Deep learning plays a vital role in broad spectrum of scientific fields, such as computer vision, speech recognition, natural language processing, and so on. In order to support deep learning, many frameworks are created with the aim of setting up artificial neural networks as quickly as possible. Such frameworks can be run on systems including either Graphics Processing Unit or Intel Xeon Phi the second generation - Knights Landing coprocessor. However, very few deep learning frameworks can be run on legacy systems containing Intel Xeon Phi Knights Corner. For that reason, we propose and develop pyMIC-DL which is a NumPy-like library supporting deep learning frameworks run on such legacy systems. pyMIC-DL that is an extension of offloading module pyMIC implements basic functions so that it can be easily integrated into deep learning frameworks. The experimental findings show that pyMIC-DL outperforms NumPy in terms of two hardware platforms with the similar theoretical peak performance. Moreover, pyMIC-DL has shown its high effectiveness when being integrated into a well-known deep learning framework Chainer with a very impressive performance. Hence, pyMIC-DL is a fairly promising NumPy-like library to facilitate deep learning frameworks run on the legacy systems.
|Number of pages||8|
|Publication status||Published - 28 Jun 2018|
|MoE publication type||Not Eligible|