Many kinds of statistical modeling are naturally suited for GPU computing owing to their inherent parallelism. For example, in Monte Carlo simulation there are usually many independent simulations to be done, and each can be performed by a separate thread. The challenges lie in coping with unusual features that could lead to warp divergence or poor memory coalescence, or require a global reduction operation. Nevertheless, this is an area with rapid adoption of GPU technology within universities.

Banks and other financial organizations have also been actively investigating the potential of GPUs. Computational finance has been a major HPC growth area in the past 20 years, and with the development of commercial random number-generation software for GPUs, I think we are likely to see widespread use of GPUs in the near future.

Chapter 16 , written by Thomas Bradley, Mike Giles, Jacques Du Toit, Robert Tong, and Paul Woodhams, concerns techniques for parallel random number generation. Three of the most widely used generators are implemented, and interestingly, the different mathematical features of the three generators lead to different approaches for efficient GPU execution.

Chapter 17 , written by Laszlo Szirmay-Kalos, Balazs Toth, and Milan Magdics, discusses a physical application, the Monte Carlo simulation of photon transport in an inhomogeneous medium. The implementation has to be structured carefully to avoid excessive warp divergence owing to the large variation in the lifetime of different photons. The efficiency is also enhanced by using a coarse grid to represent the inhomogeneous properties of the medium.

Chapter 18 , written by Christoph Schied, Johannes Hanika, Holger Dammertz, and Hendrik Lensch, deals with an application from computer graphics using random functions to generate fractal images. Here, the key idea is the use of presorting to eliminate branch divergence.