news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

The temperature in the optical computing chip industry is changing.

Author｜Octave

Editor: Wang Bo

"Integrated optical circuits will be a once-in-60-year opportunity in the semiconductor field to overtake others."

At the end of last year, the last paragraph of the preface of the "Photonic Era: White Paper on Photonic Industry Development", released at the 2023 Global Hard Technology Innovation Conference, contained the following sentence.

There is no doubt that optical computing chips have advantages in computing power and data transmission, and the research results of domestic teams are often seen in the newspapers, but when it comes to large-scale commercial use, many people still have doubts. In 2023, a graduate student engaged in the research of micro-nano photonic devices posted a post on Zhihu to discuss issues related to optical computing chips and bluntly said: "Since it is not commercially available, there must be big problems."

So, has the optical computing chip emerged from the paper?

Matrix scale (computing power density) and single-node optical recognition (computing power accuracy) are key indicators for measuring the performance of optical computing chips. The industry-recognized matrix scale that meets commercial standards is 128×128. In 2021, two companies worldwide completed the tape-out of 64×64 optical computing chips. This bottleneck has not been broken through in the three years since then, and some companies have even switched to other tracks.

During the recent 2024 World Artificial Intelligence Conference (WAIC), "Jia Zi Guang Nian" learned thatOpto-based Technology, a domestic optical computing chip company, has completed the tape-out of the first optical computing chip with both computing power density and computing power accuracy reaching commercial standards.The matrix scale of this chip is 128×128, and its peak computing power exceeds 1700TOPS.

The optical computing chip with a matrix size of 128×128 developed by Opto-Based Technology. Image source: Opto-Based Technology

"Our current benchmark is Nvidia's A100," Cheng Tangsheng, co-founder of Opto-Based Technology, told Jiazi Guangnian. "The 128×128 chip will be made into a board, and its final computing power will be comparable to Nvidia's electronic chip products. At the same time, the power consumption of our first-generation products is about one-tenth to one-fifteenth of theirs. In addition, in terms of latency, optical computing chips will also be orders of magnitude lower."

The 128×128 matrix scale means that domestic optical computing chips are one step closer to industrialization, rather than just "staying on paper" or "living only in the laboratory."

Xiong Yinjiang, founder of Optical Benwei Technology, said that the company is currently debugging a 128×128 optical computing board and is expected to launch commercial optical computing board products within 2025, empowering large models, AI computing hardware, intelligent computing centers, embodied intelligence and other industries with higher energy efficiency and greater computing power. At the same time, the company is about to complete the research and development of optical computing chips with a larger matrix scale.

"I think we are at a critical point," an investor who focuses on the chip industry told Jia Zi Guang Nian. "People have changed from initially believing that optical computing chips are completely unrealizable to believing that it is possible."

The temperature in the optical computing chip industry is changing.

1. “Do you believe in light?”

"Do you believe in light?"

This Internet buzzword is also a true reflection of the optical computing chip industry.

Over the past 50 years, the computing power of chips has continued to grow in line with Moore's Law. In the past, the industry was more accustomed to using the computing power of electric chips, but the computing power of electric chips is currently facing a growth bottleneck.

For example, the quantum tunneling effect limits the further reduction of transistor size; as performance increases, energy consumption and heat accumulation place higher requirements on chip heat dissipation; signal delays, internal chip transmission speeds and data transmission bottlenecks limit processing speeds; and due to scalability issues, the number of transistors under the existing technical framework is close to an upper limit.

Supply is facing bottlenecks, while demand is still increasing. Zhang Jianzhong, founder and CEO of Moore Threads, said at the "AI Creation Era - 2024 Jiazi Gravity X Technology Industry New Trend" conference held by Jiazi Light Year in May this year: "Most AI startups find it difficult to build computing centers even if they have raised billions or hundreds of millions of dollars. Coupled with the scarcity of computing resources, many companies are willing but unable to do so. The problem facing the entire industry, including OpenAI, is how to solve the iteration problem caused by the lack of computing resources."

How to break the deadlock? Optical computing chips provide a new idea.

Optical computing chips are an application form of photonic chips, focusing on the optical implementation of computing tasks.It has the characteristics of ultra-high speed, ultra-strong parallelism, ultra-high bandwidth and ultra-low loss.

In fact, scientists have never stopped exploring light for decades. In the 1940s, Fourier transform was introduced into optics and Fourier optics was developed. 1980 to 2004 was the golden age of optical computing. During this period, researchers proposed a large number of optical computing technologies, including optical pattern recognition, logical optical computing, optical neural networks, optical interconnection and optical holographic storage.

