When discussing co-packaged optics (CPO), the industry most often shows an architecture diagram of "the optical engine moved next to the ASIC." It does explain why CPO can shorten electrical links and reduce some power consumption, but when the system scales from a single module to a 51.2T or 102.4T switch, what determines manufacturability and maintenance cost is often no longer that architecture diagram, but the four categories of components around it: external laser sources, polarization-maintaining fiber, detachable connectors, and thermal structures.

Engineering concept illustration generated based on the article topic
Figure 1: AI conceptual illustration for this article, generated by MiniMax based on the article topic, used to explain the application context and not representing a specific model, size, spectrum, coating system or acceptance performance.

This judgment is not derived from conceptual reasoning. The OIF 3.2T CPO module implementation agreement already defines the external laser version as using four or eight polarization-maintaining fibers; the public reference designs from Marvell and SENKO place 1,152 fibers, 32 detachable connectors, 16 external laser modules, and a cold plate in the same system.[1][2][3]

Industry insightsEstimated reading time: 9 minutesIntended readers: R&D / Process / Procurement

Key Takeaways

  • The module boundary of CPO has been standardized, but connectors have not been fully standardized.OIF defines the electrical, optical, mechanical and test interfaces of the 3.2T module, while also making clear that optical connectors are not specified by this implementation agreement, and that solutions such as MPO are only possible system implementations.[1]
  • External lasers are not simply about "moving the laser farther away."Four or eight CW lasers must enter the module through polarization-maintaining fiber and then be split within the module to four or eight transmitters; splitting alone may consume about 9 dB, modulation consumes another about 3 dB, and the remaining power must cover coupling, waveguide, multiplexing and manufacturing margins.[1]
  • Connectors begin to determine maintainability.Marvell's reference compute tray pairs each XPU with four 6.4T optical engines, and each engine connects to two 36-fiber detachable connectors; the entire tray has 32 connectors and 1,152 fibers in total. Here, repeatable alignment and cleaning are not accessory issues but system reliability issues.[3][4]
  • The production threshold is multi-physics co-design.The reliability targets given by the standard framework, the power consumption example of the 3.2T optical engine and the liquid-cooling cold plate height show that CPO competition is shifting from chip bandwidth to the coordinated capabilities of fiber, connection, thermal and testing.[2]

The "four or eight fibers" already written into the OIF standard

The OIF 3.2T CPO module is defined as a basic building block of a 51.2T switch: one system uses 16 3.2T modules placed near the switch ASIC. The protocol lists multiple module forms, including 8×400GBASE-DR4, 8×400GBASE-FR4, and internal-laser or external-laser versions.[1]

Module Form FactorLine SideLaser SourcesSystem Implications
400G-DR4 / FR4 IL8×400GIn-module integrationThe module is more compact, but the laser thermal load enters the optical engine package
400G-DR4 / FR4 EL8PMF8×400GExternal, 8 polarization-maintaining fibersThe light source is maintainable, but the number of fibers and alignment points increases
400G-DR4 / FR4 EL4PMF8×400GExternal, 4 polarization-maintaining fibersReduce the laser feed into the fiber core, but a single fiber carries higher power and bears stronger splitting loss

The protocol specifically notes that when four fibers are used, one example is about 9 dB splitting loss and about 3 dB modulation loss, with the remaining budget allocated to module input/output coupling, splitters, waveguides, modulators, and multiplexers. This budget structure shows that external lasers are not a free reliability upgrade; they shift more optical power management pressure onto fibers, connectors, and the interior of the package.[1]

Why OIF external lasers use polarization-maintaining fiber

In OIF 3.2T external laser modules, the protocol explicitly specifies PM fiber as the feed optical medium rather than standard single-mode fiber. This is a specific architectural choice and should not be extrapolated to mean that all CPO must use PMF; for systems adopting this type of external laser interface, CW light entering the silicon photonic module still undergoes splitting, modulation, and multiplexing, and the system must keep the polarization state predictable under variations in connection, coiling, temperature, and mechanical stress. PANDA fiber creates birefringence by placing stress rods on both sides of the core, and vendor materials list it for co-packaged optics, coherent DWDM, and fiber laser applications.[7]

However, "polarization-maintaining" does not mean "connect any way." For the supply chain, at least the following must be confirmed simultaneously: operating wavelength and cutoff wavelength, beat length, crosstalk or extinction ratio, cladding diameter, coating, bend radius, connector keying, and axial alignment. Heracle's PANDA datasheet offers 80 μm and 125 μm cladding options, with extinction ratio specifications given per 100 meters; this means system vendors must incorporate length, coiling radius, and connectors into the budget together, rather than copying a single polarization extinction ratio figure.[7]

