Min-Kyu Song1,Hyunseok Kim1,Junmin Suh1,Jeehwan Kim1
Massachusetts Institute of Technology1
Min-Kyu Song1,Hyunseok Kim1,Junmin Suh1,Jeehwan Kim1
Massachusetts Institute of Technology1
Conventional integration technologies have encountered issues including material deterioration due to the physical interaction between chips, fixed operation modes owing to the permanent connections to a single sensor, and limited data processing capabilities due to constraints when adding processors for various computing needs. In this study, we developed an innovative, reconfigurable 'Lego-style' 3D integration platform. This platform uses optoelectronic devices to facilitate communication between AI processors. We integrated freestanding GaAs/InGaP membranes into each chip via an epitaxial lift-off (ELO) process. Communication between chips was enabled by optical transmission between GaAs LEDs and InGaP photodiodes. By eliminating electrical interconnects, we not only avoided interference from physical contact, but also significantly enhanced chip configuration flexibility. We further incorporated memristor arrays into each stackable chip, serving as AI hardware accelerators and greatly improving processing efficiency. This innovative approach overcomes the limitations observed in previous 3D heterogeneous integration methods. By using this novel 'Lego-style' 3D heterointegration process based on a GaAs/InGaP stack, edge computing systems integrated with various sensors can provide infinite modality, exceptional computational efficiency, and adaptability for a variety of computing tasks.