P50: Energy-Efficient and Scalable Bio-Inspired Nanophotonic Computing
Abstract: This paper discusses bio-inspired neuromorphic computing utilizing nanophotonic, nanoelectronic, and NEMS technologies integrated into reconfigurable 2D-3D integrated circuits as hierarchical neural networks. The goal is to achieve ≥1000x improvements in energy-per-operation compare to the state-of-the-art implementations of neural networks on Von-Neumann based computers. We combine nanophotonic and nanoelectronic technologies to build energy-efficient (~10 fJ/b) artificial spiking neurons with required functionality (spiking, integration, thresholding, reset). Photonic interconnects exploiting 2x2 NEMS-MZIs enables distance independent propagation of signal with weighted addition among the neurons as well as possibility of on-line learning capability. Using low-leakage nanophotonic and nanoelectronic devices, and NEMS, the static power consumption of the system can be decreased down to nearly zero. Realizing 2D-3D phothonic integrated circuit technologies, the proposed system can overcome the scalability limitations of current neuromorphic computing architectures.
Award: Best Poster Finalist (BP): yes
Two-page extended abstract: pdf