Broadcom today announced the release of Tomahawk 6, the world’s first Ethernet switch with 102.4 Terabits per second of capacity on a single chip — delivering double the bandwidth of any solution on the market today.
Purpose-built for the AI era, Tomahawk 6 provides unmatched performance, efficiency, and flexibility to help data center operators and cloud providers future-proof their AI networks and optimize total cost of ownership.
Key benefits:
102.4 Tbps bandwidth in a single chip
Support for 100G and 200G PAM4 SerDes
1,024 SerDes lanes on a single device for copper-optimized deployments
AI-optimized routing, congestion control, and fault detection
Seamless scale-up to 512 XPUs; scale-out to 1M+ XPUs
CPO option for lowest power and latency in optical networks
Ethernet-based openness and flexibility across the stack
Customer Benefits: Designed to Meet Real-World AI Demands
As AI clusters scale from hundreds to over a million XPUs, network performance becomes a make-or-break factor. Tomahawk 6 directly addresses this challenge with a combination of ultra-high bandwidth, low latency, and AI-optimized routing, giving customers:
Future-ready scale: Support for more than 100,000 XPUs in two-tier networks and up to 1M XPUs in scale-out configurations.
Power efficiency at scale: Dramatic reduction in energy use and heat output, enabled by Broadcom’s advanced 200G and 100G SerDes and co-packaged optics options.
Deployment flexibility: Native support for a range of topologies (scale-up, Clos, torus, rail-only, etc.), making it ideal for any AI workload — from reinforcement learning to fine-tuning LLMs.
Open ecosystem: Freedom from vendor lock-in through support for the Scale-Up Ethernet (SUE) Framework, enabling seamless integration with any Ethernet-based XPU or NIC.
Strategic Differentiators: A Platform for End-to-End AI Networking
Broadcom’s Tomahawk 6 is not just a new chip — it’s the cornerstone of a comprehensive strategy to transform AI networking through open standards, full-stack integration, and scale-first design:
Unified scale-up and scale-out architecture: Unlike fragmented solutions, Tomahawk 6 enables hyperscalers to run both high-bandwidth server-to-server communication and broad cluster-scale networking over the same Ethernet fabric.
Cognitive Routing 2.0: Built-in telemetry, congestion-aware routing, and intelligent packet trimming enable adaptive, self-healing networks that automatically optimize performance.
Lowest TCO with passive copper and CPO: Customers can dramatically cut costs using 200G SerDes for extended passive copper reach or opt for co-packaged optics to minimize power consumption and improve reliability in dense AI clusters.
Built for the Future of AI Networking
Tomahawk 6 is launching into an ecosystem already mobilizing toward open, high-efficiency Ethernet solutions. It aligns with the Ultra Ethernet Consortium and integrates tightly with Broadcom’s full AI networking platform — including Thor NICs, Jericho switches, Agera retimers, and Sian DSPs — to provide customers a complete, interoperable solution.
