面向AI时代GPU集群的先进400G-SerDes调制与均衡器
Advanced 400G-SerDes Modulations and Equalizers for GPU Cluster in AI Era
另 1 家 · 查看机构详情
- Jinan University（暨南大学）
- Beijing Institute of Technology（北京理工大学）
- Guangdong Polytechnic Normal University（广东技术师范大学）
- Shijiazhuang Tiedao University（石家庄铁道大学）
机构由 AI 辅助整理，请以论文原文为准。
浏览论文内容
中文总结 AI 辅助
本文针对AI时代GPU集群400G-SerDes互连，研究高阶调制与均衡技术，推导PAM4/PAM6/PAM8的BER表达式，并通过仿真比较不同带宽限制下的性能，发现PAM6在多数信道下达到KP4-FEC极限所需PSNR最低，具有应用潜力。
中文摘要 AI 辅助
高速GPU集群互连正从200 Gbps向400 Gbps演进，以满足人工智能日益增长的算力需求。我们研究了面向400 Gbps互连的高阶调制与先进均衡技术。推导了4/8进制脉冲幅度调制（PAM）的理论误码率（BER）表达式，以及由32进制正交幅度调制（32QAM）生成的六种PAM6格式的近似表达式。在高斯白噪声（AWGN）信道下，由framed-cross 32QAM生成的PAM6在PAM6格式中实现了最佳BER性能。在KP4前向纠错（KP4-FEC）极限下，PAM6和PAM8相对于PAM4分别引入了约4.2 dB和7.2 dB的峰值信噪比（PSNR）代价。仿真在440 Gbps线路速率（净速率为400G）下，针对不同带宽限制条件比较了PAM4、PAM6和PAM8。随着带宽受限加剧，带宽需求对性能的影响逐渐超过噪声容限。在轻度带宽受限信道中，PAM4达到KP4-FEC极限所需的PSNR最低，而在中度受限条件下PAM6表现最佳。在重度受限条件下，PAM8因其较低的带宽需求和较小的对高频衰减敏感性而具有显著优势。采用最大似然序列估计（MLSE）后，PAM6和PAM8在所有评估信道条件下均能达到KP4-FEC极限，而PAM6在大多数信道中所需的PSNR更低，凸显了其在400G-SerDes系统中的潜力。
英文摘要
High-speed GPU cluster interconnects are evolving from 200 to 400 Gbps to meet the growing computing demands of artificial intelligence. We investigate high-order modulation and advanced equalization for 400 Gbps interconnects. Theoretical bit error rate (BER) expressions for 4/8-ary pulse amplitude modulation (PAM) and approximate expressions for six PAM6 formats generated from 32-ary quadrature amplitude modulation (32QAM) are derived. In additive white Gaussian noise, PAM6 generated by framed-cross 32QAM achieves the best BER performance among the PAM6 formats. PAM6 and PAM8 incur peak signal-to-noise ratio (PSNR) penalties of approximately 4.2 and 7.2 dB, respectively, relative to PAM4 at the KP4 forward error correction (KP4-FEC) limit. Simulations compare PAM4, PAM6, and PAM8 at a line rate of 440 Gbps (net rate of 400G) under different bandwidth limitations. As bandwidth becomes more constrained, bandwidth requirements increasingly dominate performance over noise tolerance. PAM4 requires the lowest PSNR to reach the KP4-FEC limit in mildly bandwidth-limited channels, whereas PAM6 performs best under moderate limitations. PAM8 offers a significant advantage under severe limitations because of its lower bandwidth requirement and reduced susceptibility to high-frequency attenuation. With maximum likelihood sequence estimation, PAM6 and PAM8 reach the KP4-FEC limit under all evaluated channel conditions, while PAM6 requires a lower PSNR in most channels, highlighting its potential for 400G-SerDes systems.
