Broadcom has begun shipping its latest Ethernet switch, Tomahawk Ultra, which targets improvements in high-performance computing (HPC) and AI workloads.
According to Broadcom, Tomahawk Ultra achieves a switch latency of 250 nanoseconds at a full 51.2 terabits per second of throughput. It supports line-rate switching for packets as small as 64 bytes and can handle up to 77 billion packets per second.
Go deeper with GlobalData
The switch optimises Ethernet headers by reducing overhead from 46 bytes to as low as ten bytes while maintaining compliance. This enhancement aims to boost network efficiency and allows for application-specific customisations.
Broadcom stated that it tailored the switch for HPC environments and AI clusters, providing ultra-low latency and adaptable Ethernet headers for efficient large-scale simulations and AI model training.
When used with scale-up Ethernet (SUE), the Tomahawk Ultra facilitates sub-400 nanosecond communication latency between processing units, including transit time through the switch.
By incorporating lossless fabric technology, it eliminates packet drops during data transfers. The use of Link Layer Retry and Credit-Based Flow Control ensures reliability by detecting link errors and automatically retransmitting packets, creating a lossless Ethernet fabric ideal for data-intensive workloads.
US Tariffs are shifting - will you react or anticipate?
Don’t let policy changes catch you off guard. Stay proactive with real-time data and expert analysis.
By GlobalDataAnother key feature of Tomahawk Ultra is its execution of In-Network Collectives, which addresses bottlenecks in AI and machine learning tasks. Executing operations like AllReduce and Broadcast within the switch chip reduces job completion time and maximises resource utilisation across different system architectures.
Broadcom senior vice-president and core switching group general manager Ram Velaga said: “Tomahawk Ultra is a testament to innovation, involving a multiyear effort by hundreds of engineers who reimagined every aspect of the Ethernet switch. This highlights Broadcom’s commitment to invest in advancing Ethernet for high-performance networking and AI scale-up.”
Additionally, Broadcom has introduced SUE-Lite, an optimised version of SUE designed for power-sensitive applications. SUE-Lite retains low-latency and lossless features while reducing silicon footprint and power usage, aiding integration into AI platforms. This development fits within Broadcom’s strategy to enhance AI scale-up infrastructure alongside HPC scale-out capabilities.
The previously announced Tomahawