Skip to Content
You are currently on the new version of our website. Access the old version .
PhotonicsPhotonics
  • Article
  • Open Access

24 December 2025

17 Pages

Optical Triple-Band Multiplexing Enabling Beyond-600 Gb/s Single-Photodiode Reception for Intra-AIDC Interconnects

,
,
,
,
,
,
,
,
and
1
State Key Laboratory of Photonics and Communications, Department of Electronic Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
2
State Key Laboratory of Photonics and Communications, School of Electronics, Peking University, Beijing 100871, China
*
Authors to whom correspondence should be addressed.

Abstract

Generative artificial intelligence (AI) models including GPT, Gemini, and DeepSeek are reshaping embodied agents, temporal prediction, and autonomous driving, demanding a ten-fold annual growth in training FLOPS that Moore’s law can no longer sustain. Consequently, scale-out GPU clusters require >400 Gb/s lane-rate optical interconnects within AI data-centers (AIDCs). Single-photodiode direct detection offers density, latency, and energy advantages, but DAC bandwidth remains limited to around 70 GHz. We present an optical triple-band multiplexing scheme that replaces high-frequency radio frequency (RF) mixers and local oscillators (LOs) with photonic components. A Mach–Zehnder modulator (MZM) generates 80-GBd PS-PAM-20 signal while an in-phase/quadrature (IQ) modulator driven by a wavelength-offset laser creates two independent 35-GBd PS-64-QAM bands. The proposed optical multiplexing method breaks conjugate symmetry and enhances dispersion tolerance of the direct detection system. After 200 m SSMF transmission and single 70-GHz photodiode (PD) detection, digital signal-signal beating interference (SSBI)/cross-beating compensation enables the recovery of net 543.9 Gb/s signal (line rate of 686.6 Gb/s) using only 45-GHz DACs. The optical multiplexing architecture provides a path to beyond-400 Gb/s lanes and demonstrates a scalable, energy-efficient solution for next-generation AI clusters.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.