Integrated Electro-Optic Attention: Hardware Nonlinearities for Optical Transformers

arXiv·medium signal

Mickeler, Lion, and Nardi demonstrate integrated electro-optic components that implement attention nonlinearities directly in optical hardware, moving toward transformers that run on photonic chips rather than electronic GPUs. This addresses the fundamental energy bottleneck of current AI compute by replacing electronic matrix multiplications with optical operations. Early-stage but represents a potential paradigm shift in inference hardware — relevant as AI compute costs and power consumption become the dominant scaling constraint.