Broadcom announced that its Tomahawk 6 switch silicon is now shipping in production volume, marking the commercial availability of the industry’s first 102.4 Tbps Ethernet switch designed for AI infrastructure. The Tomahawk 6 family doubles the throughput of the previous Tomahawk 5 generation and targets large-scale AI training and inference networks that connect hundreds of thousands to millions of accelerators.
The chip is designed to support both scale-out Ethernet fabrics and scale-up accelerator interconnect topologies. Broadcom said Tomahawk 6 integrates advanced load balancing and congestion management features to maintain high network utilization and minimize job completion time across large AI clusters. The device supports both 100G and 200G SerDes interfaces and offers configurations with either 512 lanes of 200G or 1,024 lanes of 100G connectivity on a single chip.
Broadcom said the platform also focuses on system-level efficiency by extending passive copper reach and reducing the need for optical modules. The architecture aims to reduce the number of switch tiers and simplify AI cluster network design while lowering latency and total cost of ownership. The company noted that Tomahawk 6 moved from initial sampling to production deployment in less than three quarters.
- 102.4 Tbps switching capacity
- Up to 512 × 200G SerDes or 1,024 × 100G SerDes per chip
- Designed for both AI scale-out Ethernet fabrics and scale-up accelerator networks
- Enables up to 128K-XPU clusters using two switch tiers
- Single switch can support up to 512 XPUs in scale-up topologies
- Advanced load balancing and congestion management for large AI workloads
- Optimized copper reach to reduce optical module requirements
“With Tomahawk 6 now shipping, Broadcom is translating its roadmap into real-world deployment,” said Sameh Boujelbene, vice president and analyst at Dell’Oro Group.
🌐 Analysis
The introduction of 102.4 Tbps switching silicon marks a major milestone in the evolution of AI data center networking. As AI clusters expand toward hundreds of thousands or even millions of accelerators, network fabrics must deliver extremely high bandwidth with low latency and efficient congestion management. Chips such as Tomahawk 6 represent the next generation of Ethernet switching silicon capable of supporting these massive distributed training environments.
Broadcom remains a dominant supplier of hyperscale switching silicon, competing with emerging alternatives such as NVIDIA’s Spectrum-X Ethernet platform and custom networking solutions deployed by hyperscalers. The rapid transition from sampling to production deployment for Tomahawk 6 reflects the accelerating demand for higher-capacity switching platforms as AI infrastructure operators design increasingly large accelerator clusters.
