Reconfigurable Heterointegration of Artificial Intelligence Chips Using GaAs/InGaP-Based Optoelectronic Devices

The deployment of conventional integration technologies still faces the issues including material degradation due to intimate contact between chips, fixed modality due to fixed connection to one single sensor, and limited data processing due to constraint for addition of processors in varying computing situation. In this work, we developed a novel reconfigurable 3D heterogeneous integration platform using optoelectronic devices for chip-to-chip communication. Freestanding GaAs/InGaP membrane was integrated into each chip by epitaxial lift-off (ELO) process. Chip-to-chip communication was enabled by optical transmission between GaAs LEDs and InGaP photodiodes. Optical communication without any electrical interconnect resulted not only in elimination of interference from physical contact, but also in great reconfigurability of the chips. Furthermore, the neuromorphic computing layers were embedded in each stackable chip as AI hardware accelerators for extremely efficient processing. This technique will solve the shortcomings of previous conventional 3D heterogeneous integration methods. With the Lego-like 3D heterointegration method based on the GaAs/InGaP stack, the sensor-computing systems can provide unlimited modality, highly efficient computing and adaptability to varying computing tasks.