Broadcom’s latest Ethernet switch, the Tomahawk 6 series, delivers record-setting bandwidth at a time when AI systems are pushing network infrastructures to their limits.
As AI deployments scale, one of the biggest bottlenecks is not compute alone but the movement of massive amounts of data between processors. Broadcom’s announcement that it is shipping the Tomahawk 6 switch family aims to address that challenge by dramatically increasing network capacity.
The Tomahawk 6 offers 102.4 terabits per second (Tbps) of switching capacity—about twice the bandwidth of competing products on the market. That leap in throughput arrives as organizations build ever-larger AI clusters and require higher-performance fabrics to maintain low latency and high utilization.
AI models are growing rapidly in size and complexity, and datacenter networks connecting thousands of specialized accelerators are feeling the strain. Broadcom designed Tomahawk 6 specifically to meet those demands, integrating features intended to boost bandwidth, efficiency, and routing intelligence for both scale-up and scale-out topologies.
Ram Velaga, Senior Vice President and General Manager of Broadcom’s Core Switching Group, described the product as more than a routine upgrade: “Tomahawk 6 is not just an upgrade – it’s a breakthrough. 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.”
Such statements reflect strong market interest. Semiconductor analyst Kunjan Sobhani of Bloomberg Intelligence highlights the wider context: “AI clusters are scaling from tens to thousands of accelerators, turning the network into a critical bottleneck while expected to deliver unprecedented bandwidth and low latency.”
Beyond headline throughput, Tomahawk 6 brings several engineering refinements that matter in large-scale deployments. One variant integrates 1,024 100G SerDes on a single chip, enabling longer copper runs without frequent conversion to optical links. That can reduce cost, simplify wiring, and lower power consumption for certain topologies.
Acknowledging the inevitability of optical networking at scale, Broadcom also offers Tomahawk 6 with co-packaged optics (CPO). Co-packaging moves optical components into the switch package, reducing interconnect power and latency compared with separate pluggable optics—advantages that compound when thousands of accelerators are interconnected.
Networks to scale AI to new heights
A notable feature in Tomahawk 6 is “Cognitive Routing 2.0,” a set of capabilities that dynamically route traffic based on real-time network conditions. For advanced AI workloads—such as mixture-of-experts models or distributed reinforcement learning—the ability to adaptively steer traffic can significantly improve performance and resource utilization.
Infrastructure engineers argue that scaling future AI systems will require rethinking networking architectures. A single technology that can support both scale-out (connecting many nodes) and scale-up (building very large individual systems) simplifies design and can enable new classes of workloads. Broadcom claims that multiple organizations are already planning deployments that will use Tomahawk 6 to support more than 100,000 XPUs (accelerated processors) across two-tier networks, and up to 512 XPUs in a single scale-up cluster.
Opening the ecosystem
Broadcom is also pushing for open standards rather than proprietary lock-in. The company introduced the Scale Up Ethernet (SUE) Framework at the Open Compute Project earlier this year, sharing specifications intended to promote interoperability and broader ecosystem adoption. This open approach aims to encourage a diverse hardware and software ecosystem around high-capacity Ethernet fabrics.
Tomahawk 6 fits into Broadcom’s broader portfolio—including Jericho switches, Thor NICs, and various optical components—positioning the company’s products as an end-to-end Ethernet platform for large-scale AI deployments. That integrated approach, combined with open specification work, is intended to give hyperscalers and cloud providers flexible options without being locked into proprietary solutions.
As AI systems scale toward previously theoretical sizes, from tens of thousands to potentially millions of processors, the networking layer will be a key determinant of overall performance and cost. If Tomahawk 6 achieves its goals—delivering 100+ Tbps of capacity, reduced power and latency through co-packaged optics, and intelligent routing to optimize traffic—it could become a foundational technology for next-generation AI infrastructure.
Want to learn more about AI and big data from industry leaders? Attend AI & Big Data Expo events held in Amsterdam, California, and London, where industry sessions explore AI infrastructure, networking, and enterprise applications.