Polarization-maintaining fiber parametersWhy it matters for CPORequired for quotation
Operating/cutoff wavelengthDetermine whether the target wavelength band is truly single-mode and whether higher-order modes occurLaser center wavelength, wavelength range, bending condition
Beat length and birefringenceDetermines polarization-maintaining capability and sensitivity to disturbancesLength, coil radius, temperature cycling
Extinction ratio/crosstalkDetermines polarization purity before entering the modulatorTest length, connector, axis definition
Geometry and coatingDetermine connector alignment, end-face quality, and long-term mechanical reliabilityCladding, coating, keying, mating durability

The value and cost of external lasers

The OIF framework notes that external lasers can improve the maintainability of CPO switches: failed lasers can be replaced, and the system can partially compensate by increasing ELS output power; integrated lasers reduce interface loss, but laser heat and failures are brought into the optical engine.[2]This is a system trade-off, not "external always better than integrated."

At the component level, SENKO positions ELSFP as a hot-pluggable external laser connection solution: one end is a host connector fixed on the CPO switch board, and the other is a module connector that plugs and unplugs with the laser module, blind-mating up to two sets of low-loss MT ferrules through a self-aligning structure. This can reduce maintenance downtime, but "hot-pluggable," "lowest insertion loss," and "more reliable" remain supplier targets and must be verified through repeated mating, contamination, end-face inspection, and random sample testing.[8]

Reliability targets also illustrate the point. The framework gives an example modeled on a 51.2T switch with 16 3.2T optical engines: simply multiplying the current pluggable module failure rate of about 300 FIT across a CPO system is insufficient, so the example sets a target of 10 FIT for the optical engine and 50 FIT for the 3.2T laser source. These are reliability design targets, not field data already achieved.[2]

The 36-channel detachable optical PIC connector announced by SENKO and Marvell in 2025 offers another approach: making the high-density optical interface detachable to support assembly, testing, and repair. The vendor emphasizes low coupling loss and repeatable alignment, but customers still need to verify it through insertion loss distribution, contamination, mating cycles, and field cleaning procedures.[4][6]

Connectors evolve from accessories into system functions

The OIF 3.2T protocol does not specify a single connector form; it only states that the system may use MPO-style connectors to an in-chassis adapter, or the implementer may choose other solutions. This "not fully specified" aspect is precisely the industrialization difficulty: when the interface is not defined down to the last connector, the system vendor's mechanical tolerances, panel density, cleaning, and maintenance strategy become differentiating capabilities.[1]

The framework document further points out that additional connections increase insertion loss; high-density MPO may not meet CPO space and density requirements; the front panel of a 51.2T switch may have 256–1,024 fibers, and the loss "long tail" of standard connectors is amplified with a large number of connections, so CPO systems often require lower-loss connectors.[2]

Competition among connectors is not only about "whether it can be plugged in," but about the following closed loop:

  1. Whether testing can be performed first at wafer level or optical engine level, avoiding bringing defective parts into a non-repairable package;
  2. Whether a detachable interface can maintain sub-micron-level repeatable alignment after multiple mating cycles;
  3. Whether end-face contamination can be inspected, cleaned, and incorporated into maintenance processes;
  4. Whether connectors, fiber-optic assemblies, and transceiver modules can jointly define the insertion loss and return loss budget;
  5. Whether a failure can be traced to the laser, connector, fiber assembly or optical engine, rather than scrapping the entire package.

The thermal and management issues behind 1,152 fibers

In its November 2025 CPO reference design, Marvell describes a 1U AI compute tray: up to four XPUs, 102.4 Tbps total, with an actual fiber count of 1,152. Each XPU connects to four 6.4T optical engines, and each optical engine connects to two 36-fiber detachable MPCs; the entire tray has 32 MPCs in total.[3]

The switch side of the same reference system uses 16 optical engines and 1,152 fibers, of which 128 are for laser feed and 1,024 are for data; 16 laser modules are moved to the panel side to improve maintainability and heat dissipation. The cold plate effective height is only 3.6 mm, thinner than traditional air-cooled solutions. The key here is not any single number itself, but that fiber management, laser maintenance, cold plate, connector cleaning, and cable bending are forced to be co-designed within the same space.[3]

System Trade-offsPublic EvidenceCannot be directly extrapolated
Density and serviceability32 MPCs, 1,152 fibersNot all CPO systems must achieve this density
Thermal and reliability3.2T@10 pJ/bit ≈ 32 W; liquid-cooling cold plate 3.6 mm effective heightNot measured field failure rates
Laser serviceability16 external laser modules placed on the panel sideDifferent architectures may use different redundancy and placement approaches
Link BudgetAdditional connections increase insertion loss, and the long tail is amplified in large-scale systemsSpecific values must be determined by product drawings and actual measurements

