As AI servers scale from hundreds of GPUs to tens of thousands, computing power is no longer the only bottleneck. Large amounts of data must move between GPUs, switches, and racks; if each high-speed channel requires long circuit board traces, signal loss, power consumption, and heat dissipation will all increase.
CPO (Co-Packaged Optics) moves the optical engine closer to the switch ASIC, shortening the transmission distance of high-speed electrical signals, and then uses optical fiber to send the data out of the device. It's not simply about miniaturizing the optical module; it's about rearranging the chip, optical components, packaging, power supply, heat dissipation, and maintenance methods. This article will start with the signal path, explaining why CPO exists, how it differs from pluggable optical modules, and what engineering issues need to be addressed before mass production.
What is CPO? The core change lies in the signal path.
CPO is an architecture that integrates a switching ASIC with one or more optical engines on the same carrier board or compact package platform. The optical engine typically includes a modulator, photodetector, driver, transimpedance amplifier, and fiber optic coupling structure; the actual composition will vary depending on the product and packaging.
Traditional pluggable optical modules are located on the device panel. The high-speed electrical signals from the switch ASIC must first pass through the package contacts, printed circuit board, and connectors before reaching the optical module. With the increase in data rate, this copper trace requires more complex equalization, retiming, and signal compensation. CPO moves the photoelectric conversion to the vicinity of the ASIC, allowing the high-frequency electrical signals to travel only a short distance, while long-distance transmission is handled by optical fiber.
This approach is part of heterogeneous integration. For a better understanding of chip stacking, substrates, and interconnects, please refer to [link/reference needed].Structural differences in high-density IC packaging technology.
How does a CPO convert electrical signals into optical signals?
When a piece of data is sent out from an exchange, it generally goes through the following process:
- The switching ASIC outputs high-speed electrical signals through SerDes (Serializer/Deserializer).
- The nearby optical engine receives electrical signals and drives the modulator to change the intensity or phase of the light.
- The modulated light is coupled into an optical fiber and transmitted to another switch, server, or cabinet.
- The optical detector at the receiving end converts the optical signal back into an electrical signal, which is then processed by the ASIC.
The key is not to "replace all circuits with light," but to allow electrical signals to handle only short-distance interconnections within or near the package. Power supply, control, and some management signals still require circuitry, and the optical fibers must be precisely aligned and secured to maintain low coupling loss.
OIF released in 20233.2 Tb/s CPO Module Implementation ProtocolIt defines eight 400 Gb/s optical interface options, 32 CEI-112G-XSR host interfaces, as well as optical, electrical, and management interfaces. This means that the CPO is not just a concept demonstration, but that the industry is also establishing the boundary conditions required for multi-vendor interoperability.
What are the differences between CPO and pluggable optical modules?
| Comparison items | Pluggable optical module | CPO co-packaged optics |
|---|---|---|
| Photoelectric conversion position | Equipment front panel | Near the switch ASIC |
| High-speed telecommunications distance | It is quite long and needs to pass through the circuit board and connector. | Shorter, mainly near the packaging area |
| Repair methods | A single module can be directly replaced. | This may involve the repair of the light engine, packaging, or the entire machine. |
| Panel Configuration | Limited by module size and heat dissipation space | More panel space can be allocated to fiber optic connections. |
| Thermal design | The heat sources of the module and ASIC are relatively dispersed. | Optical components are placed close to the high-power ASIC, resulting in stronger thermal coupling. |
| Supply chain maturity | Mature interface, many suppliers | Packaging, fiber coupling, and testing are still being integrated. |
The main benefit of CPO (Consumer-on-Purpose) solutions is shortening high-speed electrical paths, while the advantages of pluggable modules are maintenance flexibility and a mature supply chain. Therefore, the two are not simply related as new and old generations. When device bandwidth has not yet approached the limits of circuit board traces and panel density, pluggable solutions usually still have cost and operational advantages.
Why is a CPO suitable for an AI data center?
AI training requires a large amount of GPUs to exchange data synchronously. The total bandwidth, per-port rate, and energy efficiency of the switches directly affect the cluster size. When the bandwidth increases, if the power consumption of the electrical channel compensation also increases, the entire data center will not only incur electricity costs, but the cooling system will also have to dissipate these heats.
This is also the scenario where CPO (Converter of Power) receives the most attention: it can reduce losses in high-speed electrical channels and panel density pressure, allowing switches to achieve higher total throughput. However, the energy savings depend on the ASIC, SerDes, optical engine, cooling, and network topology, and cannot be represented by a single number for all products.
