The escalating demand for bandwidth and energy efficiency in AI data centers is driving optical interconnect technology to a critical juncture. An industry report released on April 16th highlights that CPO (Co-Packaged Optics) switches have officially entered mass production, with commercial products from both Nvidia and Broadcom, while TSMC's COUPE silicon photonics platform serves as the foundational core architecture.
Why is CPO so significant? In simple terms, traditional optical modules are "plugged externally into the switch," forcing signals to take a long detour. CPO, however, integrates the optical engine directly next to the switch chip, shortening the signal path from tens of centimeters to just a few millimeters. Data cited in the report indicates this change drastically reduces insertion loss from 22dB to approximately 4dB, improves signal integrity by 63 times, boosts system optical power efficiency by up to 5 times, and increases network resilience by 10 times.
Nvidia has launched two CPO product lines. Its CPO switch chip for one line is in full mass production. As the world's first fully integrated 512-lane, 200G-capable CPO Ethernet switch system, it is integrated into a network rack using a 102.4 Tb/s switch chip. Energy efficiency is a core advantage. The traditional 1.6Tbps pluggable optical module consumes about 30W, with over half used by the DSP. Nvidia's CPO solution, by integrating silicon photonics directly into the switch package and eliminating the DSP, achieves up to a 5x improvement in system-level optical power efficiency and a 10x increase in network resilience, while also reducing the number of lasers required by approximately 4 times, lowering operational costs and network outage risks. The other Nvidia line is a liquid-cooled switch system featuring four ASICs, delivering a total full-duplex bandwidth of 115.2Tbps.
Broadcom was one of the earliest vendors to implement a CPO system. Its first-generation mass-produced CPO solution offers a total bandwidth of 51.2Tbps. The next-generation platform doubles the per-modulator rate to 200Gbps. The report states this platform reduces optical interconnect power consumption by about 70% compared to traditional pluggable modules, improving system energy efficiency by over 3.5 times. Initially, Broadcom used a specific packaging scheme but encountered limitations scaling single-lane speeds beyond 100Gbps, leading them to adopt a packaging architecture based on TSMC's COUPE technology to minimize signal conditioning needs and losses.
Why have both Nvidia and Broadcom chosen the same supplier's platform? The COUPE platform is built on 3D hybrid bonding technology, essentially stacking an electronic integrated circuit directly on top of a photonic integrated circuit. This is akin to two chips having a direct conversation versus shouting across a hallway, resulting in an extremely short path with very low loss. Data shows this bonding method increases interface density by at least 16 times, reduces parasitic capacitance by about 85%, and can achieve either a 40% power reduction or a 170% speed increase at the same power. Regarding system energy efficiency, traditional copper interconnects exceed 30 pJ/bit, while traditional pluggable modules are above 10 pJ/bit; fully integrated optical engines based on COUPE can reduce this to below 2 pJ/bit, with latency reduced by over 95%.
The report also notes that as CPO packaging evolves towards larger sizes and multi-engine integration, OSAT companies with system-level advanced packaging capabilities may participate in back-end integration. One company has demonstrated the ability to integrate multiple optical engines and ASICs in large packages, achieving low power consumption, suggesting they could handle COUPE-based CPO packaging.
CPO introduces challenges beyond the chip level, notably the routing of thousands of optical fibers. A single CPO switch can have over 1000 fibers, creating a significant organizational challenge and boosting demand for optical array components. Fiber Array Units are widely used to assist the crucial fiber coupling process in CPO. In CPO applications, the number of FAUs required per switch can be 3 to 5 times that of traditional schemes. Broadcom has integrated precision FAUs into its CPO platform. Key connector products include MPO connectors, which enable parallel multi-fiber transmission, significantly increasing wiring density, with some designs allowing twice the number of connectors in the same space. Another connector type is designed for direct fiber-to-chip coupling in CPO, minimizing height and insertion loss. Furthermore, an industry group is working on implementation agreements to promote an open CPO ecosystem.
In traditional systems, lasers undergoing repeated thermal cycles are a major cause of failure. CPO addresses this by relocating lasers to separate, temperature-controlled units. One Nvidia CPO system is equipped with external laser source modules, each containing multiple lasers. A key advantage is that the total number of lasers in the data center is reduced fourfold compared to traditional designs, and each laser module supports field replacement without affecting surrounding switch infrastructure, contributing to a 10x improvement in network resilience.
What is the cost of a CPO switch? A detailed breakdown was provided for a specific model. The optical engine is the largest cost component, with the total material cost for engines alone estimated between $35,000 and $40,000. A fiber management unit for organizing over 1000 fibers costs over $3,000. The total Bill of Materials cost, including fiber and other components, is approximately $70,640. Assuming a 60% gross margin, the estimated selling price is around $176,600. Adding a allocation for a three-year service and warranty brings the total price to approximately $204,856. The total power consumption is 3,548 watts. The report notes these cost estimates are based on current production scale and may improve with higher volumes.