The OIF framework also calculates using 10 pJ/bit as an example: a 3.2T optical engine corresponds to about 32 W of electrical power. In its example environment, the room ambient temperature is 15–35°C, relative humidity up to 90%, and condensation prevention is required. This example cannot be directly taken as a thermal design specification, but it is sufficient to show that CPO thermal issues have expanded from "chip hot spots" to "cold plate, connectors, and fibers heated together."[2]

Procurement cannot just ask "how many T"

From the perspective of standards and system reference designs, CPO suppliers must include the following items in the technical confirmation sheet, rather than only quoting module rates:

Inquiry DimensionsRecommendations & RequirementsAcceptance Evidence
External laser architecture4 or 8 PM fibers, wavelength, power, redundancy, and replacement methodOptical power budget, fault injection, retesting after plugging and unplugging
Polarization-Maintaining FiberOperating/cutoff wavelength, beat length, extinction ratio, coil radius, lengthEnd-to-end testing according to the agreed length and connector
ConnectorsNumber of channels, keying, insertion loss, return loss, mating durability, cleaning methodBatch insertion loss distribution and end-face contamination inspection
Thermal DesignCold plate contact surface, fixture pressure, allowable case temperature, condensation boundarySystem-level thermal cycling and hotspot mapping
ReliabilityEOL definition, FIT target, redundancy strategy, failure localizationAccelerated life and failure analysis, without extrapolating beyond the sample range

Engineering implications for Toprays

CPO elevates "optical interconnect" from a device-level problem to a system-level problem involving fiber, connection, packaging, thermal, and testing. For supply chain participants such as Toprays in materials and fiber, the opportunity lies not only in entering the BOM of a particular CPO module, but in whether it can provide verifiable material and component interfaces:

  • Define the geometry, coating and polarization indicators of PM fiber according to wavelength, bend radius and temperature cycling;
  • Provide end-face, insertion loss, return loss, cleanliness, and mating durability data for fiber-optic assemblies;
  • Place quartz materials, fiber preforms, coatings and connector processes in the same batch traceability record;
  • Include system-level thermal cycling, condensation and vibration testing in acceptance, rather than only performing single-point measurements at room temperature;
  • Use the customer's optical power budget to derive the per-fiber power and return loss margin, rather than only comparing the module's Tbps.

This is also what distinguishes this article from “how much more rate a certain optical engine has achieved.” When CPO truly moves toward mass production, what customers buy is not an isolated optical engine but an optical link that can be assembled, tested, cleaned, maintained and traced for accountability.[5][9]

Editorial Judgment

CPO standardization work has already written external lasers, polarization-maintaining fiber, detachable connectors, and thermal management into the system boundary; vendor reference designs have pushed density and maintainability to the scale of a thousand fibers and dozens of connectors. For the supply chain, the next stage of differentiation will not be determined solely by silicon photonic chips, but by "whether materials and connections can remain verifiable under heat, contamination, bending, and batch variation."

What Toprays should establish in advance is a test data interface for fibers, quartz materials, and connector components, rather than chasing a connector name that is not yet fully standardized. What customers ultimately need is a reproducible end-to-end optical power budget, traceable materials, and an executable maintenance strategy.

REFERENCES

  1. Implementation Agreement for a 3.2Tb/s Co-Packaged (CPO) Module — OIF（standard，high）
  2. Co-Packaging Framework Document — OIF（standard，high）
  3. Co-packaged Optics: Powering the Next Wave of AI Data Center Innovation — Marvell（manufacturer，high）
  4. Marvell and SENKO Advance AI Infrastructure with Detachable 36-Channel Optical PIC Connector — SENKO（manufacturer，high）
  5. Progress in Research on Co-Packaged Optics — Micromachines 15(10), 1211（peer_reviewed，high）
  6. High-power tests of expanded beam connectors for co-packaged optics applications — SPIE Optical Interconnects XXIV（conference，high）
  7. Polarization Maintaining Fiber (Panda) — Heracle（manufacturer，high）
  8. ELSFP Connector — SENKO（manufacturer，high）
  9. Roadmapping the next generation of silicon photonics — Nature Communications 15（peer_reviewed，high）

This article distinguishes between OIF standards, vendor reference designs, and public papers: standards define interfaces and system boundaries; Marvell/SENKO figures belong to vendor reference designs; papers are used to illustrate packaging, coupling, and system bottlenecks. The 1,152 fibers, 32 connectors, and 3.6 mm cold plate are not universal specifications for all CPO systems. External materials are reference_only; actual products and acceptance are subject to customer drawings, sample testing, and final specifications. AI-generated illustrations only explain application contexts and do not represent real products.