| IP개요 |
This project proposes an accuracy configurable energyefficient approximate recurrent neural network (RNN) based long short-term memory (LSTM) processor for mobile systems. The conventional RNN-based systems have been suffered from inefficient processing which caused by recurrent processing fashion. With the conventional approach, most of previous studies only focused on improving processing speed of system, which had been applied on servers and cloud systems. In recent studies, attempt on LSTM optimization design technique has been approved by increasing demand on AI-based system. For implementing effective throughput with energy efficient processing, the bit precision-based operation can provide reasonable energy efficiency with the proper accuracy for optimized mobile systems. As mentioned above, the bit precision-based method will also control accuracy in real-time processing. The proposed LSTM processor can achieve state-of-theart energy efficient technique with the scaled accuracy. All the functionalities of the proposed work are verified by software simulations using Tensorflow and verified by RTL simulation using Xilinx FPGA. Also, we will synthesize the proposed hardware with power-aware P&R. As a result, the proposed system is fabricated in 65nm CMOS process, this system could be realization for state-ofthe art NN based processor hardware design. |
