The solution of high-dimensional PBMs using CPUs are often computationally intractable. This study focuses on the development of a scalable algorithm to parallelize the nested loops inside the PBM via a GPU framework. The developed PBM is unique since it adapts to the size of the problem and uses the GPU cores accordingly. This algorithm was parallelized for NVIDIA® GPUs as it was written in CUDA® and C/C++. The major bottleneck of such algorithms is the communication time between the CPU and the GPU. In our studies, communication time contributed to less than 1% of the total run time and a maximum speedup of about 12 over the serial CPU code was achieved. The GPU PBM achieved a speedup of about two times compared to the PBM’s multi-core configuration on a desktop computer. The speed improvements are also reported for various CPU and GPU architectures and configurations.

使用CPU解决高维PBM的方法通常难以计算。这项研究的重点是开发可扩展算法，以通过GPU框架并行化PBM内部的嵌套循环。所开发的PBM具有独特性，因为它可以适应问题的大小并相应地使用GPU内核。由于该算法是用CUDA®和C / C ++编写的，因此已针对NVIDIA®GPU进行了并行处理。这种算法的主要瓶颈是CPU与GPU之间的通信时间。在我们的研究中，通信时间仅占总运行时间的不到1％，并且在串行CPU代码上实现了约12的最大加速。与台式计算机上PBM的多核配置相比，GPU PBM的速度提高了大约两倍。