In this work we focus on approximate parallel inference in loopy graphical models using loopy belief propagation. We demonstrate that the natural, fully synchronous parallelization of belief propagation is highly inefficient. By bounding the achievable parallel performance of loopy belief propagation on chain graphical models we develop a theoretical understanding of the parallel limitations of belief propagation. We then introduce Splash belief propagation, a parallel asynchronous approach which achieves the optimal bounds and demonstrates linear to super-linear scaling on large graphical models. Finally we discuss how these ideas may be generalized to parallel iterative graph algorithms in the context of our new GraphLab framework.