Breakthrough in AI energy efficiency: Ultrafast optical chip for ChatGPT
Artificial intelligence systems like ChatGPT are known for their high energy consumption. To address this, a team from the Center for Optics, Photonics, and Lasers (COPL) has developed an innovative optical chip capable of transmitting vast amounts of data at ultrafast speeds while delivering unprecedented energy efficiency.
The core of this breakthrough, published in Nature Photonics, lies in harnessing the power of light for information transfer. Unlike traditional systems, the new chip utilizes not only the intensity of light but also the phase shift of optical radiation, adding a new dimension to the signal. This has enabled an increase in data transfer speed from 56 gigabits per second to 1 terabit per second—nearly 18 times faster! Such speed offers immense potential for training AI models: an entire training dataset, equivalent to over 100 million books, could be transferred in less than seven minutes, consuming only 4 joules of energy. The key to this technology lies in micromodulators—tiny silicon devices that manipulate light.
This new development promises a revolution for modern AI data centers, which currently occupy vast areas and consume significant energy. Thanks to the new chip, devices will be able to communicate quickly and efficiently, as if they were just meters apart, significantly reducing infrastructure and energy costs. This technology could enter the market in the coming years, with major companies like NVIDIA already beginning to adopt micromodulators. As researcher Alireza Geravandi notes, this development elevates a decade of laboratory work to a new level, bringing its real-world application much closer.
ORIENT
