Parallel implementation of a Quantization algorithm for pricing American style options on GPGPU

The Quantization Tree algorithm has proven to be quite an efficient tool for the evaluation of financial derivatives with non-vanilla exercise rights as American-, Bermudan-or Swing options. Nevertheless, it relies heavily on a fast computation of the transition probabilities in the underlying Quantization Tree. Since this estimation is typically done by Monte-Carlo simulations, it is appealing to take advantage of the massive parallel computing capabilities of modern GPGPU-devices. We present in this article a parallel implementation of the transition probability estimation for a Gaussian 2-factor model in CUDA. Since we have to deal in this case with a huge amount of data and quite long MC-paths, it turned out that the naive pathwise parallel implementation is not optimal. We therefore present a time-layer wise parallelization which can better exploit the parallel computing power of GPGPU-devices by using faster memory structures.