MicroLED optical interconnect targets GPU links to replace copper wiring

Kopin and Fabric.AI are collaborating on a MicroLED-based optical interconnect aimed at replacing copper wiring between GPUs and other high-performance processors in AI infrastructure. Fabric.AI has placed a $15 million purchase order with Kopin to fund a demonstration chipset for the effort.

The companies are calling the approach Neural I/o, built on Kopin’s MicroLED technology and its patented, bi-directional NeuralDisplay architecture. The core idea is to repurpose programmable MicroLED pixels as optical transceivers, using each pixel as a high-speed transmitter that sends digital bits optically for chip-to-chip and system-to-system connectivity.

Kopin and Fabric.AI describe the target as GPU-to-GPU data exchange at large scale while consuming “significantly less power per bit” than existing solutions. They also say the design is intended to avoid both copper interconnects and “expensive laser-based systems,” using photons instead of electrons to move data between devices.

For data center engineers, the practical promise is straightforward: if an optical interconnect can cut the power overhead of short-reach, high-bandwidth links, it can also reduce the heat that has to be removed from dense accelerator nodes. But the big gap between a demonstration chipset and production deployment is where most interconnect ideas succeed or die, especially once packaging, alignment, error rates, and manufacturability constraints show up.

Matt Kimball, principal analyst at Moor Insights & Strategy, said, “The two biggest challenges facing virtually every at-scale AI deployment are power and bandwidth,” adding that enabling connectivity “without taxing the power budget has been a persistent challenge.”

Under the companies’ agreement, Kopin owns 19.9% of Fabric.AI and will be the exclusive manufacturer of the Neural I/o chipsets. Michael Murray, Kopin’s CEO, said the interface is “expected to be uniquely capable of supporting GPU-to-GPU communication at the massive scale this market requires.” Josh Silverman, Fabric.AI’s CEO, said Kopin’s bi-directional MicroLED technology is “the foundation of our optical interconnect architecture.”

Kopin and Fabric.AI did not provide bandwidth, link distance, latency, or power-per-bit figures for the Neural I/o demonstration chipset in the information released.

Source: Kopin

Get Data Center Engineering News In Your Inbox:

Popular Posts:

Enphase-Energy's-new-solid-state-transformer-aims-to-deliver-1
Enphase Energy's new solid-state transformer aims to deliver 1.25 MW racks for AI data centers
Screenshot
Five AI data centers to reach 1 GW power capacity in 2026, new analysis shows
Xendee
Bring your own power: Why data center microgrids are shifting from backup posture to strategy
mike-nolan-PPST
How to burn-in test megawatt power equipment without drawing a megawatt from the grid
epub_Johnson-Controls-YORK-YK-HT-centrifugal-chiller_696x696_webp
Johnson Controls previews YORK YK-HT centrifugal chiller for data center water-free heat rejection

Share Your Data Center Engineering News

Do you have a new product announcement, webinar, whitepaper, or article topic? 

Get Data Center Engineering News In Your Inbox: