광섬유를 이용한 AI 훈련: 빛 기반 도약

왜 광섬유가 AI 컴퓨팅에서 전기를 대체할 수 있는가

Since the early days of 컴퓨팅, almost all computers have been based on calculations using electricity in one way or another, from antique vacuum tubes to modern nanometer-scale silicon chips.

As silicon chips get smaller and smaller, researchers have been looking at new ways to build computers that could push our capacity further than silicon chips, a topic we explored in “Top 10 Non-Silicon Computing Companies”.

These methods include using different materials, like carbon carbide, vanadium dioxide, organic materials, or graphene, for example. Another way is to change how computing is done, moving away from the binary programming of electricity-based computing, which includes quantum computing and photonics.

Photonics uses light instead of electricity to encode and transfer information. However, until now, it has still been ultimately converted into a binary signal, failing to form a purely light-based form of computation.

This has changed with the work of researchers at the Tampere University (Finland) and Université Marie et Louis Pasteur (Besançon, France). They used optical fiber for ultrafast calculations and published their findings in the scientific journal Optics Letters¹, under the title “Limits of nonlinear and dispersive fiber propagation for an optical fiber-based extreme learning machine”.

전통적인 전자 시스템을 이용한 AI 훈련의 한계

AI training and data processing are reaching limits in terms of efficiency, with AI computation increasingly constrained by energy consumption and the speed of data processing.

In contrast, light-based calculations have the potential to be thousands of times quicker and can encode data into tiny differences of energy, making it more efficient. The issue is that so far, no direct calculation using light has been performed.

The researchers’ work used a particular class of computing architecture known as an Extreme Learning Machine, (ELM) an approach inspired by neural networks.

Among some of their advantages, ELMs can learn from the training data in one step and are a relatively simple algorithm.

As a rule, ELM is unlikely to be useful for very complex tasks requiring multiple layers of AI training, but can perform very well and more efficiently for specific tasks, like visual recognition for example.

연구자들이 광섬유를 이용해 이미지를 인코딩한 방법

The researchers used femtosecond laser pulses (a billion times shorter than a camera flash) and an optical fiber confining light in an area smaller than a fraction of human hair to build an optical ELM system.

The laser pulses are short enough to contain a large number of different wavelengths or colors, creating a rich dataset.

They then sent these data into the fiber with a relative delay encoded according to an image.

출처: Tampere University

AI 처리에서 비선형 광학의 역할

This form of data encoding was transformed by the nonlinear interaction of light and glass.

Linear optics is the regular optics taught in school, where the light directly interacts with a prism, for example.

In non-linear optics, the reaction of the medium in which the light passes depends on the light’s wavelength, intensity, direction, and polarization.

Nonlinear optical components can cause photons of different frequencies to combine and create new photons at new frequencies.

“전통적인 전자와 알고리즘을 사용하는 대신, 강렬한 레이저 펄스와 유리 사이의 비선형 상호작용을 활용해 계산을 수행합니다.” Mathilde Hary and Andrei Ermolaev – Post-Doctoral Researchers

Non-linear interaction and Extreme Learning Machine (ELM) algorithm was able to train an AI to classify handwritten digits (like those used in the popular MNIST AI benchmark).

출처: Optics Letters

The best systems reached an accuracy of over 91%, close to the state-of-the-art digital methods.

What makes the result exceptional is that it was achieved in under one picosecond, or one trillionth of a second (0.000000000001 seconds).

이상적인 최적화

The best results did not occur at the maximum level of nonlinear interaction or complexity.

Instead, they required a delicate balance between fiber length, dispersion (the propagation speed difference between different wavelengths), and power levels.

“성능은 단순히 광섬유에 더 많은 전력을 가하는 문제만은 아닙니다. 빛이 처음 어떻게 정밀하게 구조화되는지, 즉 정보가 어떻게 인코딩되는지와 광섬유 특성과의 상호작용에 달려 있습니다.” Mathilde Hary – Post-Doctoral Researcher

광섬유 컴퓨터가 AI의 미래가 될 수 있을까?

