Multi-GPU Accelerated Ray Tracing Using CUDA

Ray-tracing (RT) is integral to resolving numerical radiation heat transfer processes. The radiation view factor (Fij) must be determined to quantify the heat leaving or being absorbed by a participatory surface. To capitalize on the embarrassingly-parallel nature of determining Fij, graphics processing units (GPUs) can be employed to quickly and accurately resolve Fij values for complex systems. This work outlines the methods and best-practices of implementing multi-GPU accelerated RT in NVIDIA CUDA. Herein, specific Fij values for canonical geometries (parallel planes and concentric spheres) and novel geometries (a thermoelectric generator (TEG) unicouple) are determined and compared to analytic predictions, where applicable, and previously generated numeric values; a multi-GPU accelerated Java-based RT code serves as the basis for numerical comparison. The surfaces of the geometries are represented by tessellations within stereolithography (STL) files, which are created both in binary and American Standard Code for Information Interchange (ASCII) file-type formats. While the Java implementation achieved the same results with less computational time in comparison to RT codes executed on a central processing units (CPU), linear speed-up was not achieved with increasing GPU count. The achievable speed-up was highly dependent on the number of STLs used to represent each geometry, and although non-linear, still exhibited a near order-of-magnitude decrement in computation time in comparison to CPU-base codes. The CUDA-based RT code, combined with a binary STL file-type format, achieved lesser computation time in comparison to the Java-based RT code, and exhibited near-linear speed-up with increasing GPU count for large tessellation systems. Using binary STL file-type formats, less computational overhead was required in comparison to ASCII file-type formats used within the CUDA-based RT codes. It is demonstrated that the CUDA-based RT codes are robust and fast enough to provide benchmark-quality numerical results for three-dimensional systems that consider complexities such as the shadow effect and self-intersection. The source code for the present work is available at https://github.com/shane-riley/view-factor-cuda