@inproceedings{oai:ipsj.ixsq.nii.ac.jp:00183847, author = {Ching-Nung, Lin and Shi-Jim, Yen and Jr-Chang, Chen and Ching-Nung, Lin and Shi-Jim, Yen and Jr-Chang, Chen}, book = {ゲームプログラミングワークショップ2017論文集}, month = {Nov}, note = {The performance of Deep Learning Inference is a serious issue when combining with speed constraint Monte Carlo Tree Search(MCTS). Traditional hybrid CPU and Graphics processing unit solution is bounded because of frequently heavy data transferring. This research focuses on accelerating parallel synchronized Deep Convolution Neural Network(DCNN) prediction in MCTS. This paper proposes a method to accelerate parallel DCNN prediction and MCTS execution at GPU, Intel AVX-512 CPU and Xeon Phi Corner. It outperforms the original architecture using the GPU forwarding server. In some cases, GPU speeds up 7.2 times; AVX-512 CPU increase 15.7 times speed. Xeon Phi Corner accelerates 11.1 times performance. In addition, with 64 threads in Google Cloud Platform, maximal 53.8 times faster is achieved., The performance of Deep Learning Inference is a serious issue when combining with speed constraint Monte Carlo Tree Search(MCTS). Traditional hybrid CPU and Graphics processing unit solution is bounded because of frequently heavy data transferring. This research focuses on accelerating parallel synchronized Deep Convolution Neural Network(DCNN) prediction in MCTS. This paper proposes a method to accelerate parallel DCNN prediction and MCTS execution at GPU, Intel AVX-512 CPU and Xeon Phi Corner. It outperforms the original architecture using the GPU forwarding server. In some cases, GPU speeds up 7.2 times; AVX-512 CPU increase 15.7 times speed. Xeon Phi Corner accelerates 11.1 times performance. In addition, with 64 threads in Google Cloud Platform, maximal 53.8 times faster is achieved.}, pages = {131--137}, publisher = {情報処理学会}, title = {Accelerate Parallel Deep Learning Inferences with MCTS in the game of Go}, volume = {2017}, year = {2017} }