We describe different implementations of the 3D Heisenberg spin glass model for Graphics Processing Units (GPU). The results show that the {\em fast} shared memory gives better performance with respect to the {\em slow} global memory only if a multi-hit technique is used.
@article{arxiv.1006.2566,
title = {The Heisenberg spin glass model on GPU: myths and actual facts},
author = {M. Bernaschi and G. Parisi and L. Parisi},
journal= {arXiv preprint arXiv:1006.2566},
year = {2010}
}