The development of computing technology can be roughly divided into the era of mechanical computing and the era of electronic computing. From the original mechanical computer to the current supercomputer, the development of computing technology can be described as rapid with each passing day. With the development of information technology and the deepening of social digitalization, the demand for high-performance computing in various fields continues to grow. Therefore, in recent years, the "fire" of optical chips has become more and more prosperous, and various regions have introduced policies to support the development of optical chips.
The main reason for the popularity of optical chips is their potential to improve computing power. Optical chips are the core components of optoelectronic devices, mainly used to transmit, receive, modulate and process optical signals, and are also an indispensable part of optical communications, lidar, data centers and other fields. Its working principle is based on the photoelectric effect of semiconductor materials, and the mutual conversion of electrical signals and optical signals is completed by controlling the generation, transmission and monitoring of photons.
Classification and advantages of optical chips
According to the function, optical chips can be divided into transmitter chips, receiver chips, modulator chips, integrated chips, etc. Among them, the transmitter chip is mainly used in optical fiber communication (such as 5G base station optical module), data center optical interconnection, etc. The receiver chip is mainly used for optical sensing systems, and the modulator chip is mainly used for optical communication systems (such as 100G/400G optical modules); Integrated chips are mainly used for high-end optical modules, high-density interconnection of data centers, etc. Compared with traditional chips, optical chips have significant advantages in core dimensions such as signal transmission, power consumption, and architecture, such as Joule heating loss in the transmission of traditional chip electrical signals in copper cables, skin effect at high frequency further increases power consumption, chip heating is serious (such as CPU needs to be matched with a large heat sink), and optical signals have low transmission loss in optical fiber (about 0.2 dB/km), and the power consumption of photoelectric conversion inside the chip is only 1/5-1/3 of that of electronic chips (such as 100G optical module power consumption <5W, The power consumption of the same rate electrical module >15W).
The demand for computing power is growing rapidly
At present, the number of parameters of AI models has increased dramatically, from millions of parameters in the early days to hundreds of billions or even trillions of parameters today, such as large language models such as GPT-3. AI technology is widely used in many fields such as image recognition, voice recognition, autonomous driving, and medical diagnosis. In autonomous driving, vehicles need to process a large amount of sensor information such as camera images and radar data in real time to make accurate driving decisions, which puts forward extremely high requirements for computing power. In medical diagnosis, AI-assisted medical image analysis needs to quickly process a large amount of image data such as CT and MRI, which also relies on powerful computing power. In addition, new machine learning algorithms, such as the Transformer architecture in deep learning, have made remarkable achievements in areas such as natural language processing and computer vision. These advanced algorithms often require higher computing power to achieve their performance benefits.
Figure: When the chip manufacturing process is limited, how can humans continue to improve computing power?
How optical chips solve the problem
Ultra-high data transmission speed: The propagation speed of optical signals in optical fibers is close to the speed of light, much higher than the transmission speed of electrons in wires. For example, the 100G PAM4 EML optical chip released by Changguang Huaxin can transmit massive data at a speed of 100Gbps per second. This enables the optical chip to quickly transmit data to the computing unit when processing massive data in big data analysis, which greatly shortens the data processing time. In AI model training, for example, Google's data center uses optical modules with optical chip technology to significantly improve the network performance of the data center, accelerate the data transmission speed of large-scale AI model training and inference, and improve the training efficiency of the model.
Powerful signal processing capabilities: Photonic devices have a high degree of parallelism and matrix computing capabilities, which can efficiently process complex data, and can also realize analog calculations to further improve signal processing capabilities. For example, the "Tai Chi-II." optical chip developed by Tsinghua University, the first full forward intelligent optical computing training architecture, realizes the in-situ optical training of large-scale neural networks, can increase the training speed of millions of parameters of optical networks by 1 order of magnitude, increase the accuracy of processing intelligent classification tasks by 40%, and also realize computational imaging at kilohertz frame rate, and improve imaging efficiency by 2 orders of magnitude.
Low energy consumption characteristics: The energy loss in the process of optical signal transmission is much lower than that of electronic signals, and the energy consumption of optical chips is only about one-tenth of that of traditional electronic chips. Like Microsoft's Project LightSpeed, which aims to use optoelectronics to improve network speed in data centers, through the use of optical chips, it is expected to reduce data center energy consumption and increase processing power. For long-running computing tasks in big data analysis and AI data centers, optical chips can effectively reduce energy consumption while achieving high-speed computing, which is in line with the concept of green computing and reduces operating costs.
As Moore's Law approaches the bottleneck, the explosive demand for computing power in the era of artificial intelligence is driving the global chip industry to seek architectural innovation, and building a next-generation computing system has become the strategic commanding heights of international scientific and technological competition. In this context, optical chips are regarded by the industry as a key technology to break through the limitations of traditional computing architectures due to their characteristics of high speed, low power consumption, parallel processing, etc., and have attracted industry giants such as Intel, NVIDIA, and TSMC to deploy one after another. Although optical chips have broad prospects and continue to heat up, there are still many challenges from laboratory to large-scale industrialization. Industry experts emphasized that it is urgent to strengthen basic scientific research, break through the technical barriers of semiconductor materials and processes, accelerate the industrialization process through in-depth collaboration between industry, university and research, and overcome multiple difficulties in technology and application through multi-party collaboration, so as to release the true potential of optical chips.
