Performance Study of GPU- and FPGA-Based 3-D Monte Carlo Computation Used in Dynamic Radiation Topotherapy

The development of cluster and parallel computers has increased the popularity of implementing the Monte Carlo computational method in medical and scientific applications. However, having access to these high-performance systems is not easy and a reasonable size system is not yet able to solve many popular Monte Carlo problems near real-time. The Field Programmable Gate Arrays (FPGAs) have been proved to be useful in speeding up the Monte Carlo algorithm since they are able to execute the problem with a high degree of parallelism. In recent years, the Graphic Processing Units (GPUs) have been proved to be useful for many general-purpose calculations due to their parallel and pipelined architecture, which is able to accelerate various algorithms. Since Monte Carlo simulations require intensive computations, GPU architecture can also deliver great performance for Monte Carlo simulations similarly to the FPGA. This paper explores performance characteristics of the FPGA and GPU in computational intensive 3-D Monte Carlo applications. The study is part of our current research on the study of dynamic radiation topotherapy with stationary sweeping fan beam and translational moving microbeam on stationary treatment couch. We performed our initial study with photon interactions on 3-D water phantoms and the initial results (without intensive performance tuning) showed that performance of the 16-core NVidia GeForce 8400 GS GPU was about 2 times better than single-core 3-GHz Intel Xeon and about 4 times slower than the Xilinx Virtex-5 FPGA. We learned that out-of-core memory contention is an important performance drawback. Although we performed our study on photon transport, we expected the results to be similar for charged-particle transport (our interest in dynamic topotherapy) since only core Monte Carlo calculations were implemented on the GPU and the FPGA.