Photonic neural network chips promise compact footprint, low latency, and high energy efficiency. Yet, their scale and computing throughput are fundamentally constrained by one-dimensional input interfaces, unavoidable waveguide crossings, and the resulting crosstalk and excess loss. As a result, two-dimensional (2D) image data must be serialized through limited input ports, sacrificing spatial parallelism and creating input/output (I/O) bottlenecks. Here we demonstrate a programmable three-dimensional (3D) photonic neural network chip, fabricated by femtosecond laser direct writing (FLDW) in glass, that directly processes 2D images. The cascaded architecture alternates photonic-lantern waveguide arrays and phase-shifter arrays to implement matrix operations. An 8-layer 8 × 8 device achieves a computing throughput of 6554 TOPS, surpasses leading planar photonic platforms, and delivers 93% accuracy on MNIST classification and 94% fidelity in optical pattern generation. By combining 3D spatial parallelism with programmability, this work establishes a scalable paradigm for reconfigurable photonic computing in complex inference tasks. Photonic AI chips can process information at high speed, but current designs are constrained by planar layouts. Here, authors demonstrate a programmable 3D photonic neural network chip that directly processes 2D images and achieves 93% accuracy on MNIST classification.
Cao et al. (Tue,) studied this question.