A runtime/memory trade-off of the continous Ziggurat method on GPUs

Pseudo random number generators are intensively used in many computational applications, e.g. the treatment of Uncertainty Quantification problems. For this reason, the optimization of such generators for various hardware architectures is of big interest. We present a runtime/memory trade-off for the popular Ziggurat method with focus on GPUs. Such a trade-off means that the runtime of pseudo random number generation can be reduced by investing more memory and vice versa. Especially GPUs benefit from this approach since it reduces warp divergence which occurs for rejection methods such as the Ziggurat method. To our knowledge, such a trade-off for the Ziggurat method has never been investigated before for GPUs. It is shown that this approach makes the Ziggurat method competitive against well established normal pseudo random number generators on GPUs. Optimal implementations and grid configurations are given for different GPU architectures.