Training AIs with only light is a radical departure from all the methods used until now. This is likely not going to be a method possible to use for every type of data, but for the ones where it can be applied, this could bring results that are 1,000x more efficient in terms of energy, and as much as a million times quicker.

“우리 모델은 분산, 비선형성 및 양자 노이즈가 성능에 어떻게 영향을 미치는지를 보여주며, 차세대 하이브리드 광-전자 AI 시스템 설계에 필요한 핵심 지식을 제공합니다.” Andrei Ermolaev – Post-Doctoral Researcher

Most likely, such an approach would mean that some AI calculation would be delegated to a non-linear optical fiber hardware custom-built for the task. So repetitive tasks, like visual identification, would be the best candidates more than processing new data.

“이 연구는 비선형 광섬유 광학의 기본 연구가 새로운 계산 접근법을 촉진할 수 있음을 보여줍니다. 물리학과 머신러닝을 결합함으로써 초고속 및 에너지 효율적인 AI 하드웨어를 향한 새로운 길을 열고 있습니다.” Andrei Ermolaev – Post-Doctoral Researcher

Potential applications range from real-time signal processing to environmental monitoring and high-speed AI inference.

Such work is, however, still at the demonstration of basic principles of the technique stage, and far from a commercialization step.

It nevertheless demonstrates that photonics is likely going to be an increasingly important part of the computing industry moving forward, as light can be superior to electricity for some computing applications due to fundamental physics reasons.

상장된 주요 레이저 및 포토닉스 기업

Coherent (II-VI Marlow): 레이저 혁신을 선도하는 기업

Coherent는 26,000명 이상의 직원을 보유한 대형 산업 콘글로머리트이며 레이저 기술 분야의 선두주자입니다. 이는 첨단 소재 기업 II-VI (COHR ) Marlow와 레이저 제조업체 Coherent의 합병으로 탄생했습니다.

The company is an expert in advanced materials used in lasers, optics, and photonics, such as indium phosphide, epitaxial wafers, and gallium arsenide.

It grew largely thanks to multiple acquisitions over the last decade, from $600M in revenues in 2013 to $4.7B in 2024.

The company derives 29% of its revenues from lasers directly, with the rest linked to associated equipment like optical fiber, and electronics. The instrumentation category mostly includes life

sciences and medical applications.

출처: Coherent

The presence of the company in advanced materials like thermophotovoltaics (which we discussed in a previous article), silicon carbide, lasers, and electronics helps it benefit from structural trends like the growth of precision manufacturing, additive manufacturing (3D printing), electrification, and renewable energies.

The company has recently separated its silicon carbide business into a new entity, owned at 75% by Coherent, with the rest owned equally by its partners Mitsubishi Electric (bringing silicon carbide power IP) and Denso (bringing its activity as an automotive supplier on electrification and power semiconductors).

This is because silicon carbide is increasingly its own technology, mostly used in high-power applications like EVs, batteries, and renewable energy.

Coherent is a leader in LIDAR and 3D-digital sensing, including for self-driving applications, biotech Next Generation Sequencing (NGS) Flow Cells, and lasers for semiconductor manufacturing. It expects its main markets to grow at 8-20%.

출처: Coherent

The other potential new applications of lasers, like direct energy weapons, photonic computing, nuclear fusion, and spacetech, could all equally help sustain the long-term growth of the company.

Overall, Coherent is as close as it can get to a “pure play” publicly traded laser company for investors interested in the sector, with strong vertical integration and 3,100+ patents protecting its innovations.

As photonics progresses, it will progressively increase the demand for ultra-fast, ultra-precise laser systems, as well as lasers used in optical telecommunications.

최신 Coherent (COHR) 주식 뉴스 및 개발

참조 연구

^{1. Andrei V. Ermolaev, Mathilde Hary, et al. Limits of nonlinear and dispersive fiber propagation for an optical fiber-based extreme learning machine. Optics Letters. Vol. 50, Issue 13, pp. 4166-4169 (2025) https://doi.org/10.1364/OL.562186}