A Low Power Rnn Accelerator Based On Dynamic Accuracy Approximate Computing

In recent years,recurrent neural networks have achieved superior results in many areas of artificial intelligence.However,most recurrent neural network models require extremely high computational power and cannot be deployed on low-end mobile devices,such as mobile phones and embedded systems.This hinders the application of recurrent neural networks in the daily life and the commercial area.Low power recurrent neural network accelerators can promote the application of recurrent neural networks in the field of mobile devices,which are of great significance to the construction of the Internet of Things.This thesis takes the low power recurrent neural network accelerator as the research object,innovatively proposes a dynamic accuracy approximate computing strategy for low power recurrent neural network accelerator,and completes the design of low power recurrent neural network accelerator based on dynamic accuracy approximate computing.The dynamic accuracy approximate computing strategy is divided into two layers: algorithm and hardware.At the algorithm level,network-level hierarchical quantization strategy,frame-level precision adaptive strategy and connection-level dynamic pruning strategy are proposed.At the hardware level,an approximate multiplication strategy is proposed,which uses an approximation multiplier based on approximate addition instead of the traditional multiplier.The arrays of the low power recurrent neural network accelerator based on dynamic accuracy approximate computing are composed of dynamic accuracy approximate multipliers.The dynamic accuracy approximate multipliers can be reconstructed into computing units with different bit-width and different computing precision.The control modules of different functions implement the control of dynamic accuracy at each level.The experiment shows that the effect of the proposed dynamic accuracy approximate computing on the accuracy of the recurrent neural networks is within 6%.Comparing with the non-optimized RNN accelerator,proposed low power RNN accelerator increases the energy-efficiency by up to 547.95%.In the TSMC 28 nm technology,the peak throughput of proposed accelerator achieves 989.7GOPs,the power consumption is 276.4mW,and the normalized energy efficiency is 3.58 TOPs/W.