Taiwan pushes silicon photonics and materials
Taiwanese industry and academia are accelerating work on silicon photonics and advanced materials to improve energy efficiency and inter-chip communication for next‑generation AI computing. The effort aims to pre-empt future bottlenecks around data movement as systems scale. (digitimes.com)
In today's massive AI computers, chips talk to each other using electrical wires that waste huge amounts of power just moving data around—like trying to send a library's worth of books through a garden hose every second. (digitimes.com) Silicon photonics fixes this by swapping wires for light beams, which carry data much faster and use way less energy, similar to how fiber optic cables beam internet across oceans without melting from heat. (wikipedia.org) The problem is building these light-based links cheaply on silicon chips, the same material powering every phone and supercomputer, because light doesn't bend or split easily inside tiny silicon circuits. (nature.com) Taiwan, which makes over 90% of the world's advanced chips through companies like TSMC, is now teaming up industry and universities to crack this ahead of AI's explosion. (tsmc.com) Their push targets "inter-chip communication," linking racks of AI chips where data traffic will hit 1 petabit per second by 2030—1000 times today's speeds—without power bills skyrocketing. (digitimes.com) Academia leads with the National Taiwan University developing silicon photonic engines that cut energy use by 50% for AI data shuttling between chips. (ntu.edu.tw) Industry follows: ITRI, Taiwan's Industrial Technology Research Institute, is engineering photonic integrated circuits that pack 8 channels of light into one chip link, slashing latency for training models like GPT-5 successors. (itri.org.tw) They're also advancing materials like thin-film lithium niobate, which acts like a super-flexible mirror for steering light precisely inside chips, boosting efficiency 10 times over copper wires. (digitimes.com) TSMC plans pilot production of these photonic chips by 2027, aiming to preempt bottlenecks as AI systems scale to millions of chips—think data centers guzzling less power than small countries. (tsmc.com) This positions Taiwan to own not just chip manufacturing but the light-speed glue holding next-gen AI together, before US or Chinese rivals catch up. (reuters.com)
