How GPUs Kill Threads in Neural Network Training

Mayta Quispe M.F.; Cruz F.T.; Juarez Vargas J.C. · Lecture Notes in Networks and Systems (2024) · DOI: 10.1007/978-3-031-68660-3_6

Resumen

The optimization and exploration of computing resources continue significant challenges in challenges in the field of neural network training. This paper present an analysis of the effectiveness of several computational architectures, including Graphics Processing Units (GPUs), Central Processing Units (CPUs), and CPUs using threaded techniques, in supporting neural network training. The research highlights the significant differences in training times across different platforms, demonstrating the effectiveness of GPU-based techniques. In order to guarantee a thorough assessment, the research uses two separate datasets: MNIST(Modified National Institute of Standards and Technology), which has 60,000 photos, and USPS(United States Postal Service), which has 7,291 images. These datasets provide the basis of our empirical work, representing a range of dimensions and complexity. The investigation's findings demonstrate the superiority of GPUs in accelerating neural network training, also provide insightful information on how best to use compute resources for these kinds of jobs. This study offers a thorough understanding of the dynamics involved in neural network training by analyzing the minute variations in performance across various computational architectures. Additionally, it lays the groundwork for future investigations into how to make these training procedures more effective, which might result in faster progress in the field of artificial intelligence. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2024.

Instituciones: Universidad Nacional del Altiplano