High Performance Implementation Of Multiple Machine Learning Algorithm On GPGPU

The main point of this thesis is how to implement some machine learning algorithm on GPGPU, with high performance. The algorithm includes dynamic time warping(DTW),feedforward neural network. The limits, advantages, key points of GPGPU development are discussed in this thesis. This thesis shows how to make algorithm faster and how to take advantages of GPGPU. We improved the calculation speed by 66.32% and reduced the video card memory by 91.70% on DTW implementation. The speed of multiple GPGPUs SGD implementation, the training algorithm of neural network, is 1.53 times the computing speed of a single graphics card.