Cognifiber miniaturizes photonics for edge computing

Cognifiber, a company that builds stand-alone photonic artificial intelligence (AI) accelerators for IoT analytics, has its sights set on a new goal —  to replace traditional silicon with glass. By miniaturizing its in-fiber photonics, the startup claims it can bring server-grade capabilities to the edge.

The startup focuses on photonic computing, which goes a step further than most current commercial applications that use photonics just for communications across racks in the data center. Cognifiber has now announced a “first of its kind” glass-based photonic chip specifically created for edge computing.

The company claims it has been able to miniaturize its previous designs to “a fraction” of the size. The “rack-size” system has been reduced to a 4U server with a height of 18cm, making it feasible to deploy in offices, which  also leads to a lower cost.

Photonic glass chips as a solution

The goal of edge computing is to increase the compute capabilities at the edge, which would prevent IoT devices from having to rely on the cloud without impacting their performance, while also decreasing the response time (also known as, latency) for real-time operation. As some examples of where Cognifiber imagines its solution could be deployed, it cites smart meters and home assistants, connected vehicles, automated trains and fleet management of drones. 

“Combining photonic glass chips promotes our edge solution to bring rapid AI and Machine Learning locally to edge devices, which are limited in their capacity and power allowance,” said Ze’ev Zalevsky, cofounder and CTO of Cognifiber.

The company claims its in-fiber processing can boost performance by 100 times and decrease AI/ML training cost by 80%. Cognifiber’s technology is based on exploiting the crosstalk —which, traditionally, is meant to be avoided — between fibers. This allows for programmable and reconfigurable interactions, akin to field programmable gate arrays (FPGAs). Both linear and nonlinear operations are supported, making the technology suitable for AI.

The Israeli startup, which was previously part of Intel’s Ingenuity Partner Program, announced $6 million in series A last month.  It has demonstrated a “successful” proof of concept during its seed stage and aims to bring a full system to market by the end of 2023.