As AI large models, intelligent computing clusters, and agent applications expand rapidly, computing power networks have evolved into integrated infrastructure for hierarchical coordinated scheduling. Computing interconnects are divided into three tiers: Scale-Up (vertical expansion within super-node racks), Scale-Out (horizontal expansion across multiple racks within a data center), and Scale-Across (inter-regional wide-area computing interconnect), as shown in Figure 1. With core advantages of ultra-large capacity, ultra-low latency, and long-distance transmission, optical communication has become the foundational carrier for the three-tier architecture of "National Computing Hubs – Regional Computing Centers – Metro Edge Computing."
However, most existing computing interconnect network standards originate from specifications in traditional Ethernet, optical transmission, and DCI (Data Center Interconnect) domains. They lack sufficient coverage of capabilities such as high-performance collective communication for AI distributed training, low-latency lossless transmission, computing-awareness, and computing-network collaborative scheduling.
This article systematically reviews the current status of optical communication standards for computing power networks and industry pain points, proposes an upgrade strategy for the standard system, and provides standardized support for the coordinated planning and construction of China's next-generation optical communication network and national integrated computing power network.
1. Scale-UpCurrent Status and Pain Points of Optical Communication Standards for High-Speed Interconnect Between Compute Cards Within Super-Nodes
Scale-Up refers to short-distance interconnect between XPU chips (referring to all heterogeneous computing chips involved in AI computing scheduling within data centers) inside racks or super-nodes. Link lengths are typically within several meters, aiming to interconnect numerous XPUs into super-computing nodes to support demanding collective communication tasks such as All-Reduce and All-to-All for tensor parallelism and expert parallelism.
Current Standard Status
Currently, Scale-Up interconnect still primarily relies on copper cables and PCB (Printed Circuit Board) electrical interconnect. Proprietary protocol technologies like NVLink iterate and industrialize rapidly, but chips and optical engines from different vendors cannot interoperate, resulting in a fragmented landscape characterized by "proprietary protocols leading, open standards lagging." IEEE (Institute of Electrical and Electronics Engineers) 802.3 series standards mainly target rack-to-rack interconnect, lacking definitions for chip-level ultra-low latency and hardware-accelerated collective communication.
With SerDes technology iteration, single-channel transmission rates have increased to 224 Gbit/s, highlighting bottlenecks in electrical interconnect signal attenuation and power consumption. CPO (Co-Packaged Optics) and NPO (Near-Package Optics) have become key technical routes for Scale-Up. Internationally, OIF (Optical Internetworking Forum) and IEEE 802.3 working groups have initiated research on physical layer and mechanical interface standards for CPO/NPO, defining specifications for optical engines, fiber arrays, connectors, and electrical interfaces.
Domestically, the Standards Institute of CAICT (China Academy of Information and Communications Technology) led the establishment of the domestic NPO standards working group in mid-July 2026, jointly advancing the development of standards for optical engine packaging, heat dissipation, interface interoperability, and reliability testing methods. Open interconnect industry alliances such as OISA (Omni-directional Intelligent Sensing Express Architecture) and ESUN (Ethernet for Scale-Up Networking) promote open high-speed interconnect protocols to reduce binding to proprietary bus technologies.
Core Pain Points
First, protocol ecosystem fragmentation creates technical lock-in via proprietary interconnect protocols, with insufficient maturity of open standards. Scale-Up directly interfaces with XPU chips, preventing direct interoperability between super-nodes from different vendors, which hinders pooled scheduling of computing resources.
Second, CPO and NPO optical interconnect standards are incomplete. Co-packaging optical engines with switch chips imposes new requirements on heat dissipation, power supply, fiber arrays, and field maintenance. Existing standards focus more on optical components themselves, lacking specifications for whole-system reliability, fault localization, and maintainability. Inconsistent specifications for optical engines, fiber arrays, and connectors across vendors make interconnection difficult.
Third, there is a lack of quality standards for AI collective communication. Only bit error rate and bandwidth metrics are defined, while unified evaluation systems for latency jitter, tail latency, and lossless transmission for All-Reduce and All-to-All are missing, making it impossible to quantitatively assess the support capability of optical interconnect for AI large model training tasks.
Fourth, performance metric evaluation specifications for power consumption and reliability are missing. Power consumption per unit bandwidth is a core metric for Scale-Up, but the industry lacks unified test methods. Vendor test conditions vary, making horizontal performance comparison difficult.
Upgrade Directions
One is to accelerate the standardization of open high-speed interconnect protocols for Scale-Up, reducing risks of proprietary protocol lock-in and balancing open ecosystems with chip performance needs; two is to refine CPO and NPO series standards, unifying specifications for optical engines, fiber arrays, connectors, power/heat dissipation, fault detection, and O&M interfaces, and establishing multi-vendor interoperability testing and certification systems; three is to introduce new AI collective communication metric systems and performance testing methods, incorporating computing business performance into network transport quality evaluation systems; four is to establish standardized testing specifications for power consumption and reliability of Scale-Up optical interconnect, forming a unified performance benchmarking baseline.
