Fiber Optics Links for Artificial Intelligence Computing and Networks

Abstract: Artificial Intelligence/Machine Learning solutions are some of the most discussed and debated, most impactful, and transformative technologies and areas today. AI/ML networks are the fastest growing segment of data networking pushing bandwidth demand and ever higher data rates. Fiber Optics links provide the data communications foundation for AI/ML computing and data center networks. AI/ML computing and networks are deploying 100 Gb/s, 200 Gb/s, 400 Gb/s, and 800 Gb/s Ethernet and InfiniBand data rates, utilizing 25G NRZ (Non-Return-to-Zero), 50G and 100G PAM-4 (Pulse Amplitude Modulation with 4 Levels) modulation, for switch-server, switch-to-switch, and within switch topologies, computing, and data communications. Standards and MSAs (Multi-Source Agreement) have completed specifications to support 800 Gb/s and even possible 1.6 Tb/s data rates over multiple lanes utilizing 100 Gb/s signaling for both MMF and SMF optical links to support AI/ML networks. The industry is focused on defining standards Ethernet optical links with 200 Gb/s signaling with demand for AI/ML networks in 2025 and beyond – these initial optical links will be SMF. However, there are now technology feasibility demonstrations for short-reach Ethernet 200G VCSEL-MMF links with possible solutions before the end of the decade or before. We present the parameters, features, and select channel results of the latest Multimode and Single-Mode fiber solutions to support the requirements for these challenging AI/ML computing and data center networks. We present the importance of and latest advances in stringent DMD bandwidth specifications, tightest geometry tolerances for ultra-low-loss connections, reduced microbend sensitivity for challenging cable constructions, OM3, OM4, OM4+ optimized MMF, high density and enhanced usability of Rollable Ribbon and Bend Optimized solutions for both MM and SM, and chromatic zero dispersion range for SM fiber. We review the relative benefits of AOC (Active Optical Cable), pluggable transceivers and structured cabling, and briefly cover the roadmap for Co-Packaged Optics (CPO) fiber solutions. Low power and low latency are critical requirements for AI/ML computing and networks. We present data for an ultimate low latency link with 50G NRZ, linear interface, and Hollow Core Fiber (HCF). We review the impact and benefits of these fiber solutions for enterprise data center networks. Our key conclusion is that the range of existing and new, innovative MMF, SMF, HCF, cable designs, technologies, and standards will continue to meet technical challenges and customer needs for cost-effective, low-power, low-latency, high-density, high-speed short-reach optical links to support the exponential growth in AI/ML computing and networks. About the Presenter: G. Mabud Choudhury is Standards Manager for OFS. He contributes to IEEE 802.3 Ethernet, INCITS T11 Fibre Channel, and IEC Fibre standards. He is Vice Chair for IEEE P802.3db 100 Gb/s, 200 Gb/s, and 400 Gb/s Short Reach Fiber Task Force. Mr. Choudhury has an MS degree in Mechanical Engineering from Massachusetts Institute of Technology, and a BS degree in Mechanical Engineering from Duke University. He holds 32 US patents.