For example, NVIDIA in 2025Spectrum-X Photonics Official AnnouncementThe company claims its silicon photonics switches are 3.5 times more energy efficient than traditional solutions and lists 100 Tb/s and 400 Tb/s switch configurations. This product data is specific to the manufacturer and system conditions, suitable for observing industry trends, and should not be directly applied to other CPO architectures.
When evaluating the effectiveness of a data center, it's also necessary to consider TDP, cooling energy consumption, and PUE. Further reading is available.Methods for determining the thermal requirements of TDP and PUEAvoid looking only at the nominal power consumption of the switch chip.
Co-packaging does not mean that the laser must be placed next to the chip.
Lasers are sensitive to temperature and reliability, and the switching ASIC is a high-heat flux area within the system. Therefore, some CPO architectures place the modulator and photodetector around the ASIC, while the laser source remains on the front panel, sending continuous light into the optical engine via optical fiber.
OIFELSFP External Laser Implementation ProtocolThe laser is placed on the cooler front panel and supports hot-swappable replacement. This reduces the risk of the laser being affected by the ASIC heat source and also prevents a failed light source from being scrapped along with the entire switch package.
However, external lasers increase the complexity of fiber optic routing, connectors, power distribution, and eye safety design. In other words, while CPO solves the problem of long-distance electrical interconnection, it shifts some of the complexity to optical integration and system management.
The four most difficult problems in CPO mass production
1. Heat dissipation and temperature distribution
With ASICs, drivers, and optical components placed closer together, localized hotspots can affect laser wavelength, modulator efficiency, and photodetector noise. Thermal design cannot solely focus on total power consumption; it must also consider the location of heat sources, package thermal resistance, cold plate contact surfaces, and coolant paths. If the system uses liquid cooling...Selection criteria for water-cooled plate flow channels and pressure dropIt will also affect maintainability and pump power consumption.
2. Integration of yield rate and known good products
A single CPO package may contain a high-cost ASIC, multiple optical engines, electronic chips, and a precision fiber optic interface. Failure of any component or coupling point can lower the overall package yield. Therefore, wafer-level testing, known-good die screening, and post-package testing must be planned together in the early stages of design.
3. Fiber Coupling and Assembly Tolerances
Alignment errors between optical fibers and silicon photonic components are typically reflected directly in insertion loss. Mass production lines must balance micron-level alignment, adhesive curing, thermal expansion, and long-term reliability; samples that work in the laboratory do not necessarily mean they can be mass-produced at a reasonable pace.
4. Maintenance and Fault Isolation
In the event of a pluggable module failure, field personnel can directly replace the individual module. If the CPO lacks a modular design, an optical engine failure may require replacing the switch tray or even the entire system. The monitoring interface must be able to distinguish between laser, optical path, optical engine, ASIC channel, and connector failures; otherwise, the power savings may be offset by downtime and spare parts costs.
Will CPO completely replace pluggable optical modules?
It's unlikely in the short term. CPO is best suited for switches with high port count, high total bandwidth, and power consumption and panel density that are close to their limits; for equipment with lower bandwidth, requiring frequent module upgrades, or emphasizing rapid on-site maintenance, pluggable optical modules are still reasonable.
In practice, several coexistence solutions may emerge: some high-speed ports may adopt CPO (Continuous Pluggable Optics), while lower-speed or long-distance interfaces may remain pluggable; linear pluggable optics (LPO) may also be used to reduce the power consumption of digital signal processing within the module. When selecting a device, the overall energy consumption, equipment cost, maintenance time, spare parts strategy, and supplier interoperability should be compared, rather than just comparing the power consumption of a single port.
Conclusion: First consider the system conditions, then determine whether a CPO is needed.
The essence of CPO (Converter-on-Poly) is to shorten the transmission of the most difficult high-speed electrical signals, allowing optical fibers to handle long-distance data transmission sooner. It can improve the upper limits of switch bandwidth density and energy efficiency, but at the same time, it concentrates heat dissipation, fiber coupling, packaging yield, and maintenance responsibilities into a more complex system.
When evaluating a CPO, you can ask four questions in sequence: Has the existing electrical channel become a bandwidth or power consumption bottleneck? Can the overall cooling system control the temperature of the optical components? In case of failure, to which layer can it be replaced? Does the supplier adhere to interoperable optical, electrical, and management interfaces? Only when these answers can be measured and maintained can the CPO be more than just a high-speed solution on the specification sheet, but a data center architecture that can be truly implemented.