However, due to unclear application requirements and scenarios for optical computing and the lack of matching optical hardware systems, the development of optical computing technology has been relatively slow.

In recent years, with the explosion of the artificial intelligence industry, optical computing chips have also entered a period of rapid development.

In 2021, the total valuation of global photonic products and related services reached more than US$7 trillion, accounting for approximately 11% of the world's total economy.

Now, optical computing is an emerging industry with a wide range of professional involvement and extremely high theoretical and technical requirements.It is seen as a potential area where China can start on the same starting line as its international competitors and even surpass them technologically.

2.128×128: ​​Double breakthrough in computing power and power consumption

Despite the high hopes for optical computing chips, optical computing has been a technology trapped in the laboratory for a long time.

In 2017, the first Nature article appeared in the field of optical computing, and then some companies were established. However, the entire industry chain is not mature enough, and there are not many reference examples from packaging to electronic chip coordination. In 2022, the year when Opto-Benwei Technology was founded, the industry has developed for five years, and the technology and industry chain have gradually matured, but optical computing chips have not been commercialized on a large scale like traditional electronic chips.

Xiong Yinjiang believes that there are two main reasons for this: first, the design of optical computing chips themselves was immature in the past and could not be directly compared with the current leading electrical solutions; second, the industrial chain for turning them into products is immature, and optoelectronic integration is ultimately a very challenging engineering task.

“The optical computing chips in many domestic universities are of a smaller scale, such as 3×3 or 4×4. Some start-ups can only achieve 64×64 at most, which means they do not have commercial capabilities and are still in the research and development stage,” Li Gang, head of early projects at Frees Capital, told Jia Zi Guang Nian.

This time, the optical computing chip with a matrix size of 128×128 from Opto-Base Technology was successfully taped out.Domestic optical computing chips have reached the critical point of commercialization.

To achieve large-scale commercial use, optical computing chips need to solve the problems of nonlinear computing and storage-computing integration. Both scientific research institutes and the industry believe that building an optoelectronic integration ecosystem is a must. Opto-based Technology has achieved storage-computing integration based on PCM phase change materials, with storage units and computing units fully integrated. It has now iterated the ability to design electrical chips with optical computing chips as the core, and has established in-depth strategic cooperation with top domestic packaging companies to jointly develop advanced optoelectronic packaging capabilities.

"The 128×128 matrix scale is comparable to the architectures used by Google's TPU or Huawei's Ascend. For us, optical computing is equivalent to electronic computing in terms of computing power density and matrix scale, which was not possible with optical computing before, and we have done it for the first time," said Xiong Yinjiang.

The prototype of the photon storage and computing integrated machine of Opto-Based Technology. Image source: Opto-Based Technology

The matrix size of the optical computing chip corresponds to the computing power.

For example, when the matrix size is increased from 64×64 to 128×128, the computing power increases by at least four times, but the increase in power consumption is linear.

Specifically, at a 64×64 matrix scale, the chip is equipped with 64 analog-to-digital converters (ADCs) and 64 digital-to-analog converters (DACs). When the matrix scale reaches 128×128, although the computing power increases fourfold, the number of ADCs and DACs required only increases to twice the original, which means that the computing efficiency is at least doubled.

At the same time, the accuracy of this optical computing chip has also achieved a breakthrough, reaching the standard accuracy of AI reasoning.

More importantly, each unit in the 128×128 matrix is ​​adjustable, making the chip a completely universal platform capable of handling any type of AI computing task, in stark contrast to other optical computing chips that have fixed weights and can only handle a single application.

Such technological breakthroughs did not happen overnight.RatherThe result of ten years of technological accumulation.

Since 2014, the Oxford laboratory has pioneered the research and development of phase change materials combined with silicon photonics, and successfully developed the world's first photonic storage and computing integrated chip in 2017. While studying for a doctorate at Oxford University, Cheng Tangsheng studied under Harish Bhaskaran, the world's first person in "phase change material optical computing" - Professor of the Department of Materials at Oxford University and Fellow of the Royal Academy of Engineering. He also led and participated in the research and development of phase change material optical computing chips and new ultra-low power nano phase change materials at Oxford University.

Cheng Tangsheng told "Jia Zi Guang Nian" that since he started studying for his doctorate at Oxford, he has been working on solving problems in the practical applications of optical computing technology for many years, studying how to reduce power consumption and expand scale in order to meet the computing power needs of applications such as the development of artificial intelligence.

