Two post-1995 entrepreneurs founded an optical computing chip company and developed the world's first optical chip integrating storage and computing.
Author | Wu Ruoyu
Editor | Peng Xiaoqiu
In 2022, two young people under 30: one closed his US company, and the other suspended his PhD studies at the University of Oxford. They both returned to China and founded a company called "Light Standard".
Three years have passed. Light Standard has completed the tape-out of the world's first optical computing chip with computing power density and precision meeting commercial standards. It has become the world's only optical computing chip company that has achieved in-memory computing integration and integrated a 128*128 matrix-scale optical chip onto a single die.
The one who closed his US company is Xiong Yinjiang, co-founder & chairman of Light Standard. He is now in charge of operations and commercialization. He holds a master's degree from the University of Chicago and has years of engineering experience in large model algorithms and AI agents. The one who suspended his PhD studies is Cheng Tangsheng, co-founder & CEO of Light Standard. He is now responsible for R & D and engineering implementation. Cheng Tangsheng once studied under Harish Bhaskaran, the world's leading figure in "phase-change material optical computing" and a fellow of the Royal Academy of Engineering.
The two co-founders of Light Standard. Xiong Yinjiang is on the left, and Cheng Tangsheng is on the right.
Xiong Yinjiang and Cheng Tangsheng were determined to start a business because they had been paying attention to the entrepreneurial opportunities in the field of optical computing for many years. Before starting the business, they had completed the tape-out and application verification of small matrix chips such as 3×3 and 4×4, accumulating experience in the entire process of tape-out, design, and materials. In 2022, the demand for AI exploded. Using the speed and advantages of light to innovate chip manufacturing might truly break through Moore's Law and solve the computing power anxiety. They thought the best time had come. Especially since this year, with the further rise of large models and generative AI, the industry's demand for computing power has extremely inflated. Not only in the model training itself but also in the subsequent practical reasoning stage, energy consumption is higher.
According to the "Energy and AI" report released by the International Energy Agency, in 2024, global data centers consumed about 415 terawatt-hours of electricity, accounting for 1.5% of the world's electricity consumption. It is estimated that by 2030, the electricity consumption of global data centers will more than double, reaching about 945 terawatt-hours, which will exceed the current total electricity consumption of Japan.
Meanwhile, the computing power of a chip is directly related to the number of transistors it contains. Currently, the 3-nanometer lithography process is approaching the atomic physics limit of 0.3 nanometers. To further improve the computing power performance of electrical chips in terms of process, the only way is to continue Moore's Law through advanced packaging such as stacking. However, this can only barely support, and it cannot solve the contradiction between energy consumption and the supply - demand of computing power in the long run.
Since 2017, academic discussions on optical computing at the University of Oxford and MIT have attracted domestic and international attention. People have found that in addition to stacking chip transistors, optical chips can perform signal processing using optical technology. They have the potential to increase computing power by more than 1000 times compared with electrical chips, with lower energy consumption and larger bandwidth.
In recent years, in addition to the government, the capital market, and computing power ecosystem enterprises paying attention to the progress of optical computing chips, many optical computing chip companies have emerged at home and abroad. Light Standard was established in this context. It focuses on the R & D of photonic in-memory computing chips based on the heterogeneous integration of silicon photonics and phase-change materials (PCM).
From the laboratory to commercial implementation
To promote the transition of optical computing chips from the laboratory to industrial applications, two core challenges need to be overcome: one is how to miniaturize light from the fiber scale to the chip scale; the other is to retain the chip's ability to process complex computing tasks while completing the size miniaturization, ensuring that it can meet the actual needs of scenarios such as large models and intelligent computing centers.
Light Standard's breakthrough began with an accurate judgment of the entry timing. "In 2017, although optical computing chips attracted attention due to their many advantages, the industrial chain links such as Fab manufacturing and packaging processes at home and abroad were not yet mature. Especially the shortcomings in key processes such as chip packaging made it difficult to promote the commercial implementation of optical computing chips," said Xiong Yinjiang, co-founder & chairman of Light Standard.
By 2022, with the explosion of AI and the gradual maturity of the overall industrial chain, design companies were needed to step forward and actively drive the production chain to accelerate iteration through demand, pressing the fast - forward button for commercial implementation. Therefore, Light Standard chose to enter the market at this time.
Compared with the regular rhythm of technology companies, Light Standard's productization process has advanced significantly faster.
After the company was established, it quickly completed the iteration of matrix scales such as 16x16 and 25x25. In 2023, it achieved the tape-out of a 64x64 matrix scale and made a breakthrough in the 128x128 matrix scale in June of the following year.
Optical computing chip with a 128*128 matrix scale
However, the tape-out of the 128×128 matrix is only the first step to touch the threshold of commercialization. "Simply competing in matrix values cannot build a moat. The nodes of many solutions are fixed values and can only be used for specific calculations, unable to adapt to different models," Xiong Yinjiang emphasized. The value of Light Standard's chips also lies in creating the first architecture with "more than 16,000 nodes and each node fully adjustable".
For scenarios such as large models, 128×128 is the "economically feasible" scale for current model splitting. The real - time adjustability of nodes allows the chip to adapt to the parameter changes of any model, opening up a key space for the programmability of matrix operations - this is the core symbol of the transition of optical computing chips from scientific research to commercial use.
What supports this breakthrough is Light Standard's unique choice of technical route: the heterogeneous integration of silicon photonics and phase-change materials and the unique Crossbar photonic matrix computing structure, which provides the core solutions to the two problems of "scale miniaturization" and "balance between computing power and energy consumption".
Demonstration of the Crossbar technical route
"The programmable computing unit of the traditional pure silicon architecture is large in size. At most, hundreds of units can be placed on a standard chip area. The computing power is only a single - digit Tops at the standard GHZ main frequency, far from meeting the needs of AI. More importantly, pure silicon materials need continuous power supply for 24 hours to maintain their state during computing. This part of the power consumption accounts for 70 - 80% of the total system power consumption, posing high requirements for heat dissipation and going against the trend of the era of exploding computing power demand," Xiong Yinjiang introduced.
Through research, Light Standard used phase-change materials to reduce the unit size to 1/10 - 1/20 of the traditional solution, and the accommodated matrix scale increased by 10 - 20 times. Regarding the energy consumption problem of AI chips, Light Standard also made optimizations at the design stage.
Now, the chip based on the heterogeneous integration of silicon photonics and phase-change materials only needs a single pulse of extremely low energy to complete parameter refreshing. And due to the non - volatile characteristics of phase-change materials, it does not need continuous energy to maintain the state, achieving nearly zero power consumption.
In addition, based on the storage function of PCM phase-change materials, Light Standard completed the integration of storage units and computing units, realizing in - memory computing with integrated storage and computing, further alleviating the storage pressure and improving chip performance.
How to sell optical chips?
Entering the commercialization stage, more practical requirements are placed on technology companies. Currently, Light Standard faces two types of customer groups in cloud - side scenarios, with significantly different but clearly - defined demands.
End - demanders represented by Internet giants are also the "final battlefield" of the computing power ecosystem. These customers have mature computing power clusters and large - model R & D capabilities. For them, if optical computing chips can significantly exceed in core indicators such as energy efficiency ratio and computing power density, even if the upfront investment is high, there is clear commercial rationality. Moreover, the model iteration speed is fast, and hardware needs to be ahead of "software" to be compatible with the new - generation software ecosystem.
The other type of cloud - side scenario customers are government - led intelligent computing centers, whose demands are relatively standardized, and they value the economic benefits of projects more. The differences in the demands of the two types of customers are essentially the concretization of technological value in different scenarios, which precisely constitute a complete market coordinate system for optical computing chips to transition from technical verification to large - scale commercial use.
The optoelectronic fusion computing card that Light Standard is currently developing can use the PCIe interface or other general standards for data interaction. It can be used plug - and - play and provides upper - layer application interfaces and an ecosystem that are easy to be compatible with. It has a high energy efficiency ratio and high algorithm flexibility, suitable for different application scenarios.
"In the future, our products will also be verified in edge - side scenarios. The differences in customer demands between cloud - side and edge - side scenarios are even greater. The core demands in cloud - side scenarios are high energy efficiency ratio and algorithm flexibility, and key efforts need to be made in software ecosystem construction to optimize the adaptability of hardware and algorithms. Edge - side scenarios emphasize stability. Especially for automotive - grade applications, they need to deal with complex environments such as extreme temperatures, and higher requirements are placed on hardware reliability," added Cheng Tangsheng, co-founder & CEO of Light Standard.
Photonic AI computing demo hardware after preliminary packaging
Currently, Light Standard has reached cooperation with domestic first - tier Internet giants, GPU manufacturers, intelligent computing centers, and universities, waiting for further application verification. At the same time, it is conducting in - depth technical exchanges with satellite manufacturers, communication operators, and optical module manufacturers around applications.
In terms of packaging and testing, Light Standard established cooperation with Fab manufacturers early on and jointly developed advanced packaging processes with the manufacturers to promote technological iteration. This year, Light Standard will complete the co - packaging and testing of the 128×128 matrix optical computing chip and electrical chip, forming the first - generation optoelectronic fusion computing card. At the same time, it will promote the tape - out of the 256×256 matrix optical computing chip. It is expected that the more competitive second - generation optoelectronic fusion computing card will be launched in 2026.
When the new product is about to be launched, Light Standard completed a new round of financing again in June this year. Dunhong Asset led the investment, and state - owned funds such as Pudong Science and Technology Angel Mother Fund, Suzhou Future Angel Industry Fund, and Zhangjiang Science and Technology Investment followed the investment. The old shareholder Zhongying Venture Capital increased its investment again. Among them, Zhongying Venture Capital is committed to discovering and cultivating early - stage startups in the hard - technology field with great competitive potential in the new era. The management funds come from the self - owned funds of Zhongying Innovation Group, with a total scale of 2 billion yuan. It is an ever - green fund for circular investment. It participated in the investment in Light Standard at an early stage with foresight and increased its investment multiple times.
In fact, Light Standard also completed two rounds of financing in 2024. One was in March, completing the Angel + round of financing. The other was in December of the same year, completing the financing led by Jinqiu Fund. The old shareholders Zhongying Venture Capital, Mushi Capital, and Xiaomiao Langcheng all over - subscribed for follow - on investments.
In the post - Moore era, to achieve a catch - up in optical computing chips and promote the full - process breakthrough of chips "from Demo to product - level", it is necessary to tackle the tough problem of advanced packaging for optical computing and build the relevant industrial chain.
Image source on the homepage | Provided by the enterprise
Typesetting | Liu Siya
This article is from the WeChat official account "Hard Krypton". The author is Wu Ruoyu, and it is published by 36Kr with permission.