Implementation of the Continuous Space Language Model on a Heterogeneous Mobile Processor

Mobile processors continue to increase in performance while becoming more power efficient, providing consumers with improved gaming, multi-media, and browsing, along with longer lasting device usage. To keep up with consumer multimedia demand, mobile processor manufacturers have begun to integrate Graphical Processing Units (GPU) on mobile processors, providing users with the high performance graphics required for mobile gaming applications. The addition of integrated GPUs to the mobile processors also offers new opportunities for introducing to the mobile platform computationally intensive algorithms that were formerly impractical when running on the mobile CPU processor alone. GPU manufacturers such as NVIDIA are scaling down their GPU architecture from desktop systems to those developed for use on mobile processors, such as the Tegra K1, featuring a single GPU processing core. This research effort investigates the performance of the computationally intensive Continuous Space Language Model (CSLM) algorithm on the NVIDIA Tegra K1 mobile processor and compares its performance on this platform to that of the desktop GPU platform reported by Thompson et al. [1]. The performance on the embedded GPU platform will also be compared to that of a conventional embedded CPU. However, first, the execution time of the algorithm will be observed while executing on a laptop CPU platform to provide a reference point for the performance of the Tegra K1 CPU processor(s). Next, the algorithm will be configured to execute solely on the Tegra K1 CPU processor(s), and the execution time will be observed. Finally, the execution time of the algorithm will be observed after porting only the computationally intensive portions of the algorithm to the Tegra K1 GPU while other portions remain on the embedded CPU