Over the last years, high performance computing has become a key step to introduce computer tools in the medical field, especially in image guided surgery and therapy (IGT). Among these tools, there is a special need for real time registration algorithms. By real time, we mean faster than the image acquisition to add an acceptable overhead (say under 1mn for IGT applications). To tend toward such small registration times, one usually simplify and adapt algorithms so that they become application and data specific. This process involves a lot of designing and programming work for each application, and reduces the generality and robustness of the method. Our goal in this paper is to show that a general registration algorithm (here the block-matching scheme of \citeOurselin:miccai:00) can be parallelised on a cheap and standard parallel architecture with a reasonably small amount of additional programming work, thus keeping intact the algorithm performances and generality. For medical applications, we show that a cheap cluster of bi-processor PCs connected by an Ethernet network is a good trade-off between the power and the cost of the parallel platform. Portability, scalability and safety requirements led us to choose OpenMP to program multi-processor machines and MPI to coordinate the different nodes of the cluster. The resulting computation times are very good on small and medium resolution images (resp. 19 seconds and 45 seconds on 5 bi-processors), and they are still acceptable on high resolution MR images (resp. 1mn35 for 5 bi-pro and 70 seconds for 10 bi-pro). One can obtain even smaller times by stopping the algorithm one step before the final level in the multi-scale pyramid, at the cost of a slightly degraded accuracy (relative precision below the voxel size): the registration of high resolution MRI is down to 1 mn on a cluster of only 2 bi-processor PCs.