Broadcom's recently launched Tomahawk 6-Davisson (TH6 Davisson) co packaged optical (CPO) Ethernet switching chip redefines the performance standards of data center networks with a switching capacity of 102.4 Tb/s. As the industry's first CPO solution to achieve this bandwidth, it integrates 16 6.4 Tb/s Davisson DR optical engines and relies on TSMC COUPE (Compact Universal Photon Engine) technology and 3-nanometer process technology to significantly optimize the scalability and link stability of large-scale AI clusters while improving port density and energy efficiency.
1、 Technical difficulties and innovative solutions
Traditional pluggable optical modules face three major challenges in dealing with high-speed signal transmission: high power consumption, poor signal integrity, and maintenance complexity.
Power consumption and heat dissipation bottleneck: In the pluggable solution, signals need to be transmitted over long distances to the front panel, resulting in electrical losses of up to about 30W per port. TH6 Davisson integrates the optical engine and switch chip heterogeneously, embedding the optical conversion unit into the package to shorten the electrical signal path, reduce power consumption by about 70% (up to 9W per port), and reduce dependence on complex heat dissipation systems.
Signal attenuation and jitter control: The electrical loss of the 200Gbps channel reaches 22dB and requires high-power DSP compensation. Broadcom utilizes the COUPE silicon optical platform and substrate level multi chip packaging to reduce losses to 4dB, while improving AI training stability through link jitter optimization (achieving 1 million hours of "jitter free" operation in Meta testing).
Maintainability challenge: The laser in CPO design is vulnerable and affects its lifecycle. TH6 Davisson adopts a field replaceable ELSEP laser module, with the light source set as a field replaceable unit (FRU), balancing integration advantages and serviceability.
2、 Product Core Features
TH6 Davisson achieves multidimensional breakthroughs in performance, integration, and compatibility:
Bandwidth and Port Density: Single chip capacity of 102.4 Tb/s, supporting 64 800G ports or equivalent configurations, with a channel rate of 200Gbps per channel.
Energy efficiency and thermal management: reduces power consumption by 70% compared to pluggable optical modules, and is compatible with high-density racks through liquid cooling design.
● Scalability: Supports full interconnection of 512 XPUs in a single layer, expands to over 100000 XPUs in a dual layer network, and reduces congestion through the Cognitive Routing 2.0 intelligent traffic engine.
Compatibility and Standardization: Compliant with IEEE 802.3 specifications, supports interoperability with 400G/800G devices, and adapts to the NPO/CPO ecosystem.
3、 Comparison of Competitors and Market Positioning
Competitive analysis:
● Performance Leadership: TH6 Davisson's 102.4 Tb/s capacity and 200Gbps channel rate surpass Intel Bailly's (51.2 Tb/s) across generations and land earlier than Nvidia Spectrum-X (115.2 Tb/s, launched in 2026).
Energy efficiency and reliability: Its CPO integration brings 70% power optimization and Meta verified link stability, which is significantly better than pluggable solutions (such as New H3C S9827).
Ecological strategy: Broadcom adheres to open Ethernet standards, while Nvidia relies on software and hardware binding, and TH6 Davisson provides customers with heterogeneous interoperability flexibility.
4、 Actual application scenarios
TH6 Davisson is designed for high computing power demand scenarios, with specific applications including:
AI Training Cluster: In a dual layer network with over 100000 XPUs, low jitter and high bandwidth are used to ensure long-term training tasks and improve GPU utilization. For example, in Meta AI data centers, CPO technology has increased training efficiency by 90%.
Horizontal Expansion of Cloud Data Centers: Supports multiple topologies such as Clos and Torus, reduces multi hop latency through Cognitive Routing 2.0, and optimizes east-west traffic.
High Performance Computing (HPC): Combining liquid cooling and heat dissipation to achieve Pbps level data exchange in supercomputing environments, suitable for intensive computing tasks such as climate simulation and genomics.
5、 Product supply situation
Broadcom has provided early customers with samples of TH6 Davisson BCM78919 and plans to fully launch them in the future. Hardware manufacturer Edgecore Networks has taken the lead in launching switches based on this chip (AS1817-64O/128O), confirming the feasibility of mass production. In addition, the fourth generation CPO platform of Broadcom (400Gbps per channel) is currently under development, continuing the technological iteration.
Conclusion
Tomahawk 6-Davisson from Broadcom has solved the power consumption, signal, and scalability bottlenecks in high-speed networks through CPO integration, silicon optical optimization, and maintainable design. Its 102.4 Tb/s capacity sets a new benchmark for AI and cloud data centers, driving Ethernet towards a more efficient and open future.