2. Scale-OutCurrent Status and Pain Points of Optical Communication Standards for Cluster Interconnect Within Data Centers
Scale-Out refers to interconnect ranging from tens of meters to several kilometers between racks and pods within the same intelligent computing data center, constructing topologies like CLOS to support thousand-card or ten-thousand-card level AI cluster distributed training. Scale-Out optical interconnect is currently the largest deployed optical communication scenario in intelligent computing centers, primarily using pluggable optical modules.
Current Standard Status
The standardization and industrial maturity of 400G/800G/1.6T optical modules in Scale-Out are relatively high. Internationally, IEEE 802.3 is the basic standard for Scale-Out Ethernet physical layers, having completed 400G and 800G Ethernet standards and currently developing 1.6T Ethernet standards. OIF and IPEC (International Photonics & Electronics Committee) develop implementation agreements for 400G/800G DR (Datacenter Reach) and FR (Fiber Reach) series short- and medium-reach pluggable optical module physical layers. Packaging forms such as OSFP (Octal Small Form-factor Pluggable) and QSFP-DD (Quad Small Form-factor Pluggable-Double Density) are specified by MSA (Multi-Source Agreement) organizations. Domestically, CCSA (China Communications Standards Association) TC6 working groups simultaneously develop industry standards for high-speed optical modules in data centers, strengthening preliminary research on standards for new optical interconnect module devices in intelligent computing centers.
Core Pain Points
First, engineering barriers remain in multi-vendor interoperability of optical modules. Although IEEE and MSA define optical module physical layers, differences in lasers, receiver chips, and DSP (Digital Signal Processing) algorithms among vendors often lead to issues with signal margin and interoperability stability in high-speed 800G and 1.6T cross-vendor scenarios. Additionally, standards for optical module fault diagnosis and link performance monitoring are incomplete, resulting in fragmented O&M methods.
Second, computing business awareness capabilities are lacking. Current optical transmission and Ethernet standards belong to the data plane; the network does not know whether it carries large model training traffic or inference traffic, preventing dynamic adjustment of bandwidth and priority based on computing business intent. This causes network resource scheduling to fail in quickly adapting to upper-layer computing scheduling demands.
Third, there are no network quality evaluation standards for large-scale clusters. For ten-thousand-card cluster scenarios, the industry lacks unified testing standards to quantify Incast congestion, tail latency, and collective communication throughput. Cluster performance assessment relies on vendor-specific testing tools, resulting in poor comparability of evaluation results.
Upgrade Directions
One is to refine interoperability standards for high-speed pluggable optical modules, supplementing system-level interoperability testing specifications and establishing unified standardized interfaces for link diagnosis, performance monitoring, and fault management; two is to promote the implementation of computing-aware network standards, defining computing business identifiers and business intent interfaces so that optical bearer networks can identify AI training and inference services and provide differentiated deterministic network services; three is to formulate network performance testing standards for intelligent computing clusters, establishing unified evaluation methods for collective communication, tail latency, and congestion performance.
3. Scale-AcrossCurrent Status and Pain Points of Optical Communication Standards for Inter-Regional Wide-Area Computing Interconnect
Scale-Across refers to wide-area computing interconnect across machine rooms, regions, and even computing hubs, distinct from traditional DCI data links. The goal of Scale-Across is to integrate multiple geographically dispersed intelligent computing centers into a logically unified super AI computing infrastructure, supporting cross-region distributed large model training and inference applications. Link distances range from tens to hundreds of kilometers, relying on high-speed, large-capacity WDM (Wavelength Division Multiplexing) transmission systems and OTN (Optical Transport Network) technology as the backbone computing transmission channels of the national integrated computing power network.
Current Standard Status
The traditional DCI standard system primarily consists of OIF's ZR (Zero Dispersion Reach) series and ITU-T's OTN series. OIF has released 400ZR and 800ZR specifications and will soon release 1600ZR and ZR+ specifications, defining single-wavelength 1.6T coherent pluggable modules for short- and medium-reach data center interconnect. ITU-T's G.709 series standards specify traditional OTN and FlexO frame structures, mapping protocols, and forward error correction interface functions. In July 2026, the SG15 plenary meeting reached consensus on B1T (Beyond 1 Terabit) OTN technology standards and completed related revisions. G.709.1 defines the general architecture for B1T networks, while G.709.5, G.709.6, and G.709.7 specify short-reach, long-reach, and chip-to-chip MFI interfaces respectively, supporting 100G timeslots and significantly optimizing overhead ratios to improve carrying efficiency for 100GE~1.6TE Ethernet services, adapting to backbone network computing carrying needs. IETF (Internet Engineering Task Force) has released standards for SRv6 (Segment Routing over IPv6), network slicing, and CCAMP (Common Control And Measurement Plane) optical network control, recently researching computing-aware optical network architectures aimed at linking computing orchestration with optical network control planes. CCSA TC6 working groups continue to develop WDM and OTN series optical transmission standards and have released group standards for computing network transport capacity index evaluation and delay circle monitoring technology, supporting the construction of backbone transmission channels for the national integrated computing power network.
