Broadcom has begun shipping its latest Ethernet switch, the Tomahawk Ultra, a device built specifically for the performance needs of large-scale AI and high-performance computing (HPC) systems. With a switch latency of just 250 nanoseconds and support for 51.2 terabits per second (Tbps) throughput, the Tomahawk Ultra is designed to handle the kind of tightly coupled, data-heavy workloads found in next-generation data centres.
The switch also introduces several technical features aimed at overcoming long-held assumptions about Ethernet in HPC and AI environments, including concerns around latency, packet loss, and header overhead.
“Tomahawk Ultra is the result of years of focused engineering,” said Ram Velaga, Senior Vice President and General Manager of Broadcom’s Core Switching Group. “It reflects our investment in evolving Ethernet to meet the needs of AI-scale performance.”
Built for Low Latency and High Throughput
At the core of Tomahawk Ultra is a reworked switch architecture that delivers line-rate performance even with small 64-byte packets, enabling up to 77 billion packets per second.
The switch introduces a new Ethernet header format that can reduce overhead from the typical 46 bytes to just 10, while staying within Ethernet protocol standards. This frees up bandwidth and allows for more efficient packet delivery, especially in AI clusters where performance per watt and per byte matters.
Tomahawk Ultra also supports Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC), two features that help eliminate packet drops. Together, they form a lossless fabric designed to support consistent, deterministic communication between accelerators.
Optimised for AI Scale-Up
AI workloads are shifting toward tightly connected accelerator-based clusters. These systems require precise timing and high throughput, particularly for synchronised model training and inference. Tomahawk Ultra is tailored for this environment, allowing for sub-400ns XPU-to-XPU latency, including switch transit, when used with Broadcom’s Scale-Up Ethernet (SUE) specification.
To further extend flexibility, Broadcom has introduced SUE-Lite, a stripped-down version of the SUE spec that maintains low-latency and lossless operation while reducing silicon area and power consumption. This makes it easier to integrate Ethernet into power-sensitive designs such as AI XPUs and CPUs.
Collective Operations On-Chip
One of the key bottlenecks in large-scale AI systems is the overhead of collective operations like AllReduce or Broadcast. These are typically handled at the system level and can consume significant compute resources.
Tomahawk Ultra shifts this work into the switch hardware itself. Its in-network collective capability enables the switch to handle synchronisation operations internally, without involving XPUs. That means faster job completion, better utilisation of compute resources, and improved overall system efficiency. Crucially, this functionality is endpoint-agnostic, meaning it works across a wide range of platforms and doesn’t require proprietary integration.
Topology-Aware and Backward-Compatible
The Tomahawk Ultra supports a range of advanced HPC topologies, including Dragonfly, Mesh, and Torus, making it suitable for both AI and traditional scientific computing.
It also complies with the UEC standard, reinforcing its interoperability in open Ethernet environments. For existing customers, the switch is pin-compatible with Tomahawk 5, allowing for quick drop-in upgrades with minimal redesign effort.
Deployment and Roadmap
Tomahawk Ultra is already being deployed in rack-scale AI training clusters and supercomputing platforms. It complements Broadcom’s Tomahawk 6, a 102.4 Tbps switch, allowing customers to build unified Ethernet infrastructures for both scale-up and scale-out workloads.
By bridging low-latency switching, reliable data transport, and in-network compute, Broadcom is positioning Ethernet not just as a fabric for general-purpose networking, but as a viable backbone for large-scale AI and HPC.
Conclusion
The Tomahawk Ultra shows how far Ethernet has come, and where it’s headed next. With 250ns switching, in-network collectives, and lossless operation, Broadcom is giving AI system designers a powerful alternative to proprietary fabrics. This could be a turning point for open, high-performance interconnects in one of the most demanding areas of modern computing.
Industry Reaction
Michael KT Lee, Senior Vice President, Research and Development Center, Accton
“Networking needs within an XPU node are as critical as those between nodes. With 51.2 Tbps Ethernet switching, 250-ns latency, credit-based flow control, and configurable optimized header, the Tomahawk Ultra is a perfect solution for building high-bandwidth, high-reliability, high-efficiency, and low-latency lossless systems ready for scale-up AI and HPC applications. Accton is excited to embrace the launch and collaborate closely with Broadcom to bring the Tomahawk Ultra solutions to the market.”
Forrest Norrod, Executive Vice President and General Manager, Data Center Solutions Group, AMD
“Low latency is essential to unleashing the full potential of AI, from reducing training times to powering real-time inference. By combining Broadcom’s new Tomahawk Ultra switch with AMD Instinct™ GPUs and EPYC™ processors, we’re enabling high-performance, standards-based Ethernet solutions for AI infrastructure. Together, we’re advancing an open ecosystem that brings our vision of AI everywhere, for everyone, closer to reality.”
Simon Capper, Principal Engineer for AI networking, Arista Networks
“Arista appreciates the combination of ultra-low latency and scale-up Ethernet innovations of Tomahawk Ultra for AI networking. Once again Broadcom is setting the pace in the AI and the switch industry.”
Shekar Ayyar, Chairman and CEO, Arrcus
“The launch of Broadcom’s Tomahawk Ultra marks a groundbreaking advancement in Ethernet innovation, particularly for AI and HPC scale-up environments. With ultra-low latency and a lossless fabric, it significantly accelerates job completion times, critical for modern AI workloads. At Arrcus, we’re proud to champion an open, standards-based networking ecosystem. Combined with our high-performance ArcOS network operating system, customers can unlock scalable infrastructure that is both flexible and future-ready.”
