Distributed artificial intelligence via energy-efficient optical interconnects

In current data centers, short-range chip-to-chip communication is achieved via electrical wires, while long-range communication is handled via optical interconnects. However, the growth of data center energy consumption calls for energy-efficient short-haul optical interconnects, and current optical interconnects face significant challenges in terms of energy, bandwidth, and cost.

This project proposes a free-space photonic interconnect system that utilizes an unprecedented number of dense channels operating at the native bandwidth of the electronics. This approach dramatically reduces power consumption (by more than an order of magnitude) and latency (by more than two orders of magnitude) of large AI systems while reducing the cost per bandwidth.