Cross-border integration, intelligent computing for the future! iFOC 2024 Xunshi Symposium was successfully held - Xunshi Optical Communication Network (iccsz.com)

Abstract: Cross-border integration, intelligent computing for the future! iFOC 2024 was successfully held

ICC News Cross-border integration, intelligent computing for the future! On September 9-10, 2024, the 22nd Xunshi Optical Communication Market and Technology Symposium (iFOC) was held 2024) was successfully held in Shenzhen, with a total of 6 thematic forums, 5 roundtable discussions, 4 salon interactions, 3 technical trainings, 2 keynote forums and 1 400G live demonstration, as well as 24 enterprise product displays. The number of participants in this iFOC exceeded 1,200, 420 enterprises, scientific research and investment institutions, and more than 80 optoelectronic experts from optoelectronics, optical communication industry chain and academia published industry reports, focusing on AI large models, AI computing power, OCS/optical switching, 400G/800G/1.6T/3.2T, DSP\LPO\LRO, OIO, CPO, backbone optical network, coherent sinking, WSS, parallel optics, wavelength division multiplexing, advanced packaging, lidar, Market demand such as intelligent driving, hollow-core optical fiber, and market analysis and forecasting, closely follow the global trend of optical communication, integrate optoelectronic technology hotspots, and explore industrial innovation and development paths.

Professor Chen Yixin, lifelong consultant of ICC Xunshi Consulting, former professor of the Department of Applied Physics of Shanghai Jiao Tong University, and lifelong member of IEEE, said in his speech that the era of intelligence and intelligent computing has been accelerating, and iFOC has gathered industry elites to discuss the new generation of network interconnection and realize a new round of development of the optical communication industry. Professor Shen Ping, Director of the Optoelectronic Intelligence Laboratory of Southern University of Science and Technology and President of the Photonics Society, also encouraged industry colleagues to seize the opportunity of intelligent computing and work together to meet more challenges in the AI era.

In Keynote Forum 1: Optical Communication Enabling Computing Base, experts from China Mobile, iFLYTEK, Broadcom, China Telecom, Hengtong Optoelectronics, HiSilicon Optoelectronics, YOFC, China Unicom, and Huawei delivered reports on AI intelligent computing power and optical communication.

Dr. Li Han of China Mobile published a report entitled "Discussion on the Development of T-bit Optical Communication Technology and Devices for Computing Power Networks", which puts forward higher requirements for optical transmission networks in terms of ultra-large bandwidth, ultra-long distance, and ultra-low latency for typical scenarios of computing networks such as "Eastern Data and Western Computing". China Mobile proposed and built a new all-optical network technology architecture based on 400G+OXC. At the same time, China Mobile has promoted the development of 400G/800G technology and industry, focusing on the exploration and thinking of the key technologies of 1.6T ultra-high-speed optical transmission system and the evolution of optical devices, as well as the research progress and application of hollow-core optical fiber in high-speed systems, hoping to jointly promote the end-to-end technological innovation and ecological construction of T-bit all-optical networks with the industry.

Dr. Luo Yuan, Chief Architect of iFLYTEK Data Center, published a report titled "Requirements and Challenges of Optical Connectivity in Large-scale High-Speed Intelligent Computing Clusters", saying that optical connectivity plays a crucial role in high-speed environments such as large model training clusters, and its stability and performance indicators will directly affect the use of the entire cluster. At the same time, iFLYTEK introduced the use and operation of optical connections in the Vanka clusters that have been put into production, the planning of connections in the Vanka clusters currently under construction, and discussed the requirements, challenges and more application scenarios of optical connections in 100,000 cards or even larger clusters.

Tzu-Hao Chow, Senior Commercial Manager at Broadcom, presented "Optical Technologies in AI Networks: Optical Technologies in AI Networks: Scaling with High Density Since OpenAI released their ChatGPT platform, AI applications have entered a phase of explosive growth. This triggers a huge hardware drain on GPU servers and networks. This presentation focused on the optical vision of AI networks. There are three key points to focus on, which are VCSEL, EML, and CPO. VCSEL is used in 100-meter short-distance transmission scenarios, and now distributed computing scenarios are also discussing whether to use VCSEL chips. To achieve speeds in excess of 100Gbps, advanced Vcsel technology and fiber bandwidth will be considered. In order to support 2km long-distance transmission, EML chips provide the fastest market response scheme and have a proven track record of reliability and scale. In addition, CPO technology is being discussed for switch and xPU application architectures in order to enable densely integrated optical connections. As the scale of the Cluster xPU increases, the use of single-fiber bidirectional (BiDi) technology will save infrastructure resources.

