Optimizing GPU Performance in Machine Learning with Resizable BAR: An Analytical Study

The performance of machine learning systems is influenced by several factors, including algorithm complexity, memory capacity, and processing components like CPUs or GPUs. Generally, GPUs outperform CPUs in machine learning tasks due to their ability to execute multiple simultaneous computations. This parallel processing capability allows for efficient distribution of training processes, significantly accelerating machine learning operations. Since 2021, a new GPU technology called resizable BAR (Base Address Register) has emerged, enabling the GPU's entire graphics frame buffer to be accessed simultaneously by the CPU. This research aims to investigate whether enabling the resizable BAR feature enhances GPU performance compared to its default state. We conducted experiments using various machine learning models, including Resnet-50, VGG-19, DeepLab and LSTM with different purpose. The primary metric recorded was the difference in start and finish times, which was then analyzed using Wilcoxon signed-rank to test the hypothesis of performance improvement. If the hypothesis testing indicates a significant difference, further analysis will be conducted to quantify the performance gain using the n-gain test. By understanding the impact of the resizable BAR on GPU performance, this research aims to provide valuable insights for enhancing the efficiency of machine learning processes. The findings should serve as a guide for future developments and optimizations in GPU technology, particularly in contexts that require extensive computational power and speed, thereby enhancing the overall performance of machine learning algorithms.