Broadcom has begun shipping its Tomahawk Ultra Ethernet switch, setting a new benchmark for data center interconnects targeting high-performance computing (HPC) and AI scale-up workloads. Engineered for ultra-low latency, massive bandwidth, and lossless transport, the 51.2 Tbps switch achieves 250ns latency at full throughput and delivers 64-byte line-rate performance at up to 77 billion packets per second. By integrating in-network collectives and advanced flow control mechanisms, Tomahawk Ultra redefines Ethernet as a viable, standards-based fabric for tightly coupled AI and HPC environments.
The switch introduces a programmable, application-specific Ethernet header format that reduces overhead from 46 bytes to just 10 while maintaining full Ethernet compliance. This enables higher efficiency and lower latency for XPU-to-XPU communication. With support for topology-aware routing in Dragonfly, Mesh, and Torus configurations, and compliance with the Ultra Ethernet Consortium (UEC) specifications, the chip is optimized for both scale-up and scale-out architectures. Tomahawk Ultra also supports Scale-Up Ethernet (SUE), delivering sub-400ns end-to-end latency, and the new SUE-Lite variant for space- and power-constrained AI accelerators.
To enhance reliability in congested, high-throughput environments, the switch implements Link Layer Retry (LLR) with Forward Error Correction to recover packet errors without drops, alongside Credit-Based Flow Control (CBFC) to prevent buffer overflows. Tomahawk Ultra also integrates in-network collective operations such as AllReduce and AllGather directly into the switch silicon, freeing up valuable compute resources and reducing training times. The result is a scale-up AI fabric that rivals proprietary interconnects while remaining standards-based and vendor-neutral.
Broadcom pairs Tomahawk Ultra with its previously announced 102.4 Tbps Tomahawk 6 to offer a unified Ethernet architecture. While Tomahawk Ultra handles scale-up AI clusters and HPC backplanes, Tomahawk 6 is designed for massive scale-out infrastructure. The new switch is pin-compatible with the Tomahawk 5, enabling rapid deployment and system upgrades with minimal reengineering.
Tomahawk Ultra Highlights
- 51.2 Tbps switching bandwidth
- 250ns switch latency at full load
- 64B line-rate switching; 77B packets/sec performance
- 10B optimized Ethernet header (vs. standard 46B)
- In-network collectives for AI training efficiency
- Link Layer Retry + CBFC = true lossless Ethernet fabric
- Topology-aware routing for Dragonfly, Mesh, and Torus
- Full compliance with UEC and SUE/SUE-Lite frameworks
- Pin-compatible with Tomahawk 5; now shipping for AI training clusters
“Tomahawk Ultra is a testament to innovation, involving a multi-year effort by hundreds of engineers who reimagined every aspect of the Ethernet switch,” said Ram Velaga, senior vice president and general manager of Broadcom’s Core Switching Group.
Broadcom’s ecosystem partners are already lining up to endorse and integrate Tomahawk Ultra:
- Accton: “The Tomahawk Ultra is a perfect solution for building low-latency, lossless systems ready for scale-up AI and HPC applications.” – Michael KT Lee, SVP
- AMD: “Combining Tomahawk Ultra with AMD Instinct GPUs and EPYC processors enables standards-based Ethernet AI infrastructure.” – Forrest Norrod, EVP
- Arista Networks: “Ultra-low latency and Scale-Up Ethernet make Tomahawk Ultra ideal for AI networking.” – Simon Capper, Principal Engineer
- Arrcus: “Ultra-low latency and a lossless fabric significantly accelerate job completion times.” – Shekar Ayyar, CEO
- Delta Electronics: “A game-changer for AI scale-up and HPC. Our 51.2T Ethernet platform is ready.” – Wangson Wang, GM
- HPE Networking: “Its ultra-low latency and in-network collectives align with what today’s workloads demand.” – Praveen Jain, SVP
- Intel: “Enables 76.8TB/s HBM bandwidth across 64 Gaudi 3 accelerators per rack.” – Saurabh Kulkarni, VP
- Inventec: “Industry’s lowest switch latency and high energy efficiency make it a key part of our AI strategy.” – Vincent Lin, GM
- IPI: “A bold leap forward in AI and HPC clusters, built on open Ethernet.” – Kiyo Oishi, CEO
- Micas Networks: “Delivers performance, reliability, and efficiency for AI at scale.” – Andrew Qu, CEO
- Nexthop AI: “Drives the next generation of scale-up Ethernet solutions with Nexthop SONiC.” – Anshul Sadana, CEO
- QCT (Quanta Cloud Technology): “Pushes the next frontier of AI with Ethernet-based infrastructure.” – Mike Yang, President
- UfiSpace: “Pin-compatible with Tomahawk 5; enables seamless upgrade path.” – Vincent Ho, CEO
- Wistron: “Supports scalable Open Ethernet AI systems.” – Robert CL Lin, President
- WNC: “Delivers advanced, open Ethernet networks for future-ready deployments.” – Johnson Hsu, SVP
