A New Approach to Sustainable AI: Offloading Part of the Computational Burden Onto Light
Koç University’s Uğur Teğin and his students have built two systems that let light do part of an AI model’s image work: a colour-preserving optical processor and a fibre-laser cavity that classifies images with a few thousand trainable weights. We read both papers, the preprints and the peer-review files. The fibre-laser rig draws 32.5 W and spends about 0.54 joules per image, less efficient than an Nvidia A100 today; its 230 TOPS per watt figure assumes a modulator that has not been built in.
