Monolithic Silicon-Photonics Sign-Sign Dot-Product Circuits Toward Artificial Intelligence and Multiple-Input Multiple-Output Accelerators.

This research leverages advanced monolithic silicon-photonics integrated-circuit manufacturing capabilities to realize system-on-chip photonic-computing-based linear-algebra accelerators for a wide range of applications in artificial intelligence, machine learning, and multiple-input multiple-output wireless technology. With holistic codesign in both photonic and electronic domains, strategic electrical-to-optical signal conversion, a differential intensity-modulation technique, and a dual rail-to-rail photodetection architecture, the monolithic photonic-electronic test chip of a sign-sign dot-product accelerator achieves 8.92-Gb/s/MAC computation throughput with 2.22-pJ/b/MAC energy consumption for next-generation large-scale linear-algebra computing hardware targeting higher than one TMAC/s/mm2 computation density with only tens of fJ/MAC energy consumption.