Fiber Array Silicon Photonics: How Hirundo's Precision FA Bridges CPO and AI Supercomputing in 2026

Why Optical Interconnect Has Become the Bottleneck of the AI Era

As the parameter scale of large AI models expands toward the trillions and supercomputing centers race toward Exascale (10¹⁸ operations per second) and Zettascale (10²¹ operations per second) computing power, data transmission capacity has become a core bottleneck restricting computing performance breakthroughs. Conventional electrical signal transmission suffers from limited bandwidth and excessive energy consumption, making it unable to meet the demand for instantaneous mass data flow in supercomputing facilities. Against this backdrop, the coordinated development of Co-packaged Optics (CPO) and silicon photonics has opened a new track of optical interconnection for computing infrastructure — and fiber arrays serve as the critical optical bridge connecting these two technologies to power AI supercomputing.

Core Definition of FA Fiber Arrays

A fiber array (FA) is a standardized fiber assembly fabricated by arranging, fixing and polishing the end faces of multiple optical fibers on a high-precision V-groove substrate at fixed pitches (e.g., 250 μm). It acts as a collective transmission channel for optical signals, integrating numerous single-lane optical fibers into a multi-lane high-speed optical highway.

The fabrication process involves placing the bare fiber portion (with the fiber coating removed) into the V-groove, pressurizing it with a pressurizing device, bonding it with adhesive, and finally grinding and polishing the surface to the required precision. This level of control is precisely what allows fiber arrays to deliver their three core functions:

• Enable efficient optical coupling of photonic devices (e.g., interconnection with optical modules and wavelength division multiplexers)
• Reduce alignment errors from single-fiber mating and enhance communication stability
• Deliver standardized interfaces to lower mass production barriers

Fiber Arrays: The Indispensable Optical Bridge for CPO and Silicon Photonics

Silicon photonic chips generate and process optical signals on-chip, while CPO packages the optical engine directly alongside the ASIC on the same substrate. To make these two worlds talk to each other, a precision physical interface is required — and this is exactly the role of the fiber array.

1. Solving the "Integration and Docking" Problem

Through precise photolithography and packaging processes, fiber arrays can achieve a perfect match between fiber spacing and the optical ports of silicon photonic chips. A single FA simultaneously supports the parallel transmission of dozens or even hundreds of optical signals, accurately guiding the high-speed optical signals generated by the silicon photonic chip into the external optical link, and also efficiently transmitting external optical signals back into the chip. In effect, it becomes a "signal relay station" between the silicon photonic chip and the CPO packaged module.

2. Adapting to the Low-Latency Requirements of CPO

The core advantage of CPO technology is "near-package transmission," which reduces latency by shortening the distance between the optical module and the computing chip. Fiber optic arrays, with their compact structural design, can be directly integrated into the CPO package, enabling a "short-distance direct connection" of optical signals between the silicon photonics chip and the optical module. Compared to traditional fiber optic connections, this significantly reduces optical signal transmission losses, further amplifying the low-latency advantage of CPO and providing a "stutter-free" transmission guarantee for high-frequency data interaction in AI supercomputing.

3. Meeting the Large-Scale Demands of AI Supercomputing

The ever-expanding scale of AI supercomputing centers places stringent requirements on the "large-scale deployment" and "long-term stability" of optical transmission links. Hirundo's fiber arrays possess high reliability and high consistency, can be mass-produced with minimal performance deviation, and can meet the needs of large-scale optical link deployment in supercomputing centers. Simultaneously, their flexible array specification design can be adapted to the number of ports on the CPO and silicon photonics chips, reserving expansion space for future computing power upgrades and avoiding limitations on computing power enhancement due to transmission link constraints.

Hirundo's FA Product Portfolio: Layered Support for Every Data Center Scenario

Customized FA solutions deliver layered support for full-scenario data center networking, fully compatible with all-rate optical modules. Optical module selection varies across use cases from general cloud computing rooms to AI supercomputing hubs, and tailor-made FA designs precisely match end-to-end transmission requirements, covering three core application scenarios.

