CogniFiber: Photonic Computing

Computing with light has many advantages. It’s fast, cheap, and it doesn’t generate much heat. But it’s also difficult to make it work, and it has a number of challenges that are specific to photonics.

CogniFiber, an Israeli startup, says it has solved many of these issues. The company uses photonics over fiber, and it has introduced a glass-based chip.

“We use light as a data conveyor vehicle,” said Eyal Cohen, founder and CEO of CogniFiber. “We can represent data by pulses of light, in fibers or otherwise, and do it very quickly. Similar to optical communication (PAM4/PAM8), we use the amplitude of the light pulses traveling through fibers and glass chips to represent the data in each phase of the computation, as we do not use any dynamic memory read/write while computing. That means the input data rate determines the total throughput of the system, accelerating the computing immensely.”

Until now, the digital world has always been driven by electrons, which flow through wires at high rates of speed. But as wire dimensions shrink at each new process node, more energy is required to push those electrons through the wires. With increased resistance and capacitance, along with higher frequencies, it’s more difficult to maintain signal integrity due to cross-talk, as well as interference.

But data also can be routed using pulses of light, which typically are converted from electrons to photons, and then back again into electrons, where it is stored in memories. This is inefficient, although it does to reduce heat inside a device. CogniFiber’s approach, in contrast, is purely photonic. There is no conversion back and forth.

“The question that is raised in computing is, ‘Why not?'” said Cohen. “Instead of converting it to electronics, why don’t we leave it as a pulse of light, and use this pulse of light to compute?”

HPE considered that approach a decade ago, and research continues across the chip industry. CogniFiber, in particular, is focusing on building a computational system with a structure similar to that of the brain, where a singular nerve isn’t carrying all the weight, but all parts are working in harmony.

“We use the neuromorphic approach,” Cohen said. “It’s a very distributed system of computation.”

The company says its core technology, known as DeepLight, can transfer data as fast as 200 gigabytes per second using laser modulators or transceivers.

To facilitate DeepLight, the company developed a chip made entirely of glass. That chip is less structurally complex than a typical silicon chip, and works in conjunction with the fibers the company is using.

The chip, along with CogniFiber’s miniaturization capabilities, eliminates the silicon supply chain issues with a lower-cost option. Cohen said there likely will be multiple smaller suppliers for the glass chip.

There are drawbacks to this approach, though. Photonic channels can suffer from the same kinds of line-edge roughness as other interconnects, and heat can cause drift. In addition, light is harder to stabilize than electrons.

On top of that, these devices are not small. “The photon is larger than an electron by several orders of magnitude,” Cohen said. “A transistor is much smaller for electrons than a compartment for a photon. By comparison, the minimal size of a photonic element on a current photonic chip (about 100μm²) is about 30,000 times larger than that of a current electron-bearing transistor made in 5nm technology.”

That means transistors are unable to handle the large sizes of photons, so more transistors are required. And when there are too many transistors, reliability becomes an issue due to interference. These challenges have led companies in the past to abandon photonic computing.

This leads to a scaling issue, as chips that needed to implement millions of photonic elements will reach sizes beyond 100cm². In addition, the connectivity area — the total size of light-conducting passages — grows exponentially with the number of fully connected elements. Passing this threshold makes these chips expensive and vulnerable to production defects that decrease their yield due to high chance of production failure.

CogniFiber believes its product will find traction in Anomaly Detection initially, before ultimately making its way into automotive and other verticals. The company plans to launch its alpha prototype this month.

The company has raised $8.5 million in its first two rounds of funding, and is preparing for an additional round in coming months.

Related
Find more startup news and funding reports here.