TSMC Bets on Unorthodox Optical Tech

https://news.ycombinator.com/rss Hits: 22
Summary

In the race to an all-optical AI data center, a major player has now placed a bet on a different horse. Semiconductor manufacturing giant TSMC announced that it will work with Sunnyvale startup Avicena to produce microLED-based interconnects. The technology is a pragmatic twist on replacing electrical connections with optical ones to meet the high needs of communication among an increasing number of GPUs in a low cost, energy efficient way. Thanks to the computational demands of large language models and their cousins, AI clusters are facing unprecedented requirements regarding amounts of data, bandwidth, latency, and speed. Sooner or later, the copper wires that connect processors and memory within a single AI data center rack will have to be replaced with optics. “There’s a huge push to get optical connections as close to the board as possible,” says Lucas Tsai, a vice president at TSMC.Avicena offers a unique approach, using hundreds of blue microLEDs connected through imaging-type fibers to move data. The company’s modular LightBundle platform avoids problems with lasers and their associated complexity that threaten the reliability, cost, and power consumption of other optical chiplets. Tsai says “it’s very unorthodox!” But it is ideal for these short distance applications, and that’s precisely what makes it interesting.Laser-freeOptical connections today carry vast amounts of data tens to hundreds of meters across data centers at very high data rates. Traditionally, a pluggable module connects the optical fiber to the rack, where it converts between electrical and optical signals. Companies are making strides toward getting rid of these energy inefficient pluggable transceivers using co-packaged optics (CPO), which instead perform electrical-optical transformations adjacent to the silicon chip itself. Commercial versions exist for the network switch, and prototypes are making strides toward the GPU. The most prominent optical chiplet designs encode electronic b...

First seen: 2025-05-26 18:48

Last seen: 2025-05-27 15:56