Specific technical requirements for optical cables in AI data centers

The explosion of AI data centers (intelligent computing centers) is completely reshaping network cabling standards. The amount of internal interconnection optical fibers used is 5 to 10 times that of traditional data centers. In order to meet the rigid requirements of high throughput and zero packet loss for tens of thousands of GPU clusters, the AI ​​data center has put forward four extremely stringent technical requirements for optical cables:

1. Extreme space density: smaller, denser, and softer. Space in AI racks and pipes is extremely scarce, and a single switch rack may need to accommodate more than 20,000 optical fibers. Miniaturized outer diameter: It is generally required to use 200 micron outer diameter coated optical fiber (traditional 250 micron), and the overall thickness of the optical cable can be reduced by up to 60% to alleviate cable trough congestion and ensure air heat dissipation. New ribbon structures: Mainstream adoption of flexible ribbon cable technologies such as Corning's Flow Ribbon has greatly increased the number of bundled cores of micro-optical cables in the pipeline. Bending insensitivity: For high-density optical fibers in a limited cabinet space, G.657.A1 or G.657.A2 level bend-insensitive optical fiber must be used to ensure that no obvious signal bending loss is caused at a very small bending radius.

2. Special performance under CPO architecture: polarization maintenance and low loss. As computing power evolves to 800G and 1.6T, the transmission distance of traditional copper cable (DAC) is limited to less than 3 meters, and photoelectric co-packaging (CPO) technology has become a core trend. Large-scale application of polarization-maintaining fiber (PMF): The relationship between the external light source and the silicon photonic chip under the CPO architecture is extremely sensitive to the polarization state of the light. According to industry authoritative analysis, the demand for polarization-maintaining optical fiber in intelligent computing centers is experiencing an explosive growth of 10 to 20 times to ensure stable connections and suppress insertion loss. Large effective area and ultra-low loss: Large model training requires maintaining high power input for a long time. The optical fiber needs to have a large effective area to suppress nonlinear effects, and must have ultra-low loss (ULL) performance to reduce the overall energy consumption of the link.

3. Cross-campus interconnection (DCI) extremely low-latency AI distributed training requires that the network delay between cross-data centers (DCI) is infinitely close to zero. Hollow core fiber (HCF): As a disruptive technology in the AI ​​era, the transmission speed of light in the air of hollow fiber is about 35% faster than in traditional quartz glass. This can reduce network latency by more than 30% and has ultra-high power tolerance.

4. High-precision manufacturing and material technology, precision dispensing and reliability: Because AI optical fiber components are extremely fine and dense, major optical cable manufacturers (such as YOFC, FiberHome Communications) have introduced high-precision fluid dispensing processes such as Nordson EFD (Nordson) into production to ensure that in large-scale automated manufacturing, every tiny optical fiber bonding point has extremely high mechanical reliability and long-term stability.