Petascale计算
光子学
硅光子学
硅
计算机科学
硅片
炸薯条
能量(信号处理)
光电子学
材料科学
超级计算机
并行计算
电信
物理
量子力学
作者
Yuyang Wang,Songli Wang,Robert Parsons,Swarnava Sanyal,Vignesh Gopal,Asher Novick,Anthony J. Rizzo,Michal F. Lipson,Alexander L. Gaeta,K. Bergman
出处
期刊:IEEE Transactions on Components, Packaging and Manufacturing Technology
[Institute of Electrical and Electronics Engineers]
日期:2024-11-06
卷期号:15 (8): 1581-1591
被引量:22
标识
DOI:10.1109/tcpmt.2024.3492189
摘要
Data volume in hyperscale computing systems has surged exponentially over the past decade, notably driven by artificial intelligence (AI)/machine learning (ML) applications and the emergence of large-scale generative AI models. An urgent need arises for ultrahigh-bandwidth and energyefficient communications among compute clusters to support the application demands. Embedded silicon photonics (SiPh) promises to enable petascale system-wide connectivity by integrating optical input/output (I/O) directly into the compute socket. SiPh microresonator-based modulators and filters, known for their excellent wavelength selectivity and compact footprints, offer an elegant solution for realizing dense wavelength-division multiplexing (DWDM) links with ultrahigh bandwidth density, leveraging the latest advances in optical frequency comb (OFC) sources and 3-D integration with electronics. In this work, we present our scalable DWDM link architecture, designed with co-packaging in mind. We report device-level measurements of key components and validate comb-driven end-to-end data transmission. These results demonstrate promise in realizing co-packaged optical I/Os with shoreline and aerial bandwidth densities beyond 4 Tbps/mm and 17 Tbps/mm2 while consuming sub-pJ/b energy, paving the way for petascale photonic connectivity for energy-efficient computing.
科研通智能强力驱动
Strongly Powered by AbleSci AI