A Power-Efficient Specific Emitter Identification Hardware Accelerator With SNR-Aware Adaptive Precision Reconfiguration

Abstract

Specific emitter identification (SEI) plays a crucial role in spectrum management applications. Recently many SEI algorithms based on deep convolution neural network (DCNN) have been proposed. However, the high computational complexity of DCNN leads to large power consumption. Also, there is a lack of dedicated SEI hardware design. In this work, we proposed a SNR-aware adaptive scalable SEI algorithm based on DCNN, and implemented a power-efficient SEI hardware accelerator based on it. The SEI hardware accelerator can adaptively reconfigures itself between 16-bit weight computing mode and binary weight computing mode by sensing signal-to-noise ratio (SNR) condition to improve power efficiency while maintaining high accuracy. The experiment results show that our accelerator can achieve a high power efficiency of 6.51Gops/W for the high SNR condition and a high accuracy of 80% in the low SNR condition, compared with conventional non-adaptive designs with 16-bit or binary computing mode.