Work

メモリインテンシブアレイアクセラレータを用いた畳み込みニューラルネットワーク処理

Year: 2015
Type: dissertation
Abstract: Convolutional neural network (CNN) is a major approach to achieve high accuracy on image recognition.However, computation of convolution value in CNN is time and energy consuming.Although GPU is widely used for accelerating CNN, inherent performance cannot be exploited due to memory access and preprocessing overheads, if the convolution is not sufficiently large.This thesis proposes a high performance execution method of CNN on a CGRA (Coarse Grained Reconfigurable Architecture).By using a CGRA with on-chip distributed memory, the memory access overhead is significantly reduced.The proposed approach can execute both forward and back propagations on CGRA by instruction remapping.I evaluated the proposed method by using various scales of CNN.The proposed method achieves up to 4.5 times higher area performance and up to 1.8 times higher power performance, compared to a state of the art mobile GPU.
Language: Japanese
Open Access status: green
OpenAlex ID: W7144535095
Cites: 33
Cited by:
Related to: 10