Integrated Photonics and Electronics for Optical Transceivers Supporting AI/ML Applications

Abstract

The recent proliferation of artificial intelligence and machine learning applications relying on large language models is fueling unprecedented demand for compute capacity. Associated with this is a need to scale capacities of short-reach optical transceivers towards multiplex Terabit/s, while maintaining integration density (frontpanel or beachfront density) and energy efficiency (pJ/bit). One option to scale transceiver capacity is to increase the bandwidth per lane from today's 200 G to 400 G or even higher: coherent transceiver technology is then expected to play an ever more important role. Photonics and electronics with higher bandwidths beyond 100 GHz will play a crucial role. Integration of thin-film LiNbO3 modulator onto a Silicon Photonics platform is shown to be a viable option to meet the needs for new generations of optical transceivers. Front-end electronics such as linear modulator drivers and transimpedance amplifiers can rely on traveling-wave design approaches to allow continued bandwidth scaling despite (relative) slowing transistor speeds. Novel wireline data converter architectures can be used to overcome limitations of existing implementations. Maintaining signal integrity from photonics and electronics can be facilitated using both 2.5D and 3D integration approaches. While the introduction of novel materials and architectures will require time to further mature, optical transceivers operating at baudrates up to and beyond 200 Gbaud are now just beyond the horizon.