An Energy-Efficient GAN Processor for Mobile Image Translation

An energy-efficient generative adversarial network (GAN) accelerator is proposed for mobile image-to-image translation. In image translation, low precision bits below 8 were not employed due to significant output image quality degradation. In addition, due to the zero injection of transposed convolution, effective PE utilization decreased up to 89%. To address these problems, this paper proposes two key features: 1) we apply layer-wise dynamic fixed-point quantization and implement bit-combined PE to increase throughput by 2×; 2) By analyzing the matching pattern of transposed convolution, data remapping for transposed convolution that simultaneously computes 4 outputs is proposed. The proposed processor is implemented on ZCU 104 FPGA, achieving energy efficiency of 76.38 GOPS/W while consuming 6.08 W of power for mobile image-to-image translation.