As AI workloads scale beyond anything seen before, the demand for bandwidth, efficiency, and reliability inside GPU clusters is skyrocketing.
In this interview, Jim Carroll of Converge Digest speaks with Manish Mehta and Rajiv Pancholy of Broadcom about the company’s new Tomahawk 6 “Davisson” Ethernet Switch, the industry’s first 102.4 Tbps Co-Packaged Optics (CPO) platform. They discuss why link stability, thermals, and manufacturability now define the next frontier of data-center networking.
🕒 Timestamps 00:00 – Introduction Broadcom unveils the Tomahawk 6 “Davisson”, a 102.4 Tbps CPO Ethernet switch—the third generation of its co-packaged optics platform.
00:30 – The Need for CPO Explosive east-west AI traffic has pushed traditional pluggable optics to their limits. Broadcom began CPO development years before the AI surge but now finds it ideally suited for today’s cluster architectures.
01:00 – From Concept to Production Scale Initial designs targeted front-end cloud networks; the AI boom validated the CPO architecture. Discussion of design maturity, packaging yield, and link reliability.
02:30 – Power and Thermal Efficiency CPO enables dramatic power savings per bit versus pluggables—critical as rack densities climb. Thermal management innovations keep optics stable under GPU-class heat loads.
03:40 – Link Stability and Manufacturing Quality The Davisson generation delivers unprecedented link-error performance and long-term optical stability. Rajiv Pancholy explains Broadcom’s automated alignment and assembly process that ensures volume reliability.
05:10 – Ecosystem Integration and Standards How Broadcom collaborated with leading hyperscalers and optics partners to ensure interoperable electrical and optical interfaces—preparing for Ultra-Ethernet and 800G/1.6T ecosystems.
06:30 – Future Outlook Manish Mehta outlines the roadmap beyond 102.4 Tbps: scaling to even higher radix systems while maintaining power efficiency and manufacturability across data-center fabrics.
07:30 – Closing Thoughts Reflections on lessons from the CPO journey—yield learning curves, thermal co-design with switch ASICs, and what this means for the next generation of AI infrastructure.