Traditional DCI standards were designed to carry cloud computing and other businesses for data backup. However, Scale-Across targets cross-region distributed training, requiring the carrying of large model gradient synchronization and other businesses, imposing higher requirements for deterministic bandwidth, low latency jitter, lossless transmission, and dynamic optical channel scheduling.
Core Pain Points
First, standards for computing-network collaborative orchestration and operational platforms are missing. Computing scheduling systems cannot perceive optical link latency and bandwidth resources; optical networks cannot identify computing task types and priorities, failing to dynamically establish or adjust optical transmission channels in real-time based on task requirements.
Second, norms for cross-operator computing dedicated line services are inconsistent. Activation processes and interface specifications for computing dedicated lines across network domains vary, leading to long activation cycles that cannot meet elastic computing scheduling demands.
Third, computing transmission quality evaluation systems are missing. Optical transmission networks only monitor bit error rate, packet loss rate, and link availability, lacking quantitative evaluation standards for the impact of latency, jitter, and service interruption time on various AI computing businesses during training and inference tasks, constraining cross-region computing market trading.
Fourth, network resilience standards for backbone multi-route protection await improvement. Computing businesses require multi-route backup, but existing network protection standards lack mechanisms for fast switchover, lossless protection, and task-aware protection specific to computing businesses.
Upgrade Directions
One is to build computing-network collaborative control standards, formulating unified interface specifications between computing orchestration and optical networks to achieve distribution of computing business intents and reporting of link status, supporting on-demand dynamic creation of deterministic computing optical channels and promoting standardization of computing-aware optical network technology; two is to unify cross-operator computing dedicated line standards, clarifying SLA (Service Level Agreement) metric systems, service activation, and fault handling processes for computing dedicated lines, enabling cross-operator computing scheduling capabilities; three is to formulate transmission quality evaluation standards for computing businesses, distinguishing indicator thresholds for offline training, online inference, and other services to support computing market trading; four is to improve optical network resilience standards, optimizing optical-layer multi-route protection, fast fault switchover, and lossless protection mechanisms to ensure stable operation of cross-region computing businesses.
4. Overall Upgrade Strategy for Optical Communication Standardization System for Computing Power Networks
The collaborative optical interconnect networks of Scale-Up, Scale-Out, and Scale-Across constitute a complete computing bearing foundation. Building an optical communication standardization system for computing power networks requires strengthened top-level design, hierarchical implementation, cross-layer coordination, and overall layout.
First, construct a top-level framework for optical communication standards oriented towards computing power networks. Clarify business requirements and metric systems, compile standard system guidelines, and build a complete standard system based on four categories: basic physical layer standards, network protocol standards, computing-network collaborative control standards, and performance testing and SLA standards, guiding standard project planning.
Second, adhere to hierarchical differentiated technical standard breakthroughs. Scale-Up focuses on overcoming open high-speed interconnect and CPO/NPO optical engine interoperability standards; Scale-Out focuses on ultra-large-scale lossless clusters, high-speed pluggable optical modules, and computing-aware standards; Scale-Across prioritizes high-speed large-capacity coherent optical transmission, computing-network collaborative orchestration, and cross-operator computing dedicated line standards.
Third, strengthen the development of computing-network collaborative standards. Break through traditional hierarchical thinking, establish cross-layer interface standards for "Computing Business – Network – Optical Transmission," enable unified cross-domain transmission of computing task information in optical networks, and support end-to-end computing business SLA guarantees.
Fourth, build a unified testing, verification, and certification system. Establish a national-level intelligent computing optical interconnect testing and verification platform, formulate interoperability testing and performance evaluation specifications for different scenarios, conduct equipment and solution certifications, and resolve multi-vendor interoperability challenges.
Fifth, improve network security and resilience supporting standards for computing power. Incorporate network reliability, link security encryption, and lossless protection mechanisms into the optical communication standard system, combining with the planning and construction of the national "New Eight Vertical and Eight Horizontal" cable network and computing transmission channels to improve specifications for optical network survivability and disaster recovery.
5. Conclusion
While optical communication technology standards continue to iterate, they face challenges such as protocol fragmentation, insufficient computing-network collaboration, and missing computing business evaluation systems when confronted with the stringent communication demands of AI large model distributed training. To meet the construction and development requirements of the national integrated computing power network, it is necessary to establish a top-level standard framework for next-generation optical communication technologies, advance hierarchical breakthroughs in intelligent computing optical interconnect technology standards, strengthen computing-network collaborative standard development, and build a unified testing and certification system. By prioritizing standards, we can break down computing scheduling barriers, support the realization of computing transmission channels for the national integrated computing power network, promote free cross-region computing scheduling, and solidify the computing foundation for the development of new quality productive forces.
This article was published in *Communications World*, Issue 10, 2026; Original Title: "Current Status, Pain Points, and Upgrade Directions of Optical Communication Standardization System for Computing Power Networks."
Author: Li Fang, Institute of Technology and Standards, China Academy of Information and Communications Technology