Broadcom Tomahawk 5: 800G ethernet switch silicon for AI-scale networks

Responsible: ad hoc news Lifestyle & Consumer Desk. Reviewed prior to publication on June 12, 2026 at 2:02 PM ET. Details in the imprint.

Broadcom’s **Tomahawk 5** ethernet switch silicon is one of the company’s flagship offerings for AI-scale networking, delivering up to 51.2 Tbps of bandwidth and 800G ethernet port speeds in a single chip for cloud and hyperscale data centers. With AI clusters pushing into tens of thousands of GPUs, the chip is designed to move training data and model parameters quickly enough to keep accelerators busy while managing power and congestion at scale. According to Broadcom, Tomahawk 5 provides double the bandwidth of the prior Tomahawk 4 generation while maintaining similar power per bit, a key metric for operators who now see the network as a major contributor to rack-level energy use.

What Broadcom Tomahawk 5 is built to do

Tomahawk 5 is a high-end, programmable ethernet switch ASIC targeted at top-of-rack, leaf, and spine switches in large cloud and AI data centers. It supports up to 64 ports of 800G or 128 ports of 400G in a single device, enabling network architects to build dense, non-blocking fabrics with fewer chips and chassis per cluster. The silicon underpins systems from networking OEMs and white-box vendors that sell fully built switches to hyperscalers, who then deploy them in Clos-based fabrics that connect servers and GPU nodes running large language models and other AI workloads. Broadcom positions Tomahawk 5 as a foundational building block for ethernet-based AI networking, alongside its Jericho 3-AI routing silicon and custom XPU accelerators.

To meet AI performance requirements, the chip integrates advanced congestion control and telemetry features intended to reduce tail latency and minimize packet drops during collective operations like all-reduce, which are central to distributed training. Broadcom’s architecture supports end-to-end congestion control protocols and in-band network telemetry that let operators monitor traffic hotspots in real time and tune buffer and queue behavior per flow class. The company has highlighted that Tomahawk 5 is engineered to work with RoCE (RDMA over Converged Ethernet) and other transport optimizations, which are widely used to get low-latency, loss-minimized communication on ethernet without moving to proprietary interconnects.

Power efficiency is another focus. Broadcom states that Tomahawk 5 maintains approximately the same power per 400G port as Tomahawk 4 while doubling total fabric capacity, which can help reduce the number of switches and optics required per AI pod. Fewer devices translate into lower aggregate power, simplified cabling, and potentially smaller network footprints per GPU cluster. Given that hyperscale operators now track power per bit as closely as raw throughput, the ability to scale to 51.2 Tbps within a single ASIC can be an important lever in their total cost of ownership models.

In terms of programmability, Tomahawk 5 supports modern SDN and network automation stacks that hyperscalers use to roll out features and policies across millions of ports. The chip is designed to integrate with Broadcom’s SDKs, merchant switch OS partners, and cloud providers’ in-house software, enabling custom routing, telemetry, and security functions to be pushed into the data plane. This degree of programmability is increasingly relevant for AI clusters, where operators are experimenting with job-aware routing and priority schemes aligned to different training workloads and service-level targets.

Broadcom ties Tomahawk 5 into its broader AI portfolio, which includes high-radix ethernet and InfiniBand alternatives, custom accelerators, and optical interconnect components. The company has recently highlighted that AI-related products, including networking silicon, are a major driver of its semiconductor revenue growth, citing strong demand from large cloud customers scaling out new AI regions. In this context, Tomahawk 5 serves as a core element of end-to-end ethernet fabrics that must keep up with rapid increases in GPU count and model size.

From a market perspective, Tomahawk 5 competes primarily with other 51.2 Tbps-class merchant silicon and proprietary switch ASICs used in hyperscale environments. Network architects weigh features such as buffer architecture, latency profiles, RoCE behavior, and ecosystem maturity when choosing a platform. Broadcom’s long-standing position in hyperscale switching, combined with widely adopted software tooling, gives Tomahawk 5 a strong installed-base advantage, especially for cloud providers already standardized on earlier Tomahawk generations. For buyers, the decision often comes down to how seamlessly the new silicon can slot into existing fabrics, optics portfolios, and automation workflows without disrupting AI cluster deployment timelines.

Within Broadcom, Tomahawk 5 sits in the networking segment that, along with other AI-focused products, has been singled out by analysts as a key contributor to recent revenue growth and to the company’s positioning in AI infrastructure. Shares of Broadcom Inc. (US11135F1012, ticker AVGO) traded at $385.57 on Nasdaq on June 12, 2026.

This article was created with a.i. assistance and editorially reviewed. Product information is provided without warranty; prices and availability may change at any time. Not investment advice, not a buy or sell recommendation. Trading in securities carries risks up to the total loss of capital.