CAMBRIDGE, Mass. — As global energy consumption from artificial intelligence datacentres commands urgent international scrutiny, a consortium of optical physicists and microchip architects has unveiled an operational photonic tensor processing unit that computes mathematical matrix multiplications at the speed of light. Operating at room temperature without cryo-cooling, the optical processor delivers 100-fold gains in energy efficiency compared to modern silicon GPUs.
Computing at Light Speed Without Resistance
Traditional semiconductor processors rely on electric currents pulsing through nanoscopic copper interconnects, generating massive resistive heat and requiring gigawatts of cooling infrastructure. In contrast, the photonic architecture splits laser beams across microscopic silicon waveguides. Phase modulators alter the amplitude of the light, allowing billions of complex calculations to occur instantaneously as the light waves intersect.
“Electrons have mass and resistance; photons do not. By replacing electrons with coherent light for heavy matrix math, we have fundamentally decoupled computing horsepower from catastrophic electrical consumption.”
— Prof. Alistair Chen, Lead Author, Photonic Systems Laboratory
Compatibility With Existing Semiconductor Fabs
Crucially, the consortium confirmed that their optical chips are manufactured using standard 300mm complementary metal-oxide-semiconductor (CMOS) foundry lines in Taiwan and Germany. This allows existing high-volume microchip fabrication plants to produce the optical accelerators without requiring hundreds of billions of euros in specialized tooling.
Implications for the AI Infrastructure Sector
Major cloud infrastructure providers and autonomous vehicle manufacturers have already initiated early integration testing, forecasting that photonic coprocessors will enter commercial enterprise datacentres by the second half of 2027.