Wei Leping, director of the Science and Technology Committee of China Telecom Group, published a report on "Opportunities for Optical Communication in the Era of Large Models", saying that the development of large models is inseparable from the support of high-quality networks, and the basic role of new networks is to ensure the computing efficiency of GPU groups. The era of large models will open a new wave of optical input and copper retreat, and three of the trends are the development of CPO technology, silicon photonics technology, and all-optical switching/routing technology. In addition, the large model will consume a large number of DCN and DCI optical modules, and only one GH200 supercomputing system will consume 3072 800G optical modules, which gives huge development opportunities to optical modules, and drives the development of high-speed all-optical networks and hollow-core optical fibers, making full use of the advantages of low-latency, nonlinearity and low loss, wide spectrum and large core diameter of hollow-core fibers.

Dr. Sun Wei, R&D Director of Hengtong Optical Fiber, published the report "Research and Application Discussion on Multi-core Optical Fiber Technology for Space Division Multiplexing Large-capacity Communication", the capacity of global computing power networks and communication networks continues to expand, and traditional communication optical fiber technology is difficult to support the surge in capacity demand. The development of space division multiplexing multi-core optical fiber technology has gradually shown application potential in both land and sea fields. At present, Hengtong is becoming more and more mature in the preparation technology of multi-core optical fibers, carrying out experimental research with partners in long-distance and large-capacity communications, and cooperating with brother units in the field of marine communications, and has made remarkable progress.

Dr. Jiangwei Man, chief researcher of HiSilicon, published a report titled "From General Computing to Intelligent Computing: Discussion on the Evolution of AI Optical Interconnect Module Technology", saying that with the rapid development of cloud data centers, traditional optical interconnection modules in general computing data centers are being used in carrier-class applications of operators. At the same time, compared with the general computing data center, the AI intelligent computing data center also puts forward higher requirements for the low latency, intelligence, high reliability, and low power consumption of optical interconnection modules. The industry chain needs to pay attention to the challenges of the application of optical interconnection modules in AI intelligent computing centers, so as to improve the availability and reliability of optical interconnection in AI intelligent computing centers.

With the rapid development of artificial intelligence, cloud computing, and the Internet, the demand for optical fiber communication in various application scenarios continues to expand, which promotes the rapid development of ultra-high-speed and large-capacity optical communication and large-scale data center services. The demand for larger bandwidth and higher speed has put forward new requirements and challenges for optical fiber transmission network technology. This report mainly introduces the comprehensive solutions and applications of special optical fibers in ultra-high-speed and large-capacity optical communication systems and supercomputing data centers, and considers the development trends and challenges of special optical fibers in the next generation of communication technology and the broad prospects of its diversified applications.

Dr. Tang Xiongyan, Vice President and Chief Scientist of China Unicom Research Institute, delivered a report on "Opportunities and Challenges of Optical Networks in the Era of Artificial Intelligence", saying that in the era of artificial intelligence, the demand for intelligent computing is growing rapidly. Optical networks will play an increasingly important role in high-throughput wide-area interconnection and data center internal networking. For high-speed lossless interconnection between intelligent computing, it is necessary to continuously improve the bandwidth, capacity, transmission distance, cost per bit, and energy efficiency per bit of optical networks to build more flexible, intelligent, and reliable optical networks to meet the growing demand for intelligent computing interconnection. For the internal interconnection of intelligent computing centers, optical network technology needs to be gradually introduced to build a high-speed, efficient, green, low-carbon, intelligent and agile intelligent computing data center.

Tan Jingxin, Chief Solution Architect of Optical Transport at Huawei, released a report titled "Discussion on the Concept and Key Technologies of Optical Transport Target Networks in the Intelligent Computing Era", which stated that in the era of intelligent computing, the service development trend of optical transport networks is mainly reflected in the efficient collaboration between computing devices to achieve the ultimate user experience. In the DCI scenario, distributed computing is used to drive inter-computing interconnection, and in the DCA scenario, deterministic network is achieved, ubiquitous quality access, and computing resources are moderately reduced. The evolution direction of optical transport network includes network servitization, line broadbanding, media scenario-based, access diversification, quality collaboration, and optical-layer digitalization. Build an end-to-end optical network for computing services, evolve to 400G+high-baud rate/128QAM high-order modulation/C+L+S+E around the single-fiber capacity improvement, and build a large-capacity transmission system using air-core fiber and multi-core fiber systems.

