The convergence of electrical and optical technologies in data centers is accelerating, bringing the commercialization of photonics-electronics convergence and Co-Packaged Optics (CPO), which could overcome the limits of AI infrastructure, within reach. Japanese companies are said to hold critical technologies such as InP lasers and ultra-precision assembly.
This article explains the physical limits facing data center infrastructure as generative AI becomes more widespread. Even as switch ASICs achieve higher data-processing capabilities, communications between chips and constraints on power consumption and heat generation are intensifying. It identifies photonics-electronics convergence and CPO as the leading solutions attracting attention for addressing these challenges.
Photonics-electronics convergence is the concept of bringing the boundary between computing and long-distance transmission closer together, while CPO refers to an implementation format in which optical engines are directly mounted on switches or GPUs. Compared with conventional pluggable and near-packaged solutions, CPO is expected to shorten electrical interconnects and reduce power consumption, although technical challenges including thermal design and alignment have also been identified.
In the current technology comparison, implementation is advancing in stages, from pluggable optical transceivers to NPO and then CPO. CPO in particular could substantially reduce communications power consumption, but it also presents challenges involving yield in the event of failures and thermal expansion.
Globally, NVIDIA and TSMC have outlined commercialization roadmaps, and CPO is expected to become the industry's main battleground. Japanese companies are said to control critical technologies in the physical layer and materials, including InP lasers, ultra-high-count cabling and ultra-high-precision molding, creating an opportunity to target a new revenue layer in AI infrastructure.