A prototype of an IBM optical module for connecting chips with optical fiber
Ryan Lavine for IBM
A fiber-optic technology can help chips communicate with each other at the speed of light, allowing 80 times more information to be transmitted using conventional electrical connections. This can significantly speed up the training times required for large artificial intelligence models (from months to weeks), while reducing data center energy and emissions costs.
The most advanced computer chips still communicate using electrical signals carried on copper wires. But as the tech industry races to train large AI models, the process requires networks AI superchips to transfer large amounts of data – companies are eager to connect chips using light-speed fiber optic communication.
This technology is not new: the Internet already relies on fiber optic cables spanning thousands of kilometers between continents. To transmit data between the fingernail-sized chips, however, companies must connect as many hair-thin optical fibers as possible to the edge of each chip.
“As we all know, the best communication technology is fiber optic, and that’s why fiber optic is used everywhere else for long distance communication,” he said. Mukesh Khare At an IBM Research technology forecast press conference. “This innovation in copackaged optics essentially brings the power of fiber optics on a chip.”
Khare and his colleagues have developed an optics module that would allow chipmakers to add up to six times more optical fibers to the edge of a chip compared to current technologies. The module uses a structure called an optical waveguide to connect 51 optical fibers per millimeter. It also prevents light signals from one fiber from interfering with those around it.
“What IBM has really done here is use all of their materials and packaging technology, that history of leadership, to really disrupt how you make high-density fiber optics using waveguides,” he says. Dan Hutcheson At TechInsights, a semiconductor technology research company headquartered in Canada. “For me, that was when I saw the big breakthrough.”
As a result, the push for chip-to-chip communication allows AI developers to train a large language model in three weeks instead of three months. Switching from electrical wires to optical fibers for on-chip communication could also mean a fivefold reduction in the energy cost of training those AI models.
IBM has already subjected the optical module to high humidity and stress testing from -40°C (-40°F) to 125°C (257°F). Hutcheson expects that boss semiconductor manufacturing companies may be interested in licensing the technology.
“We’re in the early days of all of this, but it’s the hottest area of semiconductor technology right now in terms of high-performance computing and AI technology,” he says.
Topics:
