Abstract: In view of Restricted Boltzmann Machine problem of slow training and take too long time, on the basic of cloud computing platform Hadoop distributed task processing and GPU acceleration mechanism, the realization of RBM training based on Hadoop-GPU is designed. By researching the producers of Hadoop and training steps of RBM, the training of RBM is decomposed into two steps, and combined by Hadoop combination-job mechanism. As the basic model RBM training, the Map function in Hadoop runs Gibbs block sampling and Reduce function executes parameters update. The MNIST handwriting digit database is conduct on the test of this new realization. The MNIST experiment results illustrated that the novel algorithm has good feasibility and is advantageous for hug amount of data, especially with the increase of the data size, the more significant speedup showing growth and during RBM training the speed-up rate reach at least 20.