An Introduction to Co-Packaged Optics (CPO)
AI is red-hot right now — especially after the man in the leather jacket visited Taiwan in June 2023 and set off a whirlwind that drove growth and demand across the entire optical communications industry. Co-packaged optics (CPO) has become an especially hot topic, so in this post I'll give a brief introduction to CPO and where it stands today.
1. Why we need CPO
Rising data requirements, together with the new AI wave, are driving demand across optical communications: higher bandwidth and lower latency.
Optical communications data rates double every two years. Over the past 12 years, total bandwidth has risen 12x, but power consumption has risen 22x — and 26x for the optical module portion. Excessive power consumption requires more electricity for cooling, raising overall data center costs, and the heat it generates is also very unfriendly to active optoelectronic components (especially lasers).
Electrical signal loss from the traditional ASIC to the optical module is too high. In the existing architecture, the bandwidth limit and signal loss come mainly from the path between the ASIC and the optical module. At high bandwidths, the loss of every electrical trace and every contact has an enormous impact on overall signal quality. The higher the bandwidth, the greater the electrical loss, and both the ASIC and the optical-electrical conversion module need more powerful ICs and power supplies to drive and compensate for that loss.
CPO tackles the problem at the architecture level: electrical traces must be drastically reduced to have a chance of increasing data throughput further. Put another way, light has to get closer to the ASIC. The figure on the lower left shows the roadmap from copper in the past to 3D-packaged CPO in the future. Traditional copper relied entirely on electrical signaling; optical communications then replaced those modules with optical modules, bringing light to the switch port. Next came the concept of on-board optics and then CPO, where the optical signal moves directly inside the switch; with 3D CPO, the goal is for the ASIC's electrical signal to be converted to an optical output right inside the chip.
2. The fundamental technology behind CPO
The most important concepts of CPO are:
Optical-electrical conversion must be miniaturized so it fits inside the switch and increases density
Optical-electrical conversion should be as close to the ASIC as possible to reduce electrical signal loss
This concept gives silicon photonics a much better chance to be realized. For the past 20 years, silicon photonics has been regarded as the only way to break through Moore's Law, and research institutes and companies poured substantial resources into its R&D. But productization made little real progress until a few years ago, when Intel applied the technology to optical transceivers; more companies then joined in, and standardization gradually began.
Simply put, silicon photonics uses silicon processes to integrate the active and passive components an optical signal needs into a silicon chip, using semiconductor processes to integrate all the components. This approach can shorten the distance between the optical signal and the IC and raise transmission rates, and the integration also lowers power consumption. Manufacturing can be done on today's very mature semiconductor processes.
The process is essentially the same as an IC process, so it has a cost advantage at high volume. Its high level of integration should, in theory, also deliver good reliability.
In recent years, more and more companies have entered silicon photonics, and the whole supply chain — design, chips, epitaxy, manufacturing, packaging and systems — has steadily matured. That supply-chain maturity is a key reason the CPO concept could be put forward.
3. CPO: current status and challenges
Laser integration
The most important key to realizing CPO with silicon photonics is how to integrate the laser. Why is laser integration so critical? The laser is the source of the optical signal and is considered the single biggest factor in an optical transceiver's quality. But silicon itself doesn't emit light; it relies on III-V materials as the light emitter, and silicon and III-V processes are not compatible.
Intel overcame the incompatibility between silicon and III-V processes and used heterogeneous integration to realize a fully integrated chip. Of course, full integration isn't the only solution — I'll cover silicon photonics in more technical detail in a separate post.
Beyond successfully applying silicon photonics to traditional optical modules, Intel also used it to build the first 12.8T switch, composed mainly of modules containing 16 lasers at 106 Gbps each. In the figure below, you can see that the switch's external connectors have been replaced with optical connectors instead of traditional optical modules; optical connectors are smaller and make more efficient use of the switch's space.
The routing inside the switch is no longer the PCB itself but optical fiber; over such short distances, fiber loss is very small, and the fiber can be bent for routing as needed. With this technology, all the critical optical-electrical conversion is concentrated in the center of the switch. Potential drawbacks are repair after a laser failure and excessive heat concentration affecting laser performance.
Another approach, the external laser source (ELS), has been a hot topic in recent years; companies such as Rockley, Cisco and Broadcom have adopted it.
The idea is to pull the laser source out on its own: it can be built as a standalone laser module that plugs into the outside of the switch like an optical module. The laser light is routed through fiber into the central chip for signal processing, and the processed signal is then sent out through fiber.
The advantages of this architecture are:
1. External laser source modules can carry over the technology and supply chain of existing optical modules.
2. If a laser source fails, its module can be swapped out on its own for repair.
3. It avoids silicon photonics' biggest challenge — the integrated laser — letting the silicon photonics chip use a fully standard silicon process, which opens up more foundry options.
4. Lasers are most sensitive to heat, and this allows separate thermal management.
5. The downside is that space must be given up for the laser modules, and an extra fiber hop is required.
At OFC 2023, Cisco showcased a 51.2T switch; adopting a CPO approach could cut power consumption by about 25–30%.
Comparison of CPO laser integration approaches
Fully integrated laser
Characteristics: fully integrated chip solution
Pros: full integration gives an absolute cost advantage and lower optical loss
Cons: because everything is integrated, all heat sources are concentrated near the ASIC; in case of later repair or failure, parts can't simply be swapped as with today's modules.
Main proponent: Intel
External laser source (ELS)
Characteristics: modularizes the laser — the most critical and most heat-sensitive component — on its own
Pros: easy to repair, lower heat density.
Cons: higher optical loss; space must be given up for the laser modules
Main proponents: Cisco, Broadcom, Furukawa
Supply chain changes
Beyond the laser, I believe the other big challenge is the change in the supply chain. CPO technology has been discussed for many years, and its supply chain differs from that of traditional optical modules. In the traditional supply chain, users take switches from ODMs/OEMs, then source optical modules and fiber from module and fiber vendors to build a data center. But CPO depends on the silicon photonics supply chain, and I think the role of OEMs/ODMs will become much more important: they will need to package and integrate the PIC and EIC, may need to perform optoelectronic testing on semi-finished products, and may even have to enter the silicon photonics ecosystem. Amid this shift, I've also observed major system vendors gradually poaching talent from optical module makers for CPO R&D.
Comments