The founding team of Opto-Based Technology has made continuous breakthroughs in the field of optical computing. Starting a business in 2022 is even more "well prepared".

"From the beginning of the company, we have realized that the existing optical computing chip matrices on the market are small in size, which limits their computing density and scale. Therefore, from the first day of the company's establishment, we have made it clear that we must first increase the matrix size from 64×64 to 128×128, and then further expand it to 256×256 and 512×512." Cheng Tangsheng told "Jia Zi Guang Nian".

Last year, Opto-Electronic Technology used previous-generation optical computing chips to complete the full operation of the board architecture, including packaging, integration with electrical chips, and running general AI networks such as ResNet-50.

This year, Opto-Based Technology further optimized its chips. On the one hand, it increased the matrix size to 128×128 to improve computing density and scale; on the other hand, it optimized and simulated each independent optical device and carried out targeted tape-out to achieve comprehensive optimization of the entire system.

"We basically conduct tape-out every three months to six months. I believe that only by quickly producing products and continuously iterating them can the company be truly stable," said Xiong Yinjiang.

This time, the 128×128 optical computing chip is the result of three product iterations.

So, where are the commercial scenarios for such optical computing chip products?

3.Who needs optical computing chips?

For the application layer, optical computing chips will provide them with support for higher computing power and lower power consumption.

Higher computing power means that optical computing chips can process more complex models, and more complex models can give rise to better applications; lower power consumption means that both the cost of training models and the cost of inference in actual use by users will be greatly reduced. Once optical computing chips are commercialized, they will be a ray of hope in the post-Moore era, forming a virtuous circle.

Every revolution in technology application stems from a breakthrough in computing power. It is the improvement in computing power that has enabled the disruptive moments of artificial intelligence to emerge.

Therefore, the primary scenario for optical computing chips is artificial intelligence. "With the rapid development of artificial intelligence, the application and demand for this scenario have become very clear. Optical computing chips can quickly adapt and optimize for current large-scale model-related applications," said Li Gang.

In the field of large models, optical computing will help downstream customers reduce costs and improve efficiency. It can reduce the future inference cost to one thousandth or one ten-thousandth of the current cost, making it almost cost-free for users to access large models, thereby significantly improving the gross profit margin of large model companies.

Reducing computing power costs is very important for the implementation of large model applications. Otherwise, companies will have to worry about expenses every day, thereby reducing the model scale and reducing training quality.

A founder of an artificial intelligence startup told Jiazi Guangnian that lower computing power costs will help the development of application products. "I think now everyone should gradually understand more clearly that the more tokens are consumed, the better, rather than trying to make an application that saves tokens as much as possible."

at the same time,The significant increase in the computing power of optical computing chips can also provide powerful power for current popular technology scenarios such as artificial intelligence, big data analysis, and autonomous driving.

Xiong Yinjiang introduced that in the field of autonomous driving, the current L3 level computing power is 200-500TOPS, and to reach the L5 level, it is estimated that more than 2000TOPS computing power is required. If based on existing technology, the power consumption will be more than 2kW, which is a challenge to the endurance of new energy smart driving cars.

Therefore, it is necessary to find a low-power, high-computing-power solution, and optical computing chips can just meet this need.

As for the currently popular embodied intelligence, optical computing chips can perform a large amount of calculations on the end side without consuming too much energy, allowing robots to perform a large number of mechanical operations while maintaining a "high IQ".

If embodied intelligence is to achieve the capabilities and cognitive abilities of general artificial intelligence (AGI), the required computing power will far exceed the current level. The high computing power potential of optical computing chips will enable embodied intelligence to think independently, and may even have thinking logic and judgment abilities that surpass humans. This will be the ultimate potential of optical computing technology.

It is undeniable that optical computing chips, with their ultra-high speed and high energy efficiency, will definitely give birth to a new algorithm paradigm in the future, thereby promoting progress and innovation in the entire field of computational science.

4. The road to industrialization of "nail in the left hand, hammer in the right hand"

With the tape-out of 128×128 matrix-scale optical computing chips, the next two years will be a critical period for the engineering of optical computing chips.In other words, optical computing chips must truly begin to be tested in “real and complex” scenarios.

"2025 will be a breakthrough year for optical-based technology. By then, optical computing chips will enter the first year of commercialization. Optical-based technology will have large-scale contact and adaptation with downstream companies to establish commercial partnerships. In 2026, optical computing chips are expected to be shipped on a large scale," Xiong Yinjiang predicted.

