Broadcom has officially launched and begun shipping the Tomahawk 6 switch, the first Ethernet silicon to deliver 102.4 Tbps of bandwidth in a single chip—doubling the capacity of previous-generation devices. Aimed squarely at addressing the networking bottlenecks in massive AI clusters, Tomahawk 6 supports both scale-up and scale-out architectures, integrating 100G/200G SerDes, co-packaged optics (CPO), and Cognitive Routing 2.0 to handle the demands of clusters scaling to over one million XPUs. The chip is now in volume production, with broad industry support from network equipment vendors and system integrators.
With industry-leading 200G SerDes and support for long-reach passive copper, Tomahawk 6 enables high-efficiency system designs with lower power and latency. The switch also offers a groundbreaking configuration with 1,024x100G PAM4 lanes, enabling dense, cost-effective deployment of Ethernet-based GPU interconnects. Optional CPO configurations further reduce link flaps and enhance reliability—key requirements for hyperscalers deploying large, multi-rack GPU clusters.
At the network layer, Cognitive Routing 2.0 introduces real-time telemetry, adaptive flow control, and sub-500ns failover. These features ensure traffic is dynamically balanced across paths in large-scale AI workloads, including mixture-of-experts and reinforcement learning models. Tomahawk 6 supports both standard scale-out topologies like Clos, as well as scale-up networks using Broadcom’s Scale-Up Ethernet (SUE) framework. The chip is also Ultra Ethernet Consortium compliant, positioning Ethernet as a unified fabric for backend AI interconnect.
Industry adoption is already underway, with multiple partners integrating Tomahawk 6 into next-generation switching platforms. These deployments are expected to support clusters with 100,000+ XPUs in two-tier architectures, while offering flexibility for varied AI training and inference environments. The switch is part of Broadcom’s end-to-end AI networking portfolio, which includes Thor NICs, Agera retimers, and Sian optical DSPs.
• 102.4 Tbps of Ethernet switching bandwidth in a single chip
• 512-XPU scale-up clusters supported at 200G/link
• Two-tier scale-out support for 100,000+ XPUs
• 200G or 100G PAM4 SerDes with long-reach passive copper support
• Optional co-packaged optics (CPO) for reduced power and higher reliability
• Integrated Cognitive Routing 2.0 for real-time congestion management
• Supports arbitrary topologies including Clos, scale-up, torus, and rail-only
• Ultra Ethernet Consortium compliant
• Supports Broadcom’s Scale-Up Ethernet (SUE) framework for XPU-to-XPU networking
• Available now with support from leading networking ecosystem partners
“Tomahawk 6 is not just an upgrade – it’s a breakthrough,” said Ram Velaga, senior vice president and general manager of Broadcom’s Core Switching Group. “It marks a turning point in AI infrastructure design, combining the highest bandwidth, power efficiency, and adaptive routing features for scale-up and scale-out networks into one platform. Demand from customers and partners has been unprecedented.”
“Arista is proud of its 15-year partnership with Broadcom and collaboration on the latest Tomahawk 6 silicon combined with feature-rich Arista EOS,” said Hardev Singh, General Manager, Cloud Titans and AI, Arista Networks. “We are excited to leverage its low power, high radix, 1.6Tbps port speeds, and advanced packet processing capabilities for AI-aware routing, based on open Ethernet standards.”
“For organizations building next-gen AI infrastructure, combining the AMD Infinity Fabric and Broadcom’s Tomahawk 6 enables massive-scale, low-latency GPU clusters,” said Forrest Norrod, Executive Vice President and General Manager, Data Center Solutions Group, AMD. “This collaboration empowers hyperscalers to build efficient, open ecosystems for AI training and inference.”
