Empirical Study To Compare The Performance Of Novel CPU Implementation Of Deep Learning Algorithms With GPU-BASED Implementatioms

Abstract Deep learning architectures and algorithms have shown promising results in different applications. However, training deep learning algorithms is a time-consuming step and researchers used graphics processing unit (GPU) accelerators in order to achieve an acceptable execution time, especially for real life applications. Sub-LInear Deep Learning Engine (SLIDE) is a relatively new work that aims at speeding up deep learning using only central processing unit (CPUs) implementation. Additionally, TorchSLIDE used PyTorch libraries to speed up SLIDE by 2.6X. This research study attempts to provide an empirical comparison for four models which are: PyTorch baseline CPU, PyTorch baseline GPU, TorchSLIDE, and SLIDE. In the experiments, we used a server with 10 cores (Intel i9-7900X) and NVIDIA Quadro P4000 (GP104GL) GPU. Five datasets were used in the evaluation which are: Amazon-670K, Delicious-200K, AmazonCat-13K, LF-AmazonTitles-131K, and Wiki10-31K. Based on the experiments, Py- Torch baseline GPU outperforms the other three implementations and achieved the highest accuracy in less time compared with the others for all datasets. Moreover, SLIDE outperformed TorchSLIDE when using the AmazonCat-13K dataset.