crossorigin="anonymous">

IBM’s co-packaged optics: Adds more bandwidth to the connector.


On December 9, IBM unveiled a breakthrough in co-packaging optics that can reduce and accelerate GPU downtime. AI training.

IBM’s working prototype significantly increases the number of optical fibers attached to the edge of a silicon photonics chip by replacing conventional glass with polymer optical waveguides. This innovative approach could pave the way for faster, more efficient data transfers and allow data centers to handle heavier workloads.

The idea of ​​co-packaged optics has been around for a long time, but IBM’s process enables what the company calls the world’s first stress-tested prototype.

An IBM spokeswoman did not directly address questions about availability or timelines. Instead, they emphasized their intention to develop a roadmap and expressed openness to sharing design materials with foundries in the future.

“Ultimately, the chip product companies will have to ask for it, and then the product companies will design it into their chip designs, and the foundries can produce it,” Mukesh Khare, general manager of IBM Semiconductors, said in a briefing. ” “But there’s no specific need for the foundry aspect. There will be a design aspect that we can provide to the chip companies.

What characterizes a polymer optical waveguide?

Co-packaged optics with polymer optical waveguides are an alternative to copper connections and are often used to connect GPU accelerators in data centers. They sit on the edge of a chip and allow many high-density bundles of optical fiber to be squeezed into a small space with half a micron or less between the fiber and the connector. IBM said this provides a dramatic bandwidth boost between chips compared to electrical connections.

Size – a 50 micron pitch – also differentiates the prototype. 250 micron pitch is a standard size. Going smaller translates to increased bandwidth.

Polymer optical waveguides are stacked in four layers for up to 128 channels. At the “beach front” where the connector meets the chip, it offers 51 fibers per millimeter.

The waveguide lid and polymer optical waveguide are integrated into a chip. Image: IBM

“The great thing is not only that we have achieved such a large density increase for communications on the module, but we have also demonstrated that it is compatible with stress tests that optical links have not undergone in the past,” John Knickerbocker, distinguished engineer at IBM Research, said in a press release.

“This co-packaged optics innovation is essentially bringing the power of fiber optics right on a chip,” added Khare.

IBM’s polymer optical waveguide can compete with new connectivity processes such as the Ranovus Odin electronic and photonic integrated circuit or linear drive pluggable optics. Researchers are also experimenting with glass ribbons or vertical hierarchical interconnections in this area.

Knickerbocker said in the briefing: “It’s hard to say who’s ahead” between polymer optical waveguides and linear drive pluggable optics.

IBM has developed co-packaged optics with polymer optical waveguides at its Bromont testing facility in Quebec.

See: Data Centers Will Need More Power for AI Training hyperscalers offer More advanced models.

Proposed Market: Data centers used to train AI.

IBM suggests that the new connector could benefit. The evolving AI industry by:

  • Power draw reduction (up to 5 times) for medium-range electrical interconnects, including long ranges (hundreds of meters).
  • A reduction in the time it takes to train a large language model, from three months to three weeks.
  • Increased energy efficiency.

“With this breakthrough, tomorrow’s chips will communicate in the same way that fiber-optic cables carry data in and out of data centers, ushering in a new era of faster, more sustainable communications that will enable the AI ​​of the future.” can handle the workload,” SVP and director of research Dario Gil said in a press release.



Source link

Leave a Reply

Translate »