Broadcom Launches 51.2TB/s Tomahawk Ultra Switch, Leading a New Revolution in HPC and AI Networking
At the Hot Chips 2025 conference, Broadcom Inc., a global leader in semiconductor and infrastructure software solutions, unveiled its latest Ethernet Switch, the 51.2TB/s Tomahawk Ultra. Described by Broadcom as a "low-latency scale-up switch purpose-built for high-performance computing (HPC) and artificial intelligence (AI) applications," the product marks a significant step forward in Ethernet's ability to address the demands of ultra-high-performance networking.
Breaking the Tradition, Built for HPC and AI
The Tomahawk Ultra aims to redefine the role of Ethernet switches in HPC and AI workloads. Ethernet has long been considered limited in terms of latency and lossless data transmission, but Tomahawk Ultra overcomes these limitations through a series of innovative technologies. At a fully loaded 51.2Tbps throughput, it achieves an industry-leading switching latency of 250ns, enabling near-instantaneous data transfer for latency-sensitive applications such as HPC simulations and AI inference. Even when processing packets as small as 64 bytes, the Tomahawk Ultra maintains a line-speed switching capacity of 77 billion packets per second, ensuring efficient and stable network transmission even under extreme conditions. This is particularly critical for the dense transmission of small packets frequently encountered in AI training, significantly improving training efficiency.
Optimizing Ethernet Headers to Improve Transmission Efficiency
The Tomahawk Ultra utilizes an adaptive, optimized Ethernet header design, reducing the traditional 46-byte header overhead to a minimum of 10 bytes while fully maintaining compliance with Ethernet standards. This revolutionary design significantly improves network transmission efficiency, reduces the additional burden on data transmission, and frees up valuable bandwidth resources for actual data transmission. For AI applications with large volumes of data and frequent transmission, this optimized header design significantly reduces the risk of network congestion and improves overall system performance. Building a Lossless Network to Ensure Reliable Data Transmission
To meet the stringent data transmission reliability requirements of HPC and AI applications, the Tomahawk Ultra implements Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC) technologies, creating a truly lossless network fabric. LLR, through forward error correction, detects and automatically retransmits erroneous packets in real time, ensuring data accuracy. CBFC, by monitoring the status of the receiver's buffer in real time, dynamically adjusts the data transmission rate, effectively preventing packet loss due to buffer overflow. The combination of these two technologies fundamentally eliminates the risk of packet loss and ensures reliable data transmission for HPC and AI workloads.
Supporting Scale-Up and Scale-Out Strategies, Expanding the Boundaries of Ethernet Applications
The Tomahawk Ultra is the successor to Broadcom's 102.4TB/s Tomahawk 6 switch. Its launch further complements Broadcom's strategic plan to promote the widespread adoption of Ethernet in scale-up and scale-out applications. For scale-up, the Tomahawk Ultra is deeply optimized for the tightly coupled, low-latency communication model found in HPC systems and AI clusters. Through ultra-low latency switching and adaptively optimized Ethernet headers, it delivers predictable and efficient performance for applications such as large-scale simulations, scientific computing, and simultaneous AI model training and inference. When combined with Broadcom's Scale-Up Ethernet (SUE) specification, the Tomahawk Ultra achieves sub-400ns XPU (Accelerated Processing Unit)-to-XPU communication latency, including switch transfer time, setting a new performance benchmark for large-scale, simultaneous AI computing.
For scale-out, the Tomahawk Ultra, with its powerful processing power and flexible configuration options, also meets the needs of large-scale data center scale-out, enabling the construction of larger, more efficient network architectures and providing enterprises and cloud service providers with more scalable and cost-effective network solutions. Industry experts highly praise the broad market prospects.
"AI and HPC workloads are increasingly converging into tightly coupled accelerator clusters, which demand supercomputer-level latency, making them crucial for inference, reliability, and the network's own intelligence," said Kunjan Sobhani, chief semiconductor analyst at Bloomberg Intelligence. "Broadcom's Tomahawk Ultra demonstrates that open-standard Ethernet can now achieve sub-microsecond switching, lossless transmission, and chip-level collective operations. This is a key step in meeting the needs of AI scale-up systems, a market expected to reach billions of dollars in the coming years."
The launch of the Tomahawk Ultra not only brings a high-performance network device to the HPC and AI sectors but also injects new vitality into the development of Ethernet technology as a whole. As AI and HPC technologies continue to advance, the demand for high-performance networks will continue to grow. The Tomahawk Ultra is expected to occupy a key position in this booming market, driving the related industries into a new stage of development.