Broadcom Inc. has claimed the industry’s first 102.4-Tbits/s Ethernet switch chip, shipping the Tomahawk 6 series with double the bandwidth of Ethernet switches currently on the market. The AI-optimized Tomahawk 6 targets scale-up and scale-out AI networks with support for 100G and 200G SerDes and co-packaged optics (CPO).

The new Ethernet switch, with scale-out and scale-up networking support, offers a comprehensive set of AI routing features and interconnect options, designed to meet the requirements of emerging 100,000 to 1 million AI accelerator (XPU) clusters.

“We’re giving customers the option for 200G SerDes or 100G SerDes, depending on the speed of their GPUs, depending on the optics that they want to leverage for their networks, and depending on the copper reach they want,” Pete Del Vecchio, product manager for the Tomahawk switch family at Broadcom, said during a virtual event.

The first breakthrough of the Tomahawk 6 is the high-level bandwidth at 102.4 Tbits/s, Del Vecchio said. “It is the only switch chip at that bandwidth and it is twice the bandwidth of any other Ethernet silicon.

“We also have a variety of I/O options with 200G SerDes and 100G SerDes and we will offer a co-packaged optics option,” he continued. “So we give our customers a variety of interconnects so they can build out different types of systems.”

Del Vecchio also said the Tomahawk 6 is not just about the bandwidth—the speeds and feeds—it’s also about the telemetry. “We have load balancing, so we can add a lot of efficiency to the networks and really advance the industry toward that goal of AGI [artificial general intelligence] by making that whole GPU cluster more efficient.”

Broadcom sees a complete reimagining of how scale-up racks will be built in the next couple of years with the combination of the Tomahawk 6 and its interconnect options, including copper, linear pluggable optics (LPO), and CPO, and a new paradigm that it will enable for both scale-out and scale-up on Ethernet platforms for very large AI clusters, said Ram Velaga, senior vice president and general manager for the core switching group at Broadcom, during a panel event.

Tomahawk 6 200G Ethernet switch (Source: Broadcom Inc.)

The Tomahawk 6 supports a scale-up cluster size of 512 XPUs, offering a >7× scale-up cluster size compared with alternatives, and 100,000+ XPUs in a two-tier scale-out network at 200 Gbits/s per link.

Next-generation XPUs will commonly have 200G PAM4 interfaces, Del Vecchio said. “For those XPUs, you can connect 512 of them in a single hop in a scale-up network, so now you can have a scale-up domain of 512 XPUs, which is 7× the scale of any other scale-up technology.”

Two-tier networks provide several benefits. These include fewer optics and switches, while offering lower latency and power and higher reliability and performance than those of three-tier networks.

In the case of scale-out, the Tomahawk 6 reduces the number of tiers, which are needed in networking when deploying 100,000+ GPUs, Velaga said.

“When you have a 50-terabit switch like a Tomahawk 5, you need a three-tier network to be able to connect 100,000 GPUs,” Velaga said. “When you move to a Tomahawk 6 at 100 terabits, we’re able to reduce the number of tiers from three to two tiers, significantly reducing the amount of optics, the amount of switches, and thereby, the amount of power to support these clusters of 100,000+ GPUs.”

This might not sound like a big difference, but a three-tier network translates into 67% more optics and 3× the number of switches, so now the network power overall doubles, Del Vecchio said. “You can now decrease the power that’s going into the network and in turn put the resources into your XPUs.”

Beyond the scale-out, the Tomahawk 6 product line will be able to support a scale-up domain of 512 XPUs, each connected at 200G SerDes speeds, Velaga said. “When you’re able to do this and what we see happening going forward is reimagining these XPU racks today, which are constrained to topologies of 30 or 60 XPUs inside a physical rack.”

The Ethernet switch chip with 200G SerDes is said to provide the longest reach for passive copper interconnects, enabling high-efficiency, low-latency system design and delivering high reliability and reduced total cost of ownership.

Tomahawk 6 100G Ethernet switch (Source: Broadcom Inc.)

The Tomahawk 6 also claims a “groundbreaking” option for 1,024 100G SerDes on a single chip, enabling customers to deploy AI clusters with extended copper reach and efficient use of XPUs and optics with native 100G interfaces. It will be available with CPO for systems requiring optical connectivity, delivering low power and latency. It also will reduce link flaps and improve long-term reliability, which are critical for hyperscale AI network operators, Broadcom said.

AI-optimized routing

The Tomahawk 6 delivers AI-optimized routing, supporting scale-up and 1 million XPU scale-out networks. “For scale-out networks, it is common for today’s systems to have up to 10,000 or low tens of thousands of XPUs,” Del Vecchio said. “In the future, you’ll see that grow to 100,000 or a million XPUs.”

With high power and system efficiency for AI training and inference, the Ethernet switch features Cognitive Routing 2.0 that provides advanced telemetry, dynamic congestion control, rapid failure detection, and packet trimming. This enables global load balancing and adaptive flow control, tailored for modern AI workloads, including mixture-of-experts, fine-tuning, reinforcement-learning, and reasoning models, the company said.

Broadcom’s Cognitive Routing delivers 50% higher throughput and 10,000× faster response to link failures than standard Ethernet, according to the company.

The Tomahawk 6 is a generation ahead of the competition with 2× the bandwidth of any other Ethernet solution and it has the highest network efficiency with Cognitive Routing, Del Vecchio said. “It provides the best telemetry with deep insight and offers unprecedented scale, versatility of I/O, and system efficiency, and this will be combined with Broadcom’s third-generation CPO technology.”

By leveraging Ethernet, network operators get the benefit of a unified technology stack and consistent operation tools across the AI fabric while enabling fungible interfaces for cloud operators, allowing them to dynamically partition their XPU assets into the optimal configuration for different customer workloads, Broadcom said.

Tomahawk 6 is compliant with the Ultra Ethernet Consortium and supports modern AI transports, congestion signaling, and telemetry for large, distributed training environments. It also supports arbitrary network topologies, including scale-up, Clos, rail-only, rail-optimized, and torus.

Broadcom said multiple deployments are planned with more than 100,000 XPUs using Tomahawk 6 for both the scale-out and scale-up interconnects. It works with any NIC or XPU Ethernet endpoint.

Broadcom offers an end-to-end Ethernet AI platform that includes the Tomahawk and Jericho switch families, Thor NICs, Agera retimers, Sian optical DSPs, CPO, and software development kits. It offers open specifications for scale-up interfaces for XPUs and NICs. The Scale-Up Ethernet Framework, announced by Broadcom at Open Compute Project (OCP) Dublin in April 2025, is available at this link and will be shared with open standards development organizations, including OCP.

The Tomahawk6-200G 512 × 200G PAM4 and Tomahawk6-100G 1,024 × 100G PAM4 are shipping now.

Recommended
OFC 2025 unveils 1.6T networking innovations

The post Broadcom ships first 102.4-Tbits/s Ethernet switch appeared first on Electronic Products.