Neuromorphic hardware can mitigate the unsustainable energy demand of artificial intelligence infrastructure. Photonic approaches provide high speed, low energy, and high connectivity but existing solutions have large circuit footprints which limits scaling potential and they miss key biological functions, like inhibition. We report a nano-optoelectronic artificial neuron with at least 100-fold reduced circuit footprints compared to existing approaches and picowatt-level operating power. The device deterministically integrates excitatory and inhibitory inputs, performs a nonlinear transfer operation, and exhibits biologically relevant temporal dynamics. Neural weighting is implemented via tunable input gains, enabling controlled summation and thresholding. The architecture is compatible with commercial silicon technology, supports multi-wavelength operation, and can be used for both computing and optical sensing. This work paves the way for two important research paths: photonic neuromorphic computing with nanosized footprints and low power consumption, and adaptive optical sensing, using the same architecture as a compact, modular front end. Photonic neuromorphic hardware is fast and energy-efficient, but large circuit footprints remain a key bottleneck for scaling. Here, Sestoft et al. demonstrate a few-nanowire artificial neuron that can reduce circuit footprints by ≥100 × .
Sestoft et al. (Fri,) studied this question.