PALO ALTO, Calif., July 15, 2025 — Broadcom Inc. (NASDAQ:AVGO), a world chief in semiconductor and infrastructure software program options, in the present day introduced the cargo of its breakthrough Ethernet swap — the Tomahawk Extremely. Engineered to rework the Ethernet swap for high-performance computing (HPC) and AI workloads, Tomahawk Extremely delivers industry-leading ultra-low latency, huge throughput, and lossless networking.
“Tomahawk Extremely is a testomony to innovation, involving a multi-year effort by tons of of engineers who reimagined each facet of the Ethernet swap,” stated Ram Velaga, senior vp and basic supervisor of Broadcom’s Core Switching Group. “This highlights Broadcom’s dedication to put money into advancing Ethernet for high-performance networking and AI scale-up.”
Constructed from the bottom as much as meet the intense calls for of HPC environments and tightly coupled AI clusters, Tomahawk Extremely redefines what an Ethernet swap can ship. Lengthy perceived as higher-latency and lossy, Ethernet takes on a brand new position:
- Extremely-low latency: Achieves 250ns swap latency at full 51.2 Tbps throughput.
- Excessive efficiency: Delivers line-rate switching efficiency even at minimal packet sizes of 64 bytes, supporting as much as 77 billion packets per second.
- Adaptable, optimized Ethernet headers: Reduces header overhead from 46 bytes right down to as little as 10 bytes, whereas sustaining full Ethernet compliance —boosting community effectivity and enabling versatile, application-specific optimizations.
- Lossless material: Implements Hyperlink Layer Retry (LLR) and Credit score-Based mostly Circulate Management (CBFC) to get rid of packet loss and guarantee reliability.
“AI and HPC workloads are converging into tightly coupled accelerator clusters that demand supercomputer-class latency — crucial for inference, reliability, and in-network intelligence from the material itself,” stated Kunjan Sobhani, lead semiconductor analyst, Bloomberg Intelligence. “Demonstrating that open-standards Ethernet can now ship sub-microsecond switching, lossless transport, and on-chip collectives marks a pivotal step towards assembly these calls for of an AI scale-up stack — projected to be double digit billions in a couple of years.”
Tomahawk Extremely is optimized for the tightly coupled, low-latency communication patterns present in each high-performance computing programs and AI clusters. With ultra-low latency switching and adaptable optimized Ethernet headers, it gives predictable, high-efficiency efficiency for large-scale simulations, scientific computing, and synchronized AI mannequin coaching and inference.
When deployed with Scale-Up Ethernet (SUE specification obtainable right here), Tomahawk Extremely permits sub-400ns XPU-to-XPU communication latency, together with the swap transit time — setting a brand new benchmark for tightly synchronized AI compute at scale.
By decreasing Ethernet header overhead from 46 bytes to simply 10 bytes, whereas sustaining full Ethernet compliance, Tomahawk Extremely dramatically improves community effectivity. This optimized header is adaptable per utility, providing each flexibility and efficiency features throughout numerous HPC and AI workloads.
Tomahawk Extremely incorporates lossless material expertise that eliminates packet drops throughout high-volume knowledge switch. Incorporating LLR, the swap detects hyperlink errors utilizing Ahead Error Correction and robotically retransmits packets, avoiding drops on the wire degree. Concurrently, CBFC prevents buffer overflows that historically prompted packet loss. Collectively, these mechanisms create a really lossless Ethernet material, delivering the extent of reliability demanded by in the present day’s most data-intensive workloads.
Tomahawk Extremely additionally accelerates efficiency by In-Community Collectives fixing one of the persistent bottlenecks in AI and machine studying workloads. Somewhat than burdening XPUs with collective operations like AllReduce, Broadcast, or AllGather, Tomahawk Extremely executes these instantly throughout the swap chip. This will cut back job completion time and enhance utilization of pricey compute sources. Importantly, this functionality is endpoint-agnostic, enabling speedy adoption throughout a variety of system architectures and vendor ecosystems.
Designed with improvements in topology-aware routing to assist superior HPC topologies together with Dragonfly, Mesh and Torus, Tomahawk Extremely can also be compliant with the UEC commonplace and embraces the openness and wealthy ecosystem of Ethernet networking.
As a part of Broadcom’s Ethernet-forward technique for AI scale-up, the corporate has launched SUE-Lite — an optimized model of the SUE specification tailor-made for energy and area-sensitive accelerator purposes. SUE-Lite retains the important thing low-latency and lossless traits of full SUE, whereas additional decreasing the silicon footprint and energy consumption of Ethernet interfaces on AI XPUs and CPUs.
This light-weight variant permits simpler integration of standards-compliant Ethernet materials in AI platforms, selling broader adoption of Ethernet because the interconnect of selection in scale-up architectures.
Along with the 102.4 Tbps Tomahawk 6, Tomahawk Extremely varieties the muse of a unified Ethernet structure: enabling scale-up Ethernet for AI, and scale-out Ethernet for HPC and distributed workloads.
Tomahawk Extremely is one hundred pc pin-compatible with Tomahawk 5, guaranteeing a really quick time-to-market. It’s delivery now for deployment in rack-scale AI coaching clusters and supercomputing environments.
