IBM Introduces Power of Light to Chips, Increasing Speed by 80 Times

On Monday, IBM announced a breakthrough in optical technology that may significantly improve the way data centers train and run generative artificial intelligence (AI) models. Its new Co-packaged optics technology has essentially brought the power of optical technology into the chip. And that enables connections within data centers to reach the speed of light.

At a media briefing, IBM Semiconductor General Manager Mukesh Khare told reporters that the telecommunications industry has made significant progress in manufacturing faster chips. But the communication speed between these chips has not kept up with this pace. There is a difference of several orders of magnitude between the growth rate of computing power and the growth rate of communication between chips.

He pointed out, “Actually, at a more fundamental level, chips still communicate through electricity, using copper wire. And we all know that the best communication technology is fiber optic. And that is why fiber optic is used everywhere in long-distance communication.”

Although Co-packaging optical technology has existed for some time, IBM has developed a new type of polymer optical waveguide (PWG) technology, which has brought innovation to co packaging optics. PWG technology enables chip manufacturers to deploy six times more optical fibers in the edge area of silicon photonics chips, a precious space known as the “golden zone”. The width of each fiber optic cable is approximately three times that of a human hair. And its length can extend from a few centimeters to hundreds of meters, and it can transmit data at a speed of terabits per second.

What does all of this mean?

The company stated that this technology will enable communication bandwidth between chips to be 80 times faster than when electronic technology is currently used, and reduce energy consumption by more than five times.

In addition, it can increase the training speed of large language models (LLMs) by up to five times. It reduces the time required to train standard LLMs from three months to three weeks, and increasing performance gains with larger models and more GPUs.

In addition to enabling faster communication between GPUs and accelerators, this technology may also redefine the way the computing industry transmits high bandwidth data through circuit boards and servers.

Khare said, “We are very excited to introduce the power of light. And that fundamentally accelerates the development of the artificial intelligence world and many other applications.”

When asked when this technology will be commercialized, Khare stated that IBM’s research department is “ready to put it into use”.