Wangson Wang, General Manager of Data Networks Infrastructure, Delta Electronics
“Delta Electronics is constantly looking ahead, and we’re thrilled to confirm that our 51.2T Ethernet switch platform is ready to harness the full power of Broadcom's Tomahawk Ultra chip. We see Tomahawk Ultra as a game-changer for AI scale-up and HPC. The collaboration between Delta and Broadcom demonstrates our dedication to pushing the boundaries of what is possible in Data Center network infrastructure. Building on the success of Delta’s current 800G switches, the newly launched Tomahawk Ultra chips enable us to deliver advanced solutions that offer not only unmatched performance and efficiency, but also high reliability and scalability for the most demanding network workloads, supporting rapid AI/ML network deployments for our customers.”
Praveen Jain, Senior Vice President and General Manager, AI Clusters and Cloud Ready Data Center, HPE Networking
“HPE is committed to delivering open, high-performance and easy-to-manage Ethernet-based solutions for the modern data center. We commend Broadcom on its new offering, and its ultra-low latency, high throughput and support for in-network collectives align perfectly with what today’s workloads demand. It reflects our shared vision for building the most advanced and open data center infrastructure solutions with operational simplicity at its core."
Saurabh Kulkarni, Vice President, AI Technical Product Management, Intel
“Broadcom's Tomahawk Ultra Series with its high throughput and ultra-low latency enables all-to-all connectivity across up to 64 Intel® Gaudi® 3 AI accelerators per rack with total HBM bandwidth of 76.8TB/s, capable of scaling the connectivity across multiple racks. This rack-level bandwidth unlocks new possibilities for training and real-time inference of the most complex LLMs, redefining industry SLAs. Through our collaboration with Broadcom, Intel is showcasing the open architecture and modular design advantage and full capability of our rack scale platform built for large-scale, global AI deployments.”
Vincent Lin, General Manager, Inventec EBG
“Inventec congratulates Broadcom on the launch of Tomahawk Ultra Ethernet switch, which significantly enhances the efficiency and sustainability of AI solutions by delivering the industry’s lowest switch latency, 250 nanoseconds, and leading power efficiency with 800W at 51.2T performance. At Inventec, our vision is to develop cutting-edge artificial intelligence products that drive sustainable change for humanity and the environment through close partnership with Broadcom to deliver high-performance, scalable solutions, supporting customers’ evolving AI and high-performance computing needs.”
Kiyo Oishi, CEO, IPI
“The Tomahawk Ultra represents a bold leap forward in AI workloads and HPC clusters, delivering an unmatched combination of bandwidth, latency, and cutting-edge features like In-Network Collectives and scale-up Ethernet. By leveraging non-proprietary Ethernet, the Tomahawk Ultra will empower customers to scale their data-intensive applications with unparalleled performance, efficiency, and reliability, paving the way for groundbreaking innovations in data-intensive computing.”
Andrew Qu, CEO, Micas Networks
“Broadcom’s Tomahawk Ultra is a major step forward for scale-up Ethernet in AI and HPC. With 250ns latency, 51.2 Tbps switching, and advanced features like Link Layer Retry, In-Network Collectives, and the AI Fabric Header, it delivers the performance, reliability, and efficiency our customers need for AI at scale. Thanks to pin compatibility with Tomahawk 5, Micas can rapidly bring Tomahawk Ultra-based systems to market, enabling seamless upgrades to meet the demands of next-generation AI infrastructure.”
Anshul Sadana, Founder and CEO, Nexthop AI
“With Tomahawk Ultra, Broadcom has driven AI Networking to a new level, allowing us to enable a new generation of low latency and lossless scale-up Ethernet solutions. Along with Nexthop SONiC, we now offer some of the most efficient scale-up and UEC compatible scale-out Ethernet solutions for the world’s largest hyperscalers.”
Mike Yang, President, Quanta Cloud Technology
“At QCT, we are committed to delivering next-generation AI and HPC infrastructure that meets the demands of extreme scale, performance, and efficiency. Broadcom’s Tomahawk Ultra Ethernet switch is a game-changer for the AI era, enabling 51.2 Tbps of switching capacity with ultra-low 250ns latency to dramatically accelerate AI training and inferencing workloads. We are excited to continue collaborating with Broadcom to push the next frontier of AI with Ethernet-based infrastructure.”
Vincent Ho, CEO of UfiSpace
“The Tomahawk Ultra delivers high performance and full pin-to-pin compatibility with Tomahawk 5. This seamless upgrade path shortens our development cycle for next-generation platforms, and we’re excited to integrate it into our upcoming solutions.”
Robert CL Lin, President of Enterprise and Networking Business Group, Wistron
“Broadcom’s Tomahawk Ultra sets a new benchmark in Open Ethernet for AI and HPC. Designed for GPU scale-up, the Tomahawk Ultra achieves 250ns latency at 51.2 Tbps, supporting 64B line-rate switching and lossless fabrics. This innovation represents a significant step forward for the industry. Wistron is seamlessly aligning these scalable AI systems, and the Tomahawk Ultra solution offering.”
Johnson Hsu, Senior Vice President and General Manager, WNC
“We’re proud to partner with Broadcom on the innovative Tomahawk Ultra. Purpose-built for the demands of AI and HPC, this advanced platform combines high performance with open Ethernet flexibility, enabling our customers to deploy scalable, reliable, and future-ready networks.”
Read the original announcement here.
Related Reading:
You may also like
