⚡ 本页包含 AI 生成的分析内容,仅供参考
本文提出了一种32Gb/s/λ的密集波分复用带通滤波时钟前向光链路,采用9个200GHz间距的波长(8个数据,1个时钟),通过3D堆叠7nm电子IC和65nm硅光子IC实现,总吞吐量256Gb/s/光纤,能效2.78pJ/b,面积效率1.33Tbps/mm²。解决了高带宽密度和低功耗光互连的挑战。
Ward Lopes1, Benjamin G. Lee3, Thomas H. Greer III2, C. Thomas Gray2 Nvidia, Santa Clara, CA, 2Nvidia, Durham, NC, 3Nvidia, Ridgefield, CT 1 Abstract We present a 32Gb/s/λ dense wavelength division multiplexing bandpass-filtered clockforwarding link featuring nine 200GHz-spaced wavelengths (8 for data and 1 for clk). Fabricated in a 3D-stacked technology with an electronic IC in a 7nm process and a photonic IC in a 65nm SiPh process, our implementation achieves a throughput of 256Gb/s/fiber, delivering an energy efficiency of approximately 2.78pJ/b and an area efficiency of 1.33Tbps/mm2—while maintaining 0.46UI wide eye openings at <1E-11 BER. The rapid growth of data centers, driven by increasing demand from artificial intelligence training and inference, has highlighted the limitations of traditional copper cables in terms of density, reach, and energy efficiency. Photonic links, especially dense wavelength division
Sanquan Song1, Nandish Mehta1, Nikola Nedovic1, Angad Rekhi1, Georgios Kalogerakis1, Li Xu1, Brian Zimmer1, Stephen G. Tell2, Yoshi Nishi1, Xi Chen1,