This AI Uses Light Instead of Electricity and It's Mind-Blowingly Fast

photonic-computingneural-networksdasoptical-computingai-hardware

Abstraction: Photonic neural network accelerator processes distributed acoustic sensing in real-time

Key points:

Researchers from Nanjing University developed the TWM-PNNA (Time-Wavelength Multiplexed Photonic Neural Network Accelerator) for real-time processing of Distributed Acoustic Sensing (DAS) data
DAS uses fiber optic cables as sensor networks for earthquake detection, railway monitoring, oil exploration, and submarine surveillance but generates data too fast for conventional electronics
TWM-PNNA demonstrated 1.6 TOPS at 0.87 TOPS/watt measured; theoretical peak of 81 TOPS at 21.02 TOPS/watt, outperforming comparable electrical GPUs by orders of magnitude
Achieved >90% classification accuracy (vs. 98.3% for conventional electronic systems) by controlling modulation chirp ratio below 0.1
System maintained >90% accuracy retaining only 60% of full-connection parameters after pruning, enabling further size reduction
First successful integration of photonic neural networks with DAS systems capable of real-time processing; published in Advanced Photonics, March 2025

Connections: Nanjing University · Photonic Computing · Neural Networks · AI Hardware