At the keynote forum 2: "High-speed Optical Interconnection Weaving AI Network", optical communication experts from ByteDance, Huagong Zhengyuan, Corning Optical Communications, Intel, Haiguang Xinchuang, POET, Hisense Broadband, Tencent, Marvell, and Alibaba successively published keynote reports on the construction of AI networks empowered by optical interconnection technology.

Guo Lei, senior optical network architect of ByteDance, published a report on "800G Optical Interconnection Technology Innovation and Practice", saying that the exponential growth of computing power requires higher interconnection bandwidth, which further promotes the rapid iteration of optical interconnection technology, and comprehensively supports the construction of large-scale model network architecture around job number, bandwidth, performance, latency and cost. ByteDance will begin deploying 800G optical modules in 2024 and add its own innovation, 800G, to 800G optical modules The SR8 can achieve 100 meters of transmission over OM4 fiber, with a Flat-Top thermally conductive design and true backlit monitoring. ByteDance has also conducted tests on the LPO module, considering that LPO If you want to successfully land, you need to have the end-to-end capability of the whole chain of the system, joint optimization, dismantling and implementation. In the post-800G era, optical module failure prediction is very important, and at the same time, the connection mode of multi-stage tracks under the large model promotes the sinking of single-mode, ushering in a golden age of silicon photonics technology.

Jim Theodoras, Vice President of R&D at HGZ, published "How to Optimize Datacom for AI and LLM Applications Optical Transceivers for AI and In the next two generations of Ethernet switches, optical power will account for a larger proportion of the total power, and LPO can well offset the excess optical power consumption. LLM training will also consume a lot of time, power, and processing power. Datacom optical modules and AI optical modules are different, for example, to adapt to different application rates, the same product must be able to work on three internal networks, and each network has different considerations. In addition, a lower BER is required, and the data communication uses VCSEL for short-distance links, while AI optical modules use single-mode resources for short-distance links. At the same time, AI optical modules still need to maintain stable performance when FEC is turned off. AI will usher in greater application opportunities for LPO and silicon photonics, and 1.6T optical modules have arrived.

According to the report "Band Expansion Multimode Fiber for Cloud and AI Data Centers" published the report "Low-cost, low-power and low-power multimode link transmission based on VCSELs is critical for short-range communications, especially in AI/ML trunking interconnects." Multimode link rates are evolving to 100G per lane, and the industry expects MMF to maintain its OM3 support for 70m and OM4 for 100m capabilities. However, the mode bandwidth limitations of OM3 and OM4 at certain wavelengths pose a challenge in transmission distance. To solve this problem, we introduce Band-Extended Multimode Fiber (HDR) that supports high-speed transceivers MMF), HDR OM4 has similar bandwidth performance to OM5 in the 840~910 nm band range, and can support 850nm VCSEL wavelength offset and BiDi transceivers transmit distances of up to 100 meters and achieve lower production costs than OM5. The results of analysis and actual testing show that HDR MMF provides more flexible VCSEL wavelength selection and more economical deployment for current 800G and future 1.6T.

Marcus Yang, Senior Product Director of IPS at Intel, published the report "Silicon Photonics Technology and New Application Brief: Optical Computing Interconnect (OCI), Co-packaged Optics (CPO)", Silicon Photonics (SiPh) combines optics and electronics to use silicon for light generation, manipulation, and detection, leveraging proven CMOS process for cost-effective manufacturing. New AI applications require advanced optical I/O, integrated optics through greater integration within photonic ICs, and top-of-the-line electronic ICs Heterogeneous integration, as well as tighter host integration, provide the solution. Intel's SiPh platform relies on on-chip lasers, SOA, and powerful IP The portfolio plays a vital role in the fields of artificial intelligence, high-performance computing and optical computing. Methods include treatment of SiPh and continued investment in advanced packaging to enable CPU concepts with in-package OCI chiplets. Intel's OCI solution offers best-in-class specifications and scalability that are critical for multiple compute fabric applications such as disaggregated architectures. With a compelling long-term scalability roadmap, Intel's innovation extends to best-in-class OCI Demo and FMCW LiDAR, highlighting the transformative potential of SiPh.

Dr. Chen Xiaogang, Chief Scientist of Haiguang Xinchuang, published "The Development Road of Silicon-based Optoelectronic Chips in the AI Era - Fabless". 2.0" report, ultra-high bandwidth, ultra-low power consumption, ultra-low latency on-chip optical interconnection network, combined with core technology of three-dimensional optoelectronic integration chip, is not only a key technology for China's high-end computing chips to break through the technology blockade, but also the inevitable choice for the development of the world's most cutting-edge AI chip architecture in the future. From high-speed optical modules to on-chip optical interconnects, standardized production and packaging and testing are the only way to promote the sustainable development of the silicon photonics industry. Haiguang is committed to the construction of a silicon photonics industry chain for mass production, strives to build a national silicon photonics product development and production platform in the AI era, and actively contributes to China's AI high-speed interconnection network.

POET Dr. Jinyu Mo, Senior Vice President of Technologies, published "Semiconductor Photonics: 200G/Channel Products for 1.6T and Higher Speeds Photonics: 200G/Lane Products for 1.6T and Beyond）》，

Dr. Huang Weiping, Chairman of the Board of Directors of Hisense Broadband, published the report"Jiubihe,Hejiubifen Optical Wave Connection Technology and Industry Trends in the Intelligent Era"Report,The trend of optical connection includes3 points,First, the demand for computing power promotes the increase of node bandwidth、Energy consumption and cost decrease,Second, the form of external optical module products will be replaced by built-inCPO or integrated optical I/O,The third is to find a photoelectric conversion technology solution that can continue to improve and expand,Silicon photonics has been widely discussed and practiced。 Bandwidth density, power efficiency, and cost-effective connectivity will drive optical connectivity technology toward CPO and integrated optical I/O, optical modules will fundamentally change or die out as independent products, and photoelectric conversion will become the built-in function and design IP of switches and processor chips. The success of China's optical communication enterprises in the past 10 years is mainly due to seizing new market demand opportunities. However, the competitive advantage of any organization and country is temporary, and it is difficult to maintain it indefinitely. To win the competitive advantage of the intelligent computing industry, China urgently needs to establish and develop core technical capabilities and industry leaders such as optoelectronic fusion and independent controllable chip design, tape-out, packaging and EDA software.

Dr. Fu Sidong, Tencent's Optical Network Architect, published a report on "Data Center Silicon Photonic Chip Technology and Evolution", Tencent has carried out a number of innovative practices in the field of silicon photonic chip technology, reducing the chip area by 20% compared with commercial standard products, supporting one laser to achieve low packaging cost, and compatible with LPO applications and greater than 32 The high bandwidth of GHz is balanced and optimized for the BR4 standard. For the future silicon photonics chip technology, BR4 will be extended to BR4.2+ to achieve 800G interconnection, the integration of transmitter and receiver can achieve optimized coupling loss, in addition, the chip size of the microring modulator scheme is expected to be reduced by 60%, and the link loss can be reduced by 2.1 dB。 With the evolution of integrated optical systems, silicon photonics technology will achieve higher speeds, higher integration, and richer functions.

Dr. published a report on "Technology and Evolution of Silicon Photonic Chips in Data Centers".

LianZhao, Vice President, Optical Connectivity Marketing, Marvell According to Qin's "AI Data Center High-Speed Optical Connectivity" report, Marvell provides the industry with complete DSP, TIA and driver solutions, ranging from 200G to 1.6T, to meet the growing demand for bandwidth and speed for AI and machine learning applications. Marvell's portfolio also includes DSP-based TRO and LPO pluggable modules. As the market evolves towards 800G and 1.6T solutions to support growing data traffic. Marvell's 1.6T Nova Optical DSPs have proven that they are ready to meet these needs, delivering twice the bandwidth of today's 800Gbps transceivers. Marvell's continued innovation in high-speed optical connectivity is laying the foundation for next-generation data center infrastructure.

Wu Na, senior analyst of ICC Xunshi, published the report "Analysis and Forecast of the Current Situation of the AI Optical Market".