On Construction of a Caffe Deep Learning Framework based on Intel Xeon Phi

With the increase of processor computing power, also a substantial rise in the development of many scientific applications, such as weather forecast, financial market analysis, medical technology and so on. The need for more intelligent data increases significantly. Deep Learning as a framework that able to understand the abstract information such as images, text, and sound has a challenging area in recent research works. This phenomenon makes the accuracy and speed are essential for implementing a large neural network. Therefore in this paper, we intend to implement Caffe deep learning framework on Intel Xeon Phi and measure the performance of this environment. In this case, we conduct three experiments. First, we evaluated the accuracy of Caffe deep learning framework in several numbers of iterations on Intel Xeon Phi. For the speed evaluation, in the second experiment we compared the training time before and after optimization on Intel Xeon E5-2650 and Intel Xeon Phi 7210 . In this case, we use vectorization, OpenMP parallel processing, message transfer Interface (MPI) for optimization. In the third experiment, we compared multinode execution results on two nodes of Intel Xeon E5-2650 and two nodes of Intel Xeon Phi 7210.