Convolutional neural networks (CNNs) are derived from standard multilayer perceptron (MLP) neural networks optimized for two-dimensional pattern recognition tasks such as optical character recognition (OCR) or face recognition. CNNs are large, complex, and require significant computational resources for training and evaluation. This work discusses the possible time gain that can be achieved by comparing and analyzing three different implementations of data transfer in GPU CUDA to speed up the convolutional neural networks learning time.
Альдарф А., Шакер А. (науч. рук. Бессмертный И.А.) Evaluation of memory access methods for training convolutional neural networks on GPU CUDA // Сборник тезисов докладов конгресса молодых ученых. Электронное издание. – СПб: Университет ИТМО, [2021]. URL: https://kmu.itmo.ru/digests/article/5452