A GPU/CUDA Based Monte Carlo Code for Proton Transport: Preliminary Results of Proton Depth Dose in Water
L Su*, T Liu, A Ding, X Xu, Rensselaer Polytechnic Inst., Troy, NYWE-C-BRB-8 Wednesday 10:30:00 AM - 12:30:00 PM Room: Ballroom B
Purpose: Although several studies have reported the use of GPUs to accelerate Monte Carlo calculations for x-ray imaging and treatment planning, there is little effort to demonstrate the utility of this highly parallel yet affordable computing tool for proton treatment planning and dose verification. This paper describes a preliminary project to design a GPU/CUDA based proton transport MC code and to evaluate the timing for proton dose depth distribution.
Methods: The proton transport in media was modeled by condensed history method, in which the effect of many interactions was grouped into single condensed step. The Moliere distribution and Valilov distribution were employed to calculate angular deflection and energy loss. The CPU code was written in C++ and GPU-based code was developed in CUDA C 4.0. The hardware platform was a desktop with Intel Xeon X5660 CPU and NVIDIA Tesla™ m²090 GPU. Nuclear interactions were not included in the preliminary study and the transport medium was limited to water.
Results: The depth dose distributions of proton of different energies were simulated. It was found that 98% of the tallies had relative error less than 1%. The code was benchmarked against MCNPX and GEANT4 codes. For 200 MeV proton pencil beam incident on the water phantom, the dose difference between our code and GEANT4 (nuclear interaction disabled) was within 2% for 95% of all depths. The speedup factor of our GPU code over CPU code was x57. While compared with MCNPX (nuclear interaction on), the GPU code was x620 times faster.
Conclusions: This is one of the first reported efforts to demonstrate a GPU/CUDA-based proton transport MC code for dose calculations. Despite some limitations, this preliminary project was able to show significant gains in the GPU computing time, thus suggesting a promising role of such Monte Carlo tools in the future.