Hardware-efficient multiplier-free complex-valued shift-add neural network equalizer for 210-Gb/s 128-QAM SSB-DD links.

Abstract

The rapid growth of data-center traffic is driving demand for high-capacity and spectrally efficient optical interconnects. Single-sideband (SSB) transmission with high-order quadrature amplitude modulation (QAM) is attractive for scale-across direct-detection links, because it avoids dispersion-induced power fading. However, photodiode square-law detection introduces severe signal-signal beat interference (SSBI), which is further aggravated by front-end bandwidth limitations. Although Kramers-Kronig (KK) receivers or iterative SSBI cancellation can mitigate these impairments, their cascaded processing is not optimal for joint compensation. In this Letter, we propose a multiplier-free complex-valued shift-add neural network (CV-SANN) equalizer for 128-QAM SSB direct-detection links in an end-to-end optimization mechanism. The proposed equalizer combines waveform-domain complex-valued joint equalization with power-of-two weight quantization, so that all real-valued multiplications in the equalizer are replaced by bit-shift and addition operations with low hardware complexity. Experimentally, we demonstrate 210-Gb/s transmission over 20-km standard single-mode fiber. The proposed CV-SANN achieves nearly the same performance as its full-precision complex-valued counterpart, with only a 0.03-dB penalty in recovered signal-to-noise ratio (SNR), while providing a 0.79-dB SNR improvement over KK and the feedforward equalizer.