As large model parameters, training data, and GPU (Graphics Processing Unit) cluster sizes continue to grow, east-west traffic in intelligent computing networks is increasing rapidly. Intelligent computing business is evolving from single DC (Data Center) operations to cross-DC collaboration and end-edge-cloud coordination, expanding communication scope from within DCs to campus, metro, and cross-region networks. Traditional electrical interconnects are limited by transmission distance, power consumption, and wiring density. Optical interconnects, with advantages such as high bandwidth, low latency, and long-distance transmission, have become a critical component of intelligent computing infrastructure.
PART.01Typical Application Scenarios, Communication Characteristics, and Interconnect Requirements of Intelligent Computing Business
Typical application scenarios for intelligent computing include computing resource scheduling, data ingestion into computing, model training, model distribution, and model inference, as shown in Figure 1.
Computing Resource Scheduling
Computing resource scheduling requires coordinating computing power, storage, network status, and business demands to select computing nodes across ends, edges, and clouds. As intelligent computing resources evolve towards multi-center and edge-distributed deployment, scheduling is shifting from simple computing matching to computing-network collaboration. The demand for optical interconnects in this scenario mainly lies in network status sensing, elastic scheduling, and dynamic coordination with computing resources. China Unicom's "Xingluo" computing scheduling platform enables integrated scheduling of general-purpose, intelligent, and supercomputing resources, supporting collaborative scheduling based on strategies such as latency, computing power, and cost.
Data Ingestion into Computing
Data ingestion includes massive data ingestion and real-time data ingestion. Massive data ingestion primarily refers to the aggregation of user data into intelligent computing centers, focusing on sustained high throughput, high reliability, and transmission cost control; optical interconnects must provide large bandwidth, stable long-distance transmission, and high reliability. Real-time data ingestion places greater emphasis on latency, jitter, and business continuity. In 2024, China Unicom completed the verification of high-throughput lossless transmission of massive data over more than 3,000 km from Shanghai to Zhongwei, Ningxia, importing Shanghai's intelligent computing training data into the Zhongwei intelligent computing training cluster. TB-level data can be delivered in as little as half a minute, fully verifying rapid intelligent computing data migration in "East Data West Computing" scenarios.
Model Training
Model training mainly includes centralized training within a single DC, storage-compute decoupled remote training, and cross-DC collaborative training.
Centralized training within a single DC organizes multi-GPU collaboration through methods such as DP (Data Parallelism), TP (Tensor Parallelism), PP (Pipeline Parallelism), and EP (Expert Parallelism), utilizing collective communications to exchange gradients, parameters, and intermediate results. This imposes key performance requirements on the network, including large bandwidth, high bandwidth density, low latency, low jitter, and high reliability. Taking Llama 3.1 70B as an example, it contains approximately 70 billion parameters. If gradients use the BF16 format, the complete gradient data volume is approximately 140 GB. In DP, gradients computed by each GPU need to be aggregated via collective communications such as AllReduce, resulting in significant inter-node communication volume. With high-density GPU deployment, concurrent multi-GPU communications lead to a rapid increase in network bandwidth per unit space. For instance, the 1U compute tray of NVIDIA GB300 NVL72 integrates 4 Blackwell Ultra GPUs and configures 4800 Gbit/s compute network interfaces, achieving a theoretical bandwidth of 3.2 Tbit/s per rack unit.
Storage-compute decoupled remote training deploys computing clusters separately from storage resources such as datasets, Checkpoints, and model files, accessing them remotely via the network. This approach is suitable for scenarios where massive data migration costs are high or constrained by data security, privacy, and compliance requirements. Traffic primarily consists of continuous reading of training data and staged writing of Checkpoints, requiring optical interconnects to support high throughput and high reliability.
