Broadcom has begun shipping the Tomahawk 6, a 102.4-Tbps ethernet switch chip designed to support massive artificial intelligence (AI) workloads.
With twice the bandwidth of current solutions, the chip enables large-scale AI clusters, supporting over a million auxiliary processing units (XPUs) in scale-out architectures.
It integrates 100 Gbps and 200 Gbps SerDes, co-packaged optics, and updated routing software for real-time congestion control. Tomahawk 6 enhances power efficiency with long-reach passive copper support and reduces latency through advanced telemetry and adaptive routing.
Compatible with standard ethernet networks and multiple topologies, it simplifies scaling without proprietary systems.
The chip also aligns with open standards, including the Ultra Ethernet Consortium and the new Scale Up Ethernet (SUE) Framework.
By supporting high-density interconnects and enabling flexibility in cluster design, Tomahawk 6 will meet the growing demands of generative AI (GenAI) training, inference, and model fine-tuning. Several large-scale deployments are already planned, some involving over 100,000 AI accelerators in production environments.
