Performance analysis of 200G transceivers for next-generation intra-datacenter VCSEL-MMF links

In recent years, with the availability to the general public of large language models (LLMs), we have seen an unprecedented pressure on the development of datacenters. Training LLMs over thousands of GPUs requires parallel computation and communication: it is important to efficiently scale capacity, size and power consumption. Photonic technologies are needed to keep up with these trends. An area of applications where optical links have already been successfully employed is intra-datacenter connections. Short links connecting spine switches, leaf switches and server racks, are nowadays usually based on multi-mode fibers (MMF) and vertical cavity surface emitting laser (VCSEL) sources. Solutions with 100G per lane are becoming commercially available, also for short wavelength division multiplexing (SWDM) systems, but there is a push towards higher rates. To keep up with the increased throughput demand, 200G per lane is expected to be the next step, but this poses new challenges in the design and fabrication of the transceiver (TRX). In this paper, we define requirements for the next-generation 200G TRXs, considering different options for the modulation format and the equalization strategy, and we validate them with a statistical study over a large set of OM3 and OM4 fibers. We show that the maximum reach can be 20 m and 30 m, respectively for OM3 and OM4, on all wavelengths of the SWDM.