With the rapid development of artificial intelligence (AI) and big data technology, data transmission efficiency has become a key bottleneck restricting technological breakthroughs. With the rapid expansion of the parameter scale of deep learning models, the communication load between computing nodes continues to rise, posing unprecedented challenges to high-bandwidth and low-latency data transmission technologies. In this context, a research team from the School of Information Science and Engineering at Fudan University has successfully developed a silicon photonics integrated high-order mode multiplexer chip that supports ultra-high-speed data transmission of 38 terabits per second (Tb). This breakthrough not only brings light to the global scientific and technological community, but also provides new technical support for artificial intelligence, big data processing and other fields.
1. Technological innovation: the convergence of multi-dimensional multiplexing and on-chip optical interconnect
The core innovation of the chip is the introduction of multi-dimensional multiplexing technology into the on-chip optical interconnect architecture. Multi-dimensional multiplexing technology combines time division multiplexing (TDM), space division multiplexing (SDM), and wavelength division multiplexing (WDM) to enable multiple data streams to be efficiently transmitted in parallel within the same optical waveguide, thus greatly improving data throughput.
Core advantages of silicon photonics technology:
1. High-speed transmission: Traditional electronic interconnects are limited by the resistance and capacitance effects of copper wire transmission, making it difficult to break through the data rate bottleneck. The bandwidth of photon transmission is much higher than that of electrons, allowing silicon photonics chips to achieve higher data rates.
2. Low power consumption: Compared with electronic interconnect, optical interconnect can significantly reduce energy consumption in the process of data transmission, providing strong support for green computing.
3. Low latency: Silicon photonics interconnect avoids signal attenuation and delay caused by wire impedance during the transmission of electronic signals, and greatly improves the efficiency of data exchange.
4. Compatible with CMOS process: Silicon photonic chips can be directly mass-produced using existing semiconductor manufacturing processes, reducing costs and accelerating industrialization.
Experimental data show that the chip still maintains a very low bit error rate at a transmission rate of 38Tb/s and exhibits good signal integrity. This technology can improve the interconnection efficiency of data center servers and supercomputers compared to traditional electronic interconnection solutions, reduce power consumption, and drive innovation in high-performance computing (HPC) and artificial intelligence computing architectures.
Figure: The Fudan team developed a silicon photonics chip to achieve 38Tb/s data transmission to facilitate AI and large model training
2. Breakthrough in AI communication bottlenecks
At present, the scale of AI model training is growing at an unprecedented rate, and the number of parameters of large language models such as GPT-4 has reached trillions of levels. In the process of model training, a large amount of data needs to be exchanged between multiple computing nodes at high speed, and the traditional electrical signal transmission method can no longer meet the requirements of ultra-large-scale computing.
Taking a typical training task of deep learning as an example, it is assumed that each compute node needs to complete tens of trillions of floating-point calculations (FLOPs) within one second and synchronize parameter updates with other compute nodes. If the communication bandwidth is insufficient, computing resources will be idle due to data waiting, resulting in a "computing-communication imbalance". With an ultra-high transmission rate of 38 Tb/s, the silicon photonics chip developed by the Fudan University team can complete the transmission of 4.75 trillion parameters in 1 second, effectively alleviating the data transmission bottleneck in the AI training process and improving the overall efficiency of the computing cluster.
In addition, the chip can also be widely used in high-performance computing (HPC), cloud computing, edge computing, autonomous driving, financial modeling and other high-throughput application scenarios, helping to build next-generation intelligent computing systems.
3. Industry impact: Silicon photonics technology leads the revolution of optical communications
Breakthroughs in silicon photonics technology have not only advanced academic research, but also had a profound impact on the industry. According to market research firm Yole Development, the global silicon photonics market is expected to grow from $25 million in 2013 to $700 million in 2024, with a compound annual growth rate of 38%. With the continuous growth of demand for data centers, 5G communications, and artificial intelligence computing, the market prospect of silicon photonics chips is very broad.
The core application directions of silicon photonics technology include:
- Hyperscale data centers: Increase interconnection rates between servers, optimize data transmission energy consumption, and reduce cooling costs.
- Next-generation communication networks: Promote the development of 5G/6G optical communication technologies to improve the bandwidth and efficiency of optical fiber networks.
- Autonomous driving and intelligent transportation: Support high-precision sensor data processing and improve vehicle-to-everything (V2X) communication performance.
- High-Performance Computing (HPC): Optimize data transmission architectures for applications such as scientific computing, weather simulation, genetic analysis, and more.
Global technology giants, such as Intel, IBM, Huawei and other enterprises, have deployed in the field of silicon photonics chips. Intel's 800G silicon photonics module launched in 2022 has made a breakthrough in the high-end data center market. Leading domestic enterprises are also actively developing a new generation of high-speed silicon photonic chips to provide technical support for the development of the domestic information and communication industry.
4. Future outlook: continuous innovation and application expansion
According to the Fudan University research team, the silicon photonics chip still has the potential for further optimization. In the future, the researchers plan to improve the integration of chips and optimize the device structure to further reduce power consumption and improve manufacturing yield. At the same time, the team will explore the application of this technology in more cutting-edge fields, such as quantum computing, photonic neural networks, etc.
With the development of artificial intelligence, cloud computing, and 6G communications, the demand for data transmission will continue to rise. With its unique advantages, silicon photonics technology is expected to become the core supporting technology for future computing power networks and high-bandwidth communications. The government and enterprises should increase investment in silicon photonics technology, promote technological innovation and industrial ecological construction, and further consolidate China's leading position in the global optical communication field.
5. Conclusion
The silicon photonics integrated high-order mode multiplexer chip successfully developed by the Fudan University team is not only a breakthrough scientific research achievement, but also marks a key step in the field of optical communication and high-performance computing in China. This technological breakthrough will provide strong technical support for many cutting-edge fields such as artificial intelligence, big data processing, and cloud computing, and accelerate global scientific and technological progress. With the further development of silicon photonics technology, we have reason to believe that in the future era of intelligent computing, China's "chip" will continue to lead scientific and technological innovation and achieve "speed of light".
