AI cluster networks based on distributed WDM photonic switches

The introduction of AI into our daily lives has caused the amount of training compute to grow by 4x per year. With AI clusters having tens to hundreds of thousands of accelerators connected in a large network, the network infrastructure accounts for a consistent part of the total power consumption to make fast and low latency connections between accelerator nodes at extreme data rates possible. WDM photonic switches are excellent candidates to replace electronic switches in AI clusters. The lack of O/E/O conversion causes the power consumption and cooling requirements to be a fraction of its electronic counterparts, greatly reducing operational expenditures. Furthermore, the transparency to the input signal in terms of wavelength, data rate and modulation format, effectively allow network operators to mix interconnect generations, increasing network sustainability and reducing capital expenditures. Challenges to realize practical switch are large port count at negligible insertion loss and fast switching, all of which are greatly desired in compute clusters. The InP platform in combination with other low loss passive technologies allow for lossless switches with sub-microsecond switching times due to the use of semiconductor optical amplifiers (SOAs) as switch and amplifier. In this paper we present a modular WDM photonic switch architecture based on hybrid InP and silica technologies. Results among multiple different paths of the WDM photonic switch with 100 Gb/s PAM-4 modulated signals confirm an average power penalty at the FEC limit of 0.8 dB with an OSNR of 34 dB and a broadband operation of the switch.