Broadcom has officially begun shipping the Tomahawk Ultra, its latest high-performance Ethernet switch designed to meet the surging demands of AI and high-performance computing (HPC) workloads. Representing a significant reimagining of Ethernet architecture, the Tomahawk Ultra offers ultra-low latency, lossless networking, and powerful in-network compute capabilities—signaling a new chapter in scale-up computing.
A Reinvention of Ethernet for AI and HPC
For years, Ethernet was seen as a high-latency, lossy option in environments that demanded precise, high-speed data movement. Tomahawk Ultra changes that. Purpose-built for tightly coupled AI clusters and HPC environments, the switch delivers an industry-first: 250ns switch latency at a massive 51.2 Tbps throughput. It supports 77 billion packets per second, even at minimum 64-byte packet sizes.
This level of performance redefines the capabilities of Ethernet, bringing it into supercomputer territory. According to Kunjan Sobhani, lead semiconductor analyst at Bloomberg Intelligence, “Demonstrating that open-standards Ethernet can now deliver sub-microsecond switching, lossless transport, and on-chip collectives marks a pivotal step toward meeting the demands of an AI scale-up stack.”
Engineered for Zero Compromise
At the heart of Tomahawk Ultra are multiple innovations that help it meet the communication needs of large-scale AI and HPC systems:
-
Lossless Fabric: Traditional Ethernet suffers from packet loss under heavy load. Tomahawk Ultra eliminates this with Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC), ensuring data integrity even in high-volume transfers.
-
Header Optimization: The switch trims Ethernet header overhead from 46 bytes to as little as 10 bytes, boosting efficiency without sacrificing compliance. These headers can be dynamically adjusted per application, increasing performance across different workloads.
-
In-Network Collectives: One of the most transformative features, Tomahawk Ultra can perform operations like AllReduce and Broadcast directly within the switch. This offloads computation from XPUs (e.g., GPUs or specialized AI chips), improving utilization and reducing job completion times.
-
Topology-Aware Routing: Supporting advanced HPC network designs such as Dragonfly, Mesh, and Torus, the switch adapts to the unique routing needs of large-scale computing environments.
A Platform for Scale-Up AI
When paired with Broadcom’s Scale-Up Ethernet (SUE) specification, Tomahawk Ultra can achieve sub-400ns latency from XPU to XPU, including transit time through the switch. That level of performance is crucial for synchronized model training and inference workloads that dominate the AI landscape.
To address power-sensitive applications, Broadcom has also introduced SUE-Lite — a streamlined version of the SUE spec optimized for energy and area efficiency. SUE-Lite retains core benefits like low-latency and lossless transport, making it easier to integrate standards-based Ethernet into AI accelerators and CPUs.
Part of a Broader Vision
Tomahawk Ultra doesn’t stand alone. It complements Broadcom’s Tomahawk 6—a 102.4 Tbps switch designed for massive scale-out deployments. Together, the two form a unified Ethernet fabric capable of supporting both AI scale-up and HPC scale-out architectures.
Broadcom emphasizes that the Tomahawk Ultra is pin-compatible with Tomahawk 5, easing integration and accelerating time-to-market for customers building the next generation of supercomputers and AI data centers.
Now Shipping
Tomahawk Ultra is already being deployed in large-scale AI training clusters and HPC environments. Broadcom’s latest innovation not only meets today’s data and compute challenges but positions Ethernet as the backbone of tomorrow’s intelligent infrastructure.
For more on the Tomahawk Ultra and the full Scale-Up/Scale-Out architecture, visit Broadcom’s official product page and media kit.