400G/800G Parallel Optical Modules: MT-FA as Standard for SR/DR Series

General cloud data centers mainly deploy 400G and 800G SR8/DR8 parallel optical modules, which rely on MT-type FAs to realize 8-channel parallel transmit-receive. Each channel carries 100G PAM4 signals with insertion loss controlled below 0.35 dB, return loss ≥ 60 dB, and channel uniformity deviation under 0.2 dB. When running 24/7 at full load, the bit error rate drops to as low as 10⁻¹², ensuring lossless data transmission for massive virtual machines and cloud storage. Compared with discrete fiber assembly, optical modules equipped with standardized FAs achieve higher mass production yields and cut manual assembly costs by over 30%, well-suited for large-scale bulk procurement by cloud vendors.

Next-Generation 1.6T Silicon Photonics / CPO Optical Modules: Custom FAs Unlock New Integration Potential

1.6T will be the primary interconnection specification for next-generation AI computing clusters. CPO integrates optical engines and ASIC chips on the same substrate, shortening optoelectronic transmission paths and drastically reducing power consumption. Detachable D-FAU fiber arrays and integrated coupling FCFA arrays are essential supporting components for CPO commercialization. Custom ultra-thin FAs directly couple with silicon photonic waveguides and eliminate intermediate transfer optical paths, cutting overall module power consumption by 15% to 40%. The detachable design breaks the one-time packaging limitation of traditional FAs: optical engine maintenance and module upgrades do not require full disassembly, substantially lowering operation and maintenance costs for data centers. These products have already been mass-deployed in optical module projects supporting AI intelligent computing switches of leading manufacturers.

Medium-to-Long-Range WDM Optical Modules: WA-Type FAs Empower Wavelength Division Multiplexing

For FR/LR series wavelength division multiplexing optical modules used in cross-data-center and inter-city interconnections, dedicated WDM-FAs work with Arrayed Waveguide Gratings (AWGs) to realize multi-wavelength multiplexing and demultiplexing. A single component integrates multiple wavelength optical paths to deliver ultra-large bandwidth long-distance transmission with fewer optical fibers, helping telecom operators reduce capital expenditure on backbone fiber deployment and supporting cross-regional computing resource scheduling.

Key Features and Advantages of Hirundo's Fiber Arrays

• High precision, multi-channel: Hirundo's FA encapsulates several or even dozens of optical fibers together in a very small space without interference between them. The use of V-groove substrates ensures precise control over fiber spacing and position, achieving efficient optical signal coupling and transmission.
• Low insertion loss: Through precision machining and polishing processes, insertion loss can be controlled to within 0.05 dB, significantly improving the efficiency of optical signal transmission.
• High return loss: Return loss is typically ≥55 dB (APC), effectively reducing optical signal reflection and ensuring transmission stability.
• Excellent temperature stability: With optical fibers, glass substrates and cover plates as base materials, FAs operate and store reliably from −40℃ to +85℃, adapting to harsh environments.
• Customization: Hirundo offers arrays with 8°, 0°, 45°, and custom angles to meet diverse customer requirements, while array specifications can be flexibly configured from 1CH to 64CH.

Upgrade Computing Infrastructure First: Precision Fiber Arrays from Hirundo Optics

Hirundo specializes in supporting 400G/800G/1.6T optical modules for data centers, offering full-spectrum customizable fiber array solutions:

• Micron-level precision V-groove manufacturing, flexible customization of 1 to 64 multi-channel designs with tightly controlled pitch tolerances for drastically improved coupling yields
• Multiple end-face polishing angles (0°/8°/42.5°/45°) perfectly matched with SR/DR parallel and high-speed silicon photonic optical modules, featuring low insertion loss and excellent channel consistency
• Self-developed detachable CPO-specific FAs and polarization-maintaining special FAs, compatible with innovative co-packaged optics architectures to advance new-generation optical interconnection for AI intelligent computing centers

Outlook: Fiber Arrays as the Foundation of the Optical Computing Era

As silicon photonics, CPO and AI supercomputing converge, the fiber array is no longer a passive afterthought but the enabling interface that determines how much of a chip's optical bandwidth can actually reach the network. From exascale data centers to chip-scale vertical coupling via 45° arrays, the road ahead runs through micron-level precision V-groove manufacturing and reliable, mass-producible optical coupling. With its high-performance fiber array portfolio spanning 400G, 800G and 1.6T applications, Hirundo is positioned to help data center operators and module makers convert AI computing demand into real, deployable optical interconnect capacity.