• Broadcom is now shipping Tomahawk 6, a 102.4 Tbps Ethernet switch chip.
• It supports AI clusters of up to a million XPUs with tools to manage bandwidth and congestion.

Broadcom is now shipping its Tomahawk 6 switch chip, offering 102.4 terabits per second of bandwidth on a single chip. That’s double the capacity of any current Ethernet switch and is aimed squarely at powering larger, more complex AI networks.

Built to handle both scale-up and scale-out network designs, Tomahawk 6 supports 100G and 200G SerDes and co-packaged optics, giving cloud providers and hyperscalers flexible options when connecting clusters of over a million processing units. It also introduces new routing features that help networks respond to congestion and failure in real time—critical for AI training and inference tasks.

Broadcom says Tomahawk 6 was designed for networks handling tens of thousands to over a million AI accelerators. With that kind of scale, the network can become a bottleneck. This chip aims to clear that hurdle by boosting throughput and reducing latency, while still relying on open, standards-based Ethernet.

Power efficiency and connectivity choices

Tomahawk 6 uses Broadcom’s high-performance SerDes to support longer copper cable runs without signal boosters, helping lower both power use and system cost. Customers also have the option to choose a version of the chip with 1,024 100G SerDes lanes on a single die—useful for dense clusters that rely heavily on direct copper connections.

For those that need optical links, Tomahawk 6 includes a version with co-packaged optics. This setup is designed to cut power use and latency further while improving long-term connection reliability—something data centre operators running hyperscale AI workloads increasingly demand.

Built for large AI workloads

The chip comes with updated routing software, dubbed Cognitive Routing 2.0, which includes telemetry and adaptive congestion control. These tools are designed to manage the unique traffic patterns of AI workloads like fine-tuning, reinforcement learning, and large model reasoning.

Training large language models requires splitting tasks across many graphics processors, with each handling a portion of the overall work. This parallel setup helps speed things up, but it also depends on the chips staying in sync by sending data back and forth. That constant communication uses a lot of bandwidth. Inference can be just as demanding, since it often involves pulling data from remote storage over the network. As these workloads grow, the pressure on the network increases.

Tomahawk 6 can support up to 512 XPUs in a scale-up setup and more than 100,000 XPUs in a two-tier scale-out network using 200 Gbps links. It delivers 102.4 terabits per second of Ethernet switching on a single chip. The chip includes 100G and 200G PAM4 SerDes that support long-range passive copper connections and optional co-packaged optics for systems that require optical links. It works with any Ethernet-based network interface card or processor and supports a range of topologies, including Clos, rail-only, rail-optimised, torus, and traditional scale-up architectures. These capabilities make it suitable for large-scale AI training and inference networks that demand both high bandwidth and flexibility.

Using Ethernet as the foundation for both connection types means organisations can stick with familiar tools and processes, even as they adjust cluster layouts to handle different workloads. That flexibility helps keep operations simpler and more efficient.

Broadcom says multiple large-scale deployments are already planned, some involving over 100,000 AI accelerators using Tomahawk 6 to manage traffic within and between racks.

Open standards and interoperability

As part of the rollout, Broadcom introduced the Scale Up Ethernet (SUE) Framework, which outlines how chipmakers and system vendors can build open, scalable connections between processors and network interface cards. The framework, announced at the Open Compute Project event in Dublin this April, is publicly available and will be shared with industry standards groups.

Tomahawk 6 also complies with the Ultra Ethernet Consortium’s specifications, supporting newer transport protocols and signalling features for large-scale training systems. It works across a wide range of network designs, including scale-up clusters, Clos topologies, rail-based systems, and torus-style layouts.