EN
Every part of the network is being improved by artificial intelligence. Massive training jobs for large language models, running distributed inference engines, and syncing vast amounts of data across clusters of computers put unprecedented load on the interconnects that connect data centers. 100G is no longer good enough for AI workloads and 400G is quickly becoming the new baseline instead of the target. Enter 800G DCI-the new standard for AI data center interconnection.
The Rising Need for 800G DCI
The race to 800G DCI is being driven by a confluence of factors. Firstly, the size of AI models has ballooned. Parameters have gone from millions to billions and trillions. Spreading those models across many different pods in different data centers requires enormous bandwidth to flow between the data centers. Secondly, distributed training now places data center pods across an entire metro area into the same compute cluster, rather than within a single building. Thirdly, real-time applications such as fraud detection and autonomous systems require both tremendous throughput and consistent, deterministic low latency. Finally, with the trend towards private AI infrastructure, the cumulative load on interconnect links continues to grow. Traditional 100G and 200G DCI technologies, while acceptable for conventional workloads become bottlenecked in AI workloads. 800G DCI addresses the bandwidth need by providing the room to grow.
Bandwidth Pressure in AI Data Centers
The nature of AI training traffic puts an atypical amount of strain on interconnect links compared to normal applications. During training, model gradients must be synced across all participating compute nodes at the end of each batch. This all-reduce traffic consists of synchronized traffic bursts which have the potential to saturate even high capacity links. Data can also be checkedpointed to distributed storage at periodic intervals, adding load. Data preparation pipelines can be a constant stream moving data between tiers of storage and the compute pods. When all three events occur simultaneously, aggregate link loads become multiplied. 800G DCI addresses these stresses by sending an entire all-reduce burst across the interconnect with a single wavelength without queuing. Reducing backpressure through the interconnect allows for a linear scale-out of the AI training job.
Low Latency Matters for AI Services
Bandwidth is not enough to solve the AI interconnect problem-low latency is crucial too. Across data centers, communication latency impacts the overall training throughput because compute nodes may be forced to wait on their peers in the compute cluster, thus yielding precious GPUs. High latency can lead to a poor user experience for AI services like autonomous systems, fraud detection, and real-time inference. 800G DCI reduces latency through a faster symbol rate(reducing packet latency) and more advanced forward error correction. Furthermore, optical paths are direct, avoiding routers that add delay to the packet’s journey. When directly connecting AI compute clusters over 800G DCI, the latency between data centers can be reduced from milliseconds to microseconds-truly enable distributed training to act like a single compute cluster spread across a campus.
Key Benefits of 800G DCI Upgrade
800G DCI has an advantage on many fronts compared to lower speeds. Its spectral efficiency is double that of 400G, meaning operators can double the throughput on the same fiber infrastructure. Power consumption also decreases because 800G has a lower energy consumption per transmitted bit compared to running two 400G links. With a single interface, port density on switches and routers also increases significantly, freeing up capacity for compute connectivity. Operational efficiency is improved by simplifying wavelength management. The new 800G speeds also future proof networks because current demand can be easily met and it will take longer before operators need to upgrade their interconnects again as AI traffic increases by an expected 50% each year. For cloud and data center operators managing tenants requiring large amounts of data transfer, 800G DCI has a business advantage.
Flexible Deployment for Fast Expansion
The speed at which AI systems are being deployed means that the interconnect layer can become a roadblock. 800G DCI offers flexible deployment options. For longer reach within the metro network, optical line systems are capable of extending transmission distance while maximizing data rate. 800G pluggable optics that fit into standard switch interfaces offer the fastest deployment-just insert, connect, and provision. Organizations building out new data center campuses can leverage purpose built 800G DCI platforms with greater control over management and performance monitoring. Complementing these solutions, Sino-Telecom’s DCI-BOX device delivers an ultra-compact, pre-integrated, and cost-effective DCI solution tailored for edge data centers and rapid service rollouts. Supporting up to 800G per channel, the DCI-BOX minimizes space and power consumption while enabling plug-and-play deployment. Whether used as a standalone unit in remote POPs or as part of a larger mesh network, the DCI-BOX ensures that capacity scaling keeps pace with AI-driven demand, reducing deployment time from days to just hours. Regardless of operational scenario, 800G DCI can be deployed in hours rather than days, keeping pace with the growth and scale of AI infrastructure.
Ideal 800G Options for ISPs and DCs
Different organizations have unique requirements for 800G DCI. AI training data centers focused on scaling their compute requirements need high throughput and low latency so point-to-point 800G DCI between clusters are ideal. ISPs who provide interconnection to multiple AI clients require strict wavelength isolation to prevent traffic bursts from one client affecting others. 800G DCI platforms that are configurable with multiple transponders will allow ISPs to assign and bill wavelengths individually. Every customer also needs to ensure reliability so that bit errors do not force compute jobs to restart from the last checkpoint.
Ready-Made 800G Solutions for Growth
Sino-Telecom’s 800G DCI solutions are tailored to the needs of AI data center interconnection. For longer haul networks or multiple hops we offer DCI platforms that are integrated with amplification, dispersion compensation, and channel monitoring. Our systems offer hitless capacity adjustment and detailed diagnostic tools as well as unified management interfaces. We hold over 240+ patents and have 150+ software copyrights and work with networks across more than 30 different countries including France, Brazil, Australia and Indonesia. Our product portfolio also includes a full suite of 200G, 400G, and 800G pluggable optics for a comprehensive solution from switch port to fiber optic line.