Co-Packaged Optics: Unlocking Data Center Performance

Hyperscale data centers are confronting a performance wall, where the traditional chip-to-port connection imposes structural limits on throughput and scalability. This challenge is amplified by projections that global demand for data center capacity will more than triple by 2030.

The industry's response is co-packaged optics (CPO), a new architecture that integrates the optical input/output (I/O) directly with the chip to resolve the distance problem. This direct approach introduces a new set of engineering obstacles centered on thermal management and serviceability. The viability of a new architecture now depends on a system-level approach capable of mastering these core issues.

The established architecture of front-panel pluggable optics has reached its physical limits under the intense processing demands of AI. The inherent challenge of a traditional optical transceiver design is the significant electrical distance signals must travel from the Application-Specific Integrated Circuit (ASIC) on the main board, across lossy copper traces, to the front-panel module. At next-generation speeds like 448Gbps, propelling signals across this path requires immense electrical power to compensate for signal degradation over distance—energy that is ultimately dissipated as waste heat.

Intense power consumption is the source of the “I/O power wall,” a major industry problem where the energy required to move data begins to rival the energy used to process it. This drives up operational costs and complicates thermal management. In parallel, the physical size of pluggable modules imposes a hard limit on front-panel I/O density. As data rates increase, physical space is running out on the front panel to add more ports, creating a separate but equally critical bandwidth bottleneck at the faceplate.

CPO redefines the architecture of the optical transceiver by integrating the optical engine onto the same substrate as the main processor. This close integration substantially shortens the electrical path, reducing the power required to propel the signal. A shorter path eliminates the need for the high-power retimers and other signal-conditioning components required for long-reach copper traces, dramatically reducing the power consumed by the optical I/O. As a direct result, CPO confronts both the I/O power wall and the front-panel density limit, yielding major improvements in power efficiency and bandwidth.

The integration itself, however, introduces two primary engineering obstacles: 

• Thermal management, caused by placing the laser directly next to a high-value processor, the primary heat source in an optical system. 
• Serviceability risk, due to component failure, which may require replacing the entire processor package.

The feasibility of CPO depends on silicon photonics, a technology that integrates optical functions onto a chip using established semiconductor manufacturing. The central component of this architecture is the modulator, which encodes electrical data onto light, presenting a primary engineering trade-off.

The trade-off is evident between the two primary modulator technologies: Mach-Zehnder Modulators (MZM) and Micro-Ring Modulators (MRM). 

• Mach-Zehnder Modulators are highly stable and reliable, but their large physical footprint conflicts with CPO's density goals. 
• Micro-Ring Modulators are compact and power-efficient, but highly sensitive to thermal fluctuations, which are unavoidable when placed next to a heat-generating processor. 

This is a real-world decision that divides major industry players. Broadcom, for example, appears to favor MZM, while Nvidia has chosen MRM for its CPO applications.

The conflict between these two approaches underscores a crucial insight: thermal management is the central engineering obstacle that must be overcome in any viable CPO design.

The most effective architectural solution to CPO’s core thermal and serviceability challenges is to decouple the laser with an External Laser Source (ELS). Molex has realized this concept in a complete, first-to-market system that moves CPO from a theoretical diagram to a practical, deployable reality.

Solving the Thermal Challenge 
The External Laser Source Interconnect System (ELSIS) is a complete solution that physically separates the high-power laser from the processor and optical engine. It directly addresses the thermal challenge by housing the lasers in a separate, pluggable module. The design moves the most intense heat load away from the high-value processor and simplifies thermal management.

Mitigating the Serviceability Risk 
In addition to solving the thermal problem, the modular design mitigates serviceability risk by enabling simple, field-serviceable maintenance of the laser module. ELSIS uses a blind-mating interface where all optical and electrical connections occur inside the equipment, avoiding the costly scenario where a laser failure requires replacing the entire processor package.

A Complete System-Level Solution 
The result of this holistic approach is the final product: ELSIS is a complete, pre-engineered system comprising: 

• Electrical and optical connectors
• A press-fit cage 
• The pluggable module itself

All components are designed to function as a cohesive unit, enabling scalable, reliable deployment.

The transition to CPO represents a significant evolution in design philosophy, where thermal management and serviceability challenges are systemic. Unlocking CPO’s full performance and efficiency requires treating the laser, optics and interconnects as a single, integrated system.

Successfully implementing this architecture demands an engineering collaborator with deep expertise across optics, electrical and mechanical domains, since failure in any one domain compromises the entire system.

Molex ELSIS embodies this system-level approach, mitigating risks and providing the scalable architecture essential for the next generation of data centers.

Explore Molex ELSIS, the industry’s first complete interconnect system for co-packaged optics.