He told Jiazi Guangnian that the development of photonic technology follows a T-shaped strategy, which goes deep into the product cycle, from chips to boards, and then to downstream adaptation and final shipment. In the second half of this year, electronic chips will also be taped out, and the team will conduct packaging tests, complete 2.5D and 3D optoelectronic sealing, and eventually form photonic computing boards and deliver them to customers for adaptation, while building a software ecosystem.

Horizontally, Optical-Based Technology will broaden the R&D barriers in the field of optical computing and conduct R&D on a larger matrix scale, including but not limited to the R&D of optical computing chips with a matrix scale of 256×256 and 512×512, and utilize advanced technologies such as wavelength division multiplexing.

"This T-shaped strategy has both depth, ensuring that the product is not a castle in the air and can reach the user level; at the same time, it also ensures that we can always maintain our leading position in research and development." Xiong Yinjiang told "Jia Zi Guang Nian".

The operating principle diagram of the optical computing chip of Opto-Based Technology. Image source: Opto-Based Technology

Guangbenwei Technology regards R&D capability as the "hammer" and market demand as the "nail".It requires both continuous iteration of technology and continuous feedback on user needs.

If the two are to meet in the middle, both the market on the left and R&D on the right will require investment.

At the scientific research level, Optical Computing Technology and Fudan University have established a joint laboratory for future computing hardware to explore all possibilities of optical computing in the era of technological divergence.

In terms of engineering implementation, the team has assembled senior "industry veterans" in the fields of analog, digital, hardware, system architecture and algorithms to ensure that the product can be actually implemented in engineering.

At the same time, the company has established connections with partners in the domestic industrial chain, including strategic cooperation with the leading domestic silicon photonics company United Microelectronics, and cutting-edge research and development of 2.5D and 3D optoelectronic packaging with the top domestic packaging and testing plants.

The technological breakthrough is just a stepping stone. Next, the development of optical computing chips andOptical TechnologyFor enterprises, what they face is the challenge of industrial ecology.

At present, market customers generally use NVIDIA or other familiar GPUs. To achieve the widespread application of optical computing boards, we need to adapt with customers. "We are working hard to ensure that the optical computing boards are fully compatible with the existing ecosystem in terms of hardware." Cheng Tangsheng said.

certainly,Optical computing chips and electrical chips are not substitutes, but rather they enhance each other.

In the early stages of development, optical computing chips may play the role of tensor computing cores to assist traditional electronic chips. Given the current dominance of electronic chips in the market, optical computing chips will first take on a large number of linear computing tasks in artificial intelligence applications.

However, as customer needs gradually become richer, the entire software ecosystem will become richer, and the delivery capabilities of various optical computing companies will also improve.

Cheng Tangsheng compared the current development of optical computing chips to the "eve of AlphaGo". Compared with artificial intelligence or electronic chip platforms, optical computing is still in a relatively early stage of development. With the improvement of efficiency and the improvement of the ecosystem, optical computing will gradually play a greater role and provide strong technical support for computing, storage, transmission and perception in the intelligent era.

Mi Lei, the founding partner of Zhongke Chuangxing, proposed the "Mi 70 Law" in 2016. He believes that optical technology will be a very critical basic technology in the future, and its cost will account for 70% of the cost of all future technology products. However, it is undeniable that any technological breakthrough requires long-term accumulation and investment, as well as even longer transformation and implementation.

Human beings have benefited from light since their birth, and light will also create more possibilities for artificial intelligence.

*References:

, Jiazi Suzhou

(Cover image source: "Jia Zi Guang Nian" generated using AI tools)

【Jia Zi Live Room Preview】

This Thursday (July 18) at 14:00, Zhang Yijia, founder and CEO of Jiazi Light Year, and Nurmaimati (Xiaomai), an analyst at Jiazi Light Year Think Tank, will jointly analyze the opportunities and challenges behind the wave of AI medical care. During the live broadcast, they will combine the research results of Jiazi Light Year Think Tank and share the latest judgments on AI medical care in an easy-to-understand way at "0.5 times the speed". During the process, the two speakers will also answer questions from the audience and communicate with everyone. Don't miss it!

Scan the QR code on the poster to make an appointment for the live broadcast and enter the live broadcast discussion group in advance. At the same time, everyone is welcome to scan the code to obtain the report and master the learning materials in advance.