Abstract
We present an accuracy analysis of the Maximum Likelihood Phylogeny Reconstruction kernel used in PhyML. Historically, PhyML was relying on double precision arithmetic to avoid rounding error as well as to minimize the number of underflow. However double precision arithmetic necessitates more memory and computational resources than single precision. This concerns is particularly important when considering an execution on a multicore architecture such as a GPU. This article provides a numerical analysis of the likelihood kernel executed in PhyML. Based on this analysis, solutions to minimize the impact of rounding error and underflow are proposed.