Cross-DC collaborative training involves multiple intelligent computing centers jointly completing tasks and exchanging gradient, parameter, or intermediate state data across centers. As computing capabilities improve, model iteration computation time shortens continuously, while the proportion of communication time (data transmission time and fiber propagation delay) increases. Therefore, optical interconnects need to support large capacity, low latency, low jitter, and lossless transmission. Taking Alibaba Cloud's Qwen2.5-72B model as an example, it contains approximately 72.7 billion parameters. Calculating with BF16 gradients, one complete gradient data set is approximately 145 GB. In a simplified dual-DC scenario, assuming the transmission time for one complete gradient synchronization per round is 2 seconds, the unidirectional effective data bandwidth required is approximately 582 Gbit/s. Based on a single wavelength of 400 Gbit/s, theoretically at least two 400G wavelengths are needed, while actual requirements must also account for overheads from protocols, FEC (Forward Error Correction), link utilization, and protection. When the cross-DC distance is 300 km, 1000 km, and 1500 km, estimating fiber propagation delay at approximately 5 μs/km for standard single-mode fiber, the one-way propagation delays are approximately 1.5 ms, 5 ms, and 7.5 ms, respectively. The fiber propagation delay accounts for less than 1% of the communication time, indicating that under these bandwidth conditions, cross-DC training is primarily constrained by data transmission bandwidth. As cross-DC interconnect bandwidth further improves, data transmission time decreases, and the proportion and impact of fiber propagation delay will gradually increase, necessitating consideration of low-latency transmission node equipment and hollow-core fiber. China Unicom utilized the AINet (Artificial Intelligence Network) spanning approximately 1500 km to connect the Shanghai and Jinan intelligent computing centers, completing mixed training of hundred-billion-parameter large models, achieving equivalent computing power exceeding 95% of a single cluster.
Model Distribution
Model distribution refers to delivering trained model weights to inference clusters or user endpoints, characterized by large files, one-to-many transmission, and cross-DC transfer. After model launch, version updates and incremental parameter synchronization are also involved. Although the scale of such traffic is smaller, updates are frequent, requiring optical interconnects to prioritize high reliability and elastic bandwidth. In 2026, China Unicom built a "Model and Token Distribution Platform," leveraging the "Xingluo" advanced computing scheduling platform to provide inference distribution and "Train Central, Infer Edge" capabilities.
Model Inference
Model inference can be categorized into centralized inference within a single DC, end-edge-cloud collaborative inference, and cross-DC collaborative inference. Model inference performance is typically measured using TTFT (Time To First Token) and TPOT (Time Per Output Token). TTFT reflects the waiting time from user request arrival to the first token output, including queuing, Prefill computation, and network transmission overheads. TPOT reflects the average time interval between consecutive tokens generated during the Decode phase, where network transmission latency and jitter directly affect the generation time of each token. Taking the MLPerf Inference Llama2-70B Server benchmark as an example, its TTFT and TPOT constraints are 2 s and 200 ms, respectively, used to measure user experience aspects such as response and continuous generation in online inference.
Centralized inference within a single DC typically organizes multi-GPU computation through methods like TP and PP. Network transmission latency has a significant impact on TPOT. Adopting a PD (Prefill-Decode) disaggregated architecture further increases data transmission overhead between prefill and decode resource pools, thus requiring lower latency and jitter from optical interconnects. In end-edge-cloud collaborative inference, network transmission latency affects overall inference response, imposing higher requirements on network latency and reliability. As the distance from center to edge expands, fiber propagation latency and network processing overhead latency will further compress inference computation time. Cross-DC collaborative inference involves cross-center transmission, where routing selection in complex topologies, long-distance propagation delays, and cross-domain jitter all affect TTFT and TPOT. Therefore, besides bandwidth, attention must be paid to cross-domain latency determinism, reliability, and joint scheduling of computing and network resources.
Synthesizing the above scenarios, different intelligent computing businesses have varying emphases on optical interconnect requirements, but their core capability needs can be summarized into five aspects. First, large bandwidth and high bandwidth density: centralized training within a single DC and cross-DC training/inference generate substantial east-west and cross-domain traffic, requiring continuous improvement in optical interconnect single-port speeds, as well as bandwidth density at device, rack, and cluster levels. Second, low latency, low jitter, and determinism: excessive communication time reduces overall training/inference efficiency, necessitating reduced latency and guaranteed determinism. Third, high reliability and business continuity: long-cycle training and online inference require high interconnect stability, demanding improved reliability from components, link redundancy, and fault recovery perspectives. Fourth, low power consumption and high energy efficiency: as interface speeds and port counts increase, power consumption and heat dissipation pressure on key chips such as SerDes, DSP (Digital Signal Processor), switching, and optical modules continue to rise, requiring reduced energy consumption per bit transmitted. Fifth, elastic scaling and dynamic scheduling: facing dynamically changing computing power and business traffic, optical interconnects should possess status sensing, on-demand expansion, and dynamic configuration capabilities, collaborating with computing scheduling.
Table 1 illustrates the optical interconnect requirements for typical intelligent computing application scenarios. Overall, intelligent computing optical interconnects are shifting from solely pursuing high speeds to comprehensive optimization across multiple dimensions including bandwidth, latency, reliability, elasticity/agility, and energy efficiency, driving the continuous evolution of technologies such as high-speed optical interfaces, optoelectronic fusion, coherent transmission, and dynamic reconfiguration.
PART.02Key Technologies and Progress in Optical Interconnects for Intelligent Computing Demands
High-Speed Optical Interfaces for Large Bandwidth and High-Density Interconnects
High-speed optical interfaces directly support the large bandwidth and high-density interconnect needs of scenarios such as centralized training within a single DC and cross-DC collaborative training/inference. Their evolution directions mainly include increasing single-channel speed and parallel channel count.
In terms of increasing single-channel speed, data center optical interconnect interfaces are evolving from 400G and 800G to 1.6T, with single-channel speeds rising from 100 Gbit/s to 200 Gbit/s. IEEE (Institute of Electrical and Electronics Engineers) P802.3dj is developing Ethernet interface specifications from 200G to 1.6T based on single-channel 200 Gbit/s. For next-generation interfaces, IEEE 802.3 initiated research related to single-channel 400 Gbit/s (Project 802.3dv) in 2026. Increasing single-channel speed imposes higher requirements on high-speed SerDes, modulators, detectors, FEC, and packaging. Additionally, total bandwidth and bandwidth density can be increased by adding parallel channels. MicroLED features small size, high-density integration, and low power consumption. By using numerous low-speed parallel optical channels, it reduces the speed pressure on single-channel SerDes and devices while improving bandwidth density and energy efficiency. Avicena showcased its 1 Tbit/s LightBundle platform for intelligent computing interconnects at ECOC 2026 (European Conference on Optical Communications), integrating up to 335 independent MicroLED channels with a maximum single-channel speed of approximately 3 Gbit/s. Meanwhile, high parallelism drives pluggable optical modules towards higher bandwidth density. In March 2026, Arista co-founded the XPO MSA (eXtra-dense Pluggable Optics Multi-Source Agreement) with industry partners, proposing specifications for 64-channel, 12.8 Tbit/s high-density liquid-cooled pluggable optical modules, aiming for a front-panel bandwidth density of 204.8 Tbit/s per OCP (Open Compute Project) rack unit. TeraHop demonstrated a 12.8T XPO-8×DR8 optical module at OFC 2026 (Optical Fiber Communication Conference & Exhibition); during ECOC 2026, Accelink further showcased a 12.8T XPO DR liquid-cooled optical module adopting a 64×200G PAM4 architecture.
Therefore, high-speed optical interfaces are evolving from simply increasing single-channel speed to coordinated evolution involving single-channel speed-up, high parallelism, and high-density packaging integration, simultaneously meeting intelligent computing interconnect demands for bandwidth, bandwidth density, and energy efficiency.
Optoelectronic Fusion Technologies for Low Power Consumption and Low Latency
As interface speeds and switching capacities continue to rise, high-speed electrical interconnects between switching ASIC (Application Specific Integrated Circuit) chips and pluggable optical modules face challenges of increased link loss. To ensure signal transmission quality, traditional solutions require introducing functions such as equalization, retiming, and signal processing, bringing additional power consumption and latency overheads. To resolve this contradiction, optoelectronic fusion technologies have developed rapidly.
LPO (Linear Pluggable Optics) removes the DSP/retimer on the module side, having host-side SerDes assume more signal compensation functions. This reduces power consumption and processing latency while maintaining the pluggable form factor, but imposes higher requirements on SerDes performance, link loss, and system co-design. At OFC 2026, OIF (Optical Internetworking Forum) showcased retimed, half-retimed, and linear solutions around 224G electrical interfaces, along with various choices balancing power consumption and link margin. Alibaba Cloud has publicly disclosed the large-scale deployment practice of 400G DR4 LPO in AI intelligent computing networks, providing a paradigm for LPO engineering applications.
NPO (Near-Packaged Optics) deploys optical engines near switching ASIC chips or XPUs, reducing loss and power consumption by shortening PCB (Printed Circuit Board) high-speed electrical traces, while keeping the optical engine relatively independent from the main chip. Industrial progress in NPO accelerated significantly in 2026: Broadcom showcased a 3.2 Tbit/s VCSEL NPO solution at OFC 2026; Coherent showcased a 6.4 Tbit/s silicon photonics NPO optical engine at ECOC 2026. Alibaba Cloud disclosed the latest progress on 3.2T NPO at the 2026 ODCC (Open Data Center Committee) Spring Plenary Session and showcased a ~100T class QFabric switching system composed of four domestic 25.6T switching chips, indicating that NPO is evolving from device verification to system-level application.
CPO (Co-Packaged Optics) moves optical engines even closer to switching ASIC chips, utilizing advanced packaging technologies such as silicon photonics integration, external laser sources, and 2.5D/3D heterogeneous integration to shorten signal paths while improving I/O density and energy efficiency. Broadcom showcased a 102.4 Tbit/s Tomahawk 6–Davisson CPO switching solution at OFC 2026; in May 2026, NVIDIA announced that Spectrum-X Ethernet Photonics based on 200 Gbit/s SerDes had entered production; Coherent further showcased a single-channel 448 Gbit/s InP MZM-SOA array photonic chip for next-generation CPO at ECOC2026.
Overall, LPO focuses on reducing DSP power consumption within a pluggable form factor; NPO further shortens high-speed electrical connection distances through near-packaging but maintains some maintainability; CPO pushes optoelectronic integration to the package level. These three forms are not simple replacements but will coexist in the short term based on bandwidth, power consumption, cost, link margin, and maintainability requirements, evolving in synergy with silicon photonics integration, external laser sources, and advanced packaging technologies.
Large-Capacity Transmission Technologies for Cross-Domain Interconnects
Scenarios such as data ingestion, cross-DC training/inference, and model distribution require crossing campuses, metros, and even longer distances, imposing higher demands on optical interconnects for large-capacity, long-distance, and low-cost transmission. Shorter distances can utilize gray light, CWDM (Coarse Wavelength Division Multiplexing), and DWDM (Dense Wavelength Division Multiplexing) interconnects; as demands for long-distance transmission and capacity increase, coherent transmission technologies are gradually sinking into data center interconnect applications.
For medium-short distance scenarios such as campuses and adjacent data centers, simplified coherent solutions like Coherent-Lite reduce power consumption and cost by lowering DSP and optical component complexity while retaining the advantages of coherent transmission. Marvell jointly showcased a 1.6T Coherent-Lite campus interconnect system with Lumentum at OFC 2026, covering a radius of approximately 2–20 km; during ECOC 2026, Marvell further showcased a 2 nm, 1.6T O-band Coherent-Lite solution.
For metro and longer-distance data center optical interconnects, ZR (Zero Dispersion Reach)/ZR+ pluggable coherent optical modules further reduce the deployment complexity of coherent systems. Currently, 800 ZR/ZR+ is mature and commercially available, and OIF is advancing the formulation of the 1600 ZR specification. During OFC 2026, Marvell released the 1.6T COLORZ1600 ZR/ZR+ pluggable coherent module and 2 nm coherent DSP, pushing pluggable coherent optical modules towards single-wavelength 1.6 Tbit/s evolution.
In September 2026, China Unicom Research Institute released a unified optical interconnect solution for data centers using "Silicon Photonics Microring + Coherent Transmission." Leveraging the small size, low power consumption, and WDM adaptability of microring modulators, combined with the large capacity and long-distance advantages of coherent transmission, it explores adapting to data center interconnects at different distances with a unified architecture.
As cross-DC optical interconnect capacity grows further, optical line systems are also evolving towards higher density. During OFC2026, vendors such as Coherent and Nokia showcased Multi-Rail high-density line system solutions, sharing amplification, gain equalization, and monitoring resources across multiple fiber lines to reduce spatial occupancy and power consumption per line.
Reconfigurable Optical Networks for Elastic Scaling
The communication relationships and bandwidth demands of intelligent computing tasks exhibit obvious dynamism, making fixed optical links difficult to adequately adapt to changes in resource demands. OCS (Optical Circuit Switching) can establish end-to-end optical channels on demand, reducing forwarding power consumption and latency and enhancing network reconfiguration capabilities, showing application potential in scenarios such as training and inference within DCs. OCS includes technical routes such as MEMS (Micro-Electro-Mechanical Systems), digital liquid crystal, piezoelectric ceramics, and silicon photonics waveguides, currently evolving towards higher port density, smaller size, and lower power consumption. Combining OCS with traffic scheduling can further improve training efficiency. China Unicom Research Institute and Beijing University of Posts and Telecommunications jointly published a paper at ECOC 2026, proposing a constraint-aware topology scheduling mechanism for optoelectronic hybrid networks aimed at large model training. By identifying bandwidth and latency bottlenecks, optical connections are prioritized for data parallel or pipeline parallel communications. In experiments, the time for a single training iteration was shortened from 5.24 s to 4.80 s, and the average computing throughput increased from 78.94 TFLOPS to 86.14 TFLOPS.
Besides OCS, AWGR (Arrayed Waveguide Grating Router) can transmit signals of different wavelengths to corresponding output ports based on fixed mapping relationships between input ports and wavelengths, achieving lower-latency optical switching. In the ECOC 2026 Demo Zone live demonstration, Beijing University of Posts and Telecommunications, China Unicom Research Institute, and others jointly showcased a nanosecond-level reconfigurable optical switching platform based on REC-DFB laser arrays and AWGR, validating destination-address-based wavelength selection and dynamic optical path reconstruction among 4 computing nodes.
In cross-DC networks, ROADM (Reconfigurable Optical Add/Drop Multiplexer) can dynamically configure wavelengths and channels, enabling flexible establishment and reconfiguration of cross-site optical channels. Compared to fixed optical links, ROADM can dynamically adjust wavelength connections based on business relationships and bandwidth demands between different DCs, supporting multi-directional networking and flexible routing. This helps improve the utilization of cross-DC fiber and wavelength resources, adapting to dynamic changes in training and inference businesses. Additionally, the industry is exploring synergistic networking applications of OCS+WSS (Wavelength Selective Switch).
Overall, technologies such as optical switching are driving optical interconnects to evolve from fixed connections to on-demand reconfigurability, and further combining with computing scheduling to achieve joint configuration of computing and network resources.
PART.03Conclusion
Intelligent computing scenarios such as data ingestion, model training, model distribution, and model inference impose different demands on optical interconnects, driving them towards development in large bandwidth, low latency, high reliability, low power consumption, and dynamic reconfigurability. In the future, optical interconnects will further extend from traditional data transmission links into the interior of computing systems, coordinating with computing resource scheduling to become a critical support for large-scale intelligent computing infrastructure.
Editor/Layout: Sun Tian
Proofreader: Mei Yaxin
Supervisor: Liu Qicheng