The implementation of a Deep Recurrent Neural Network Language Model on a Xilinx FPGA
This enables efficient natural language processing on mobile/portable devices, though it is an incremental improvement in FPGA-based acceleration.
The paper tackles the challenge of deploying deep recurrent neural network language models on portable embedded systems by implementing a hardware accelerator on a single-FPGA platform, achieving a processing throughput of 20 GOPS with speedups of 70.5× and 2.75× compared to prior work.
Recently, FPGA has been increasingly applied to problems such as speech recognition, machine learning, and cloud computation such as the Bing search engine used by Microsoft. This is due to FPGAs great parallel computation capacity as well as low power consumption compared to general purpose processors. However, these applications mainly focus on large scale FPGA clusters which have an extreme processing power for executing massive matrix or convolution operations but are unsuitable for portable or mobile applications. This paper describes research on single-FPGA platform to explore the applications of FPGAs in these fields. In this project, we design a Deep Recurrent Neural Network (DRNN) Language Model (LM) and implement a hardware accelerator with AXI Stream interface on a PYNQ board which is equipped with a XILINX ZYNQ SOC XC7Z020 1CLG400C. The PYNQ has not only abundant programmable logic resources but also a flexible embedded operation system, which makes it suitable to be applied in the natural language processing field. We design the DRNN language model with Python and Theano, train the model on a CPU platform, and deploy the model on a PYNQ board to validate the model with Jupyter notebook. Meanwhile, we design the hardware accelerator with Overlay, which is a kind of hardware library on PYNQ, and verify the acceleration effect on the PYNQ board. Finally, we have found that the DRNN language model can be deployed on the embedded system smoothly and the Overlay accelerator with AXI Stream interface performs at 20 GOPS processing throughput, which constitutes a 70.5X and 2.75X speed up compared to the work in Ref.30 and Ref.31 respectively.