This paper describes three parallel strategies for Ward's algorithm with OpenMP or/and CUDA. Faced with the difficulty of a priori modelling of elicited brain responses by a complex paradigm in fMRI experiments, data-driven analysis have been extensively applied to fMRI data. A promising approach is clustering data which does not make stringent assumptions such as spatial independence of sources. Thirion et al. have shown that hierarchical agglomerative clustering (HAC) with Ward's minimum variance criterion is a method of choice. However, HAC is computationally demanding, especially for distance computation. With our strategy, for single subject analysis, a speed-up of up to 7 was achieved on a workstation. For group analysis (concatenation of several subjects), a speed-up of up to 20 was achieved on a workstation.