OCP Global Summit 2025 | Broadcom | Scale-Out Networks and Scale-Up Architectures with CPO
Introduction
At OCP Global Summit 2025, Broadcom shared its latest architectural view of data center networking in the AI era, in a talk titled "Scale-Out Networks and Scale-Up Architectures with CPO". Broadcom stressed that as GPU clusters grow rapidly, electrical interconnect alone can no longer meet the demands of training efficiency and power, and that CPO (Co-Packaged Optics) and the Optical Backplane will become key to next-generation AI infrastructure.
Content
1. From Scale-Up to Scale-Out: The Limits of Electrical Interconnect
Broadcom divides data center networking into two directions:
Scale-Up: scaling within a single rack or row, focused on high-bandwidth, low-latency short-reach connections.
Scale-Out: scaling across many racks, used to connect large clusters of tens of thousands of GPUs.
NVIDIA has used NVLink and NVL72 to build an electrical backplane, bringing "scale-up" into the rack. But Broadcom pointed out that as systems grow further to thousands or tens of thousands of GPUs, electrical signal loss and trace length become the bottleneck, and the industry must move to an optically interconnected backplane.
2. CPO and Optical Backplane: Toward an Optical-Domain Architecture
With a CPO + Optical Backplane architecture, Broadcom extends scale-up to an entire row (in-row) or even a whole data hall. By packaging the optics and the switch ASIC on the same substrate, it achieves:
Lower power (5.5 W per port)
Higher reliability (MTBF improved by more than 5x)
Zero connector loss and higher-density layouts
Meta's measured data show that at 100G/lane, a CPO system consumes 65% less power than conventional pluggable optics and 35% less than LPO, with zero link flaps over 1 million hours of operation testing.
3. Overall Efficiency: From Hardware to Training Performance
Broadcom noted that CPO's value lies not only in lower power but in "higher training efficiency".
On a cluster of 24,000 GPUs, CPO delivers a 90% improvement in training efficiency compared with pluggable modules (Pluggables).
The reason is that when a conventional link flaps, the training job has to roll back to the last checkpoint; CPO's stability greatly reduces such interruptions, allowing the cluster to run continuously for long periods.
4. Broadcom's Product Generations
Broadcom showed progress across three generations of its CPO family:
Gen1 Humboldt (Tomahawk 4): small-scale deployment with Tencent, mainly to validate packaging and serviceability.
Gen2 Bailey (Tomahawk 5): tested with Meta, further improving stability and maintainability.
Gen3 Davidson (Tomahawk 6): supports 200G/lane and 100T switching capacity, with lower power and more stable optical links, and uses TSMC Co-Packaged Process (COUPE) to integrate EIC and SiPh PIC into a production process for volume manufacturing.
Broadcom also said it is driving ELSFP standardized laser modules, so that CPO is no longer limited to custom packaging and can be built on existing module production lines, further advancing the industry's volume-production maturity.
5. Manufacturing and Volume-Production Challenges
Asked from the floor whether "CPO manufacturing is mature enough", Broadcom said some manual assembly steps remain, but it is working with TSMC and several packaging houses to expand capacity.
In the near term (the next two years), scale-out remains the main application; moving further into the scale-up tier will require more reliability data and long-term operational validation.
6. Open Ecosystem and Industry Collaboration
Broadcom emphasized that the CPO ecosystem must be open, not closed. It is working with companies including NVIDIA, Meta, Cisco and Marvell to drive standardization, and also supports "hybrid packaging architectures" - for example, half the ports using CPO and half still using electrical connections - to meet different deployment needs.
Conclusion
Broadcom's talk at OCP 2025 reiterated that CPO is no longer just a technology concept but a working architecture that can significantly improve cluster performance.
Through collaboration with ecosystem partners and maturing packaging technology, Broadcom aims to make CPO the key bridge between "scale-up and scale-out", while continuing to push standardization and process automation, ultimately making optical interconnect the mainstream infrastructure of AI data centers.
Further Perspectives
Technology impact
Broadcom's Tomahawk + CPO platform has become the de facto experimental standard for hyperscaler architectures. This "optical packaging" model will further push electro-optical co-design (EIC + PIC Co-Packaging) into the mainstream, and forces optical module makers to upgrade their process and test capabilities.
Supply chain observations
With Broadcom handing PIC packaging to TSMC, opto-electronic integration is shifting from traditional module assembly to wafer-level manufacturing, and the SiPh foundry role is formally being pulled into the mainstream optical communications supply chain.
The push for the ELSFP standard may draw vendors such as Lumentum, Innolight and Accelink into a shared manufacturing ecosystem, reducing single-source risk.
Market trends
Broadcom's choice of "90% training efficiency improvement" as its core selling point shows the market's focus shifting from power to overall AI compute productivity (Cluster Productivity).
As CPO transitions from 100G/lane to 200G/lane and pairs with optical backplane architectures, AI data centers will enter an era of megawatt-class single racks and optical domains spanning entire data halls.
Comments