Cloud Server Oriented FPGA Accelerator for Long Short-Term Memory Recurrent Neural Networks. (August 2019)

• Record Type: Journal Article
• Authors: Wang, Jiasheng; Zhou, Yu; Sun, Yuyang; Li, Keyang; Liu, Jun
• Is Part Of: Journal of physics. Volume 1284(2019)
• Publication Date: 2019-08
• Subjects: Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/1284/1/012044
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: Long Short-Term Memory network(LSTM), which is the most widely used and representative recurrent neural network architecture, plays an important role in language modeling, machine translation, image captioning, etc. However, due to its recurrent nature, general-purpose processors like CPUs and GPUs cannot achieve high parallelism, not to mention their high power consumption. FPGA accelerators can outperform them by flexibility, energy-efficiency and more delicate optimization in each phase of the algorithm. In this paper, we present a cloud-oriented FPGA accelerator for LSTM based on OpenCL. Different from previous works which are designed for embedded systems, our FPGA accelerator performs multiple time series predictions in parallel. We provide a general matrix optimization model to optimize the computation of LSTM in the cloud environments. The performance of our implementation beats both the CPU implementation and other previous hardware implementations. We present and analyze the performance results of our work.