The wave of artificial intelligence (AI) has spread around the world, and the demand for data center computing has grown at an amazing speed. However, traditional input/output (I/O) has gradually become a huge bottleneck for performance improvement....
The wave of artificial intelligence (AI) has spread around the world, and the demand for data center computing has grown at an amazing speed. However, traditional input/output (I/O) has gradually become a huge bottleneck for performance improvement. The semiconductor manufacturer Broadcom has led a profound industrial revolution with its breakthrough "Co-Packaged Optics" (CPO).
Banish Mehta, vice president of marketing and operation of Broadcom Optical Systems, said that the booming development of AI applications, especially training and reasoning of large-scale language models, has put forward unprecedented requirements for the network architecture of the data center. Traditionally, the data center performs optical signal conversion and transmission based on the Pluggable Transceivers. However, with the significant growth of the exchanger frequency width, the power consumption, cost and component problems of pluggable modules are becoming increasingly prominent, becoming the bottleneck for contract data center expansion (Scale-out).
On the other hand, the "scale-up" requirement of AI computing, that is, the high-speed connection of thousands of GPUs or XPUs (special processors) is also facing challenges. If it is only based on copper cable transmission, its physical limit limits the interconnection range to a single cabinet, which cannot meet the long-distance and high-frequency width requirements of hundreds of endpoints of AI integrated connections in the future. Optics connectivity has become the only feasible solution, but how to realize it in a more energy-efficient way has become a difficult problem that the industry needs to solve urgently.
Manish Mehta pointed out that Broadcom has been forward-looking in the research and development of CPO technology many years ago. The core concept of CPO is to integrate the optical engine with the exchanger ASIC or processor chip into the same substrate or package, greatly shortening the transmission path of the electric signal. This innovative architecture brings revolutionary benefits. Compared with traditional pluggable modules, Broadcom's CPO technology benefits include extreme energy efficiency, ultra-high integration density, and significant cost reduction.
Broadcom has released the TH4-Humboldt CPO platform that uses 25.6T Tomahawk 4 exchanger chips in 2022, and successfully cooperated with Tencent. After obtaining verification in the real data center environment, it will launch the TH5-Bailly CPO platform based on 51.2T Tomahawk 5 exchanger chips in 2023. Today, the product has achieved the first large-scale mass production and large-scale shipment in the industry, supplying goods to many customers around the world. In May 2025, Broadcom announced the launch of a third-generation CPO product that supports single channel 200G transmission rate, continuing to consolidate its technological leadership. At the same time, the fourth generation of target single channel 400G is also being actively developed.
Manish Mehta emphasizes that Broadcom's success in the CPO field is based on its profound ability to integrate semiconductor and optical technology. First of all, in the core modulator technology, Broadcom adopts the mature and reliable Maher-Zendel modulator (MZM). Compared with other new technologies, MZM has a decades of verification foundation in quality, reliability and quantity, ensuring customers' confidence in deploying key tasks.
Secondly, Broadcom has realized the most dense optical engine in the industry through advanced heterogeneous integration and 3D stacking technology. For example, in the TH5-Bailly platform, chip stacking (Chip-stack) method has been adopted to vertically integrate photon overall circuits (PIC, 65 nanometers) at different process stages with electronic overall circuits (EIC, 7 nanometers). This is the key to achieving ultra-high density and performance. Also, in terms of heat dissipation management, Broadcom's innovative Edge Coupling design framework allows the heat dissipation device or liquid-cooling module to directly contact the optical engine. Whether traditional air-cooling or high-efficiency liquid-cooling can achieve optimal heat transmission efficiency and ensure that the system operates stably under high loading.
Manish Mehta further pointed out that the successful deployment of the CPO system is beyond the development of a mature and tightly collaborative ecological system. Broadcom has established stable cooperation relationships with many partners around the world, covering various links such as optical cables, optical connectors, remote laser light source modules, slots and system integration. Many of these key supply chain partners are located in Taiwan. Taiwan's leading position in the global electronic manufacturing and semiconductor packaging field makes it an indispensable core force in Broadcom's promotion of popularization of CPO technology. This close cooperation not only accelerates the development and deployment of products, but also highlights Taiwan's strategic importance in global AI hardware-based facilities.
Manish Mehta also pointed out that from the early market leader of VCSEL and EML laser components to the current pioneer of mass production in CPO technology, Broadcom has once again proved its far-reaching and strong execution ability in the field of optics interconnectedness. Faced with the strict requirements of AI data centers on efficiency, power consumption and cost, co-packaging optics (CPO) is no longer a remote scenario, but a technological revolution that is taking place.
With the successful large-scale deployment of the second-generation 51.2T CPO products and the steady advancement of third-generation and fourth-generation technologies, Broadcom is leading optics to enter a more efficient, denser and more cost-effective new era. These innovations not only solve the I/O bottleneck in AI computing, but also lay the foundation for the future development of heterogeneous integrated computing architecture, and continue to provide the strongest power for the evolution of global AI infrastructure.
