Recently, the sixth online forum of the “2024 China High-Quality Development Forum” – the “Intelligent Optical Network Technology Session” was successfully held. The forum invited many heavyweight guests from the industry and academia to conduct in-depth discussions on the architecture and key technologies of optical networks in the era of intelligent computing, the application practices of AI large models in various scenarios such as DCN and DCI, and hot topics such as the integration of optical network sensing and communication/integration of optical computing and others. At the meeting, Li Yunbo, the director researcher of the Basic Network Technology Research Institute of China Mobile Research Institute, was invited to give a keynote speech titled “The Development and Discussion of Intelligent Optical Network Technology in the Era of Intelligent Computing”.
Li Yunbo stated that artificial intelligence is closely related and mutually reinforcing with the development of optical networks. Optical networks are the key basic technology to achieve the intergenerational leap of artificial intelligence, while artificial intelligence promotes more powerful and advanced functions and performance of optical networks, enhancing their service capabilities for upper-layer businesses. As the carrier of AI computing power, intelligent optical networks will build a solid base with high reliability, large bandwidth, low latency, and cross-domain collaboration both inside and outside the intelligent computing center.
For computing power network services: Optical communication presents four major development trends
Li Yunbo pointed out that China Mobile has been conducting continuous technological research on 400G long-distance transmission, leading the development and industrial realization of QPSK, 130G high baud rate devices, and 12T wide-spectrum transmission technologies, laying a solid foundation for building a large-bandwidth and low-latency all-optical base for computing power networks. Facing the development of computing power network services, optical communication coexists with four major development trends: “globalization and nonlinearization of theoretical research”, “broadbandization and multi-dimensionalization of transmission systems”, ” diversification and stereoscopicization of optical connections”, and “twinning and intelligentization of optical networks”. Facing the nonlinear Shannon limit of optical fibers, new disruptive hollow-core fibers can be considered to reduce the nonlinearity of traditional quartz fibers; expand from the existing C+L bands to S and U bands to increase the entire fiber transmission capacity; extend optical connections down to between chips and up to satellite optical communications to realize the extension of optical connection technology to a three-dimensional architecture; at the same time, with the continuous expansion of the scale of optical networks, make full use of the digital twin of optical networks to build a self-healing, self-optimizing, dynamic and intelligent optical network.
Facing future technological evolution, the industry is continuously promoting the development of high-speed and wide-spectrum technologies for 800G core devices, including technological research on higher baud rates, and the development of wider-spectrum fiber amplifiers and optical splitters. At the physical medium level, anti-resonant hollow-core fiber technology has become the focus of the industry. China Mobile first achieved the upper limit determination of the nonlinear coefficient of hollow-core fibers (extremely small physical quantities), developed the directional independent and free dimension of hollow-core fibers for the first time, and proposed the theory of same-frequency and simultaneous full-duplex lossless transmission of optical fibers; in June 2024, it opened the first 800G hollow-core fiber transmission technology test network, achieving a 20km single-core fiber bidirectional 128Tb/s ultra-large capacity transmission test based on hollow-core fibers, and achieving multiple core technical indicators such as a fusion loss of hollow-core to hollow-core fibers as low as 0.05dB, a connection loss of hollow-core to solid-core less than 0.3dB, and a post-laying cable loss of 0.6dB per kilometer, all reaching the world-class level.
AI Empowering Optical Networks: Digital Twin Technology Becomes Key
Li Yunbo said that with the vigorous development of new businesses and new technologies, the scale of the network has grown exponentially. Emerging network services have put forward higher requirements for the basic optical network, and there is an urgent need to promote the digital and intelligent transformation of the transmission network, achieve quality and efficiency improvement in the operation and maintenance of the optical network, and meet users’ pursuit of the ultimate experience. “We are facing the demands of large differences in networking technologies, difficult perception of analog signals, and high requirements for implementation and detection.
We need to explore the introduction of artificial intelligence to improve the operation and maintenance efficiency of the optical network.” Li Yunbo introduced that facing the current situation of the optical network and the main functional characteristics of digital twins, China Mobile designed the overall architecture of the optical network twin, proposed technical concepts such as cross-layer, cross-domain, and cross-border, as well as technical architecture ideas such as cloud-pipe-terminal collaborative training, hierarchical multi-protocol collaboration, and modular system design. In terms of cross-layer, optical-electrical collaborative perception and scheduling are achieved through the optical layer OAM; in terms of cross-domain, data synchronization is achieved through the unified standard Telemetry; in terms of cross-border, hierarchical intelligent detection and decision-making are introduced to realize the linkage between computing power services and the basic network.
At the same time, China Mobile comprehensively utilizes key technologies such as cloud-pipe-terminal collaborative training, hierarchical protocol collaboration, and modular system design to improve the generalization ability, efficient perception, and smooth evolution of network services of the entire digital twin optical network. In terms of the practice of AI empowering optical networks, China Mobile innovated technical solutions such as the same-trench and same-cable detection and health degree prediction. For example, in the field of deep intelligent identification of the risks of the same-trench optical fibers, it can achieve real-time perception of the static characteristic parameters (length, attenuation, fusion point) and dynamic characteristics (vibrations caused by vehicles, construction, etc.) of the optical fibers, without manual participation throughout the process, and the overall identification accuracy of the same route reaches 97.7%.
Four Technological Breakthrough Points: The Importance of Networking in Intelligent Computing Centers Is Highlighted
The intelligent computing center is the physical carrier of AI computing power. The intelligent computing center mainly composed of GPU server networking requires more advanced basic network services. As an important infrastructure carrying upper-layer intelligent computing services, in the AI era, optical networks keep pace with the times and continuously achieve technological breakthroughs.
Li Yunbo pointed out that the development of artificial intelligence itself is inseparable from the effective guarantee provided by optical networks. In the development process of intelligent computing centers, it is closely related to the technological development of optical modules, optical fibers, service scheduling, and intelligent computing extension. Optical modules need to further reduce power consumption, latency, and improve reliability; by introducing a new optical fiber architecture and new types of optical fibers, the latency of service transmission within and between data centers can be further reduced; by introducing optical cross-connects, the flexibility of internal service scheduling in data centers can be further enhanced, and the availability of internal servers in data centers can be improved by isolating fault points. In terms of high-speed optical modules, the rate is evolving from 400G to 800G and even 1.6 Tbps. China Mobile collaborates with the industry to actively promote standard development, jointly tackle various key technologies such as high-speed EML and silicon photonics, and promote the prosperous development of new applications such as AIGC large models. Inside the intelligent computing center, traditional pure electrical cross-connect-based switches are facing the bottleneck of capacity expansion and the reality of 100G/400G/800G hybrid networking.
Based on port-level optical cross-connect technology, low-latency and low-power cross-scheduling and hybrid networking can be achieved. At the same time, it also has the characteristics of topology reconfiguration, on-demand expansion, and redundancy protection, which can achieve more efficient and reliable service scheduling capabilities within the data center. In addition, as the scale of intelligent computing centers continues to expand, from ten-thousand-card clusters to one-hundred-thousand-card clusters, new bottlenecks are faced in terms of power supply and machine room space. Through network capabilities, a larger cluster scale can be obtained, and at the same time, computing power resource islands can be integrated. Li Yunbo said that currently, the focus should be on solving the problems of inter-building interconnection within the park and inter-site interconnection within the same city to meet the construction requirements of one hundred thousand cards; at the same time, accelerate the research and verification of inter-provincial multi-node interconnection technologies to meet the potential networking requirements of over one hundred thousand cards in the next 3-5 years. Currently, China Mobile has completed the 2/10/20/40/60/80/100-kilometer OTN-bearing distributed intelligent computing tests and is jointly conducting further research on key technologies of thousand-kilometer-level intelligent computing extension through industry-university-research and multi-disciplinary collaboration.
