling the global SAR does not guarantee suppression of hot spots. The maximum peak SAR, averaged over an N grams cube (local NgSAR), must be kept under certain safety limits. Computing the SAR over a three-dimensional domain can require several minutes and implementing this computation in a radiofrequency pulse design algorithm could slow down prohibitively the numerical process. In this article, a fast optimization algorithm is designed acting on a limited number of control points, which are strategically selected locations from the entire domain. The selection is performed by comparing the largest eigenvalues and the corresponding eigenvectors of the matrices which locally describe the tissue's amount of heating. The computation complexity is dramatically reduced. An additional critical step to accelerate the computations is to apply a multi shift conjugate gradient algorithm. Two transmit array setups are studied: a two channel 3 T birdcage body coil and a 12-channel 7 T transverse electromagnetic (TEM) head coil. In comparison with minimum power radiofrequency pulses, it is shown that a reduction of 36.5% and 35%, respectively, in the local NgSAR can be achieved within short, clinically feasible, computation times. Magn Reson Med, 2012. © 2011 Wiley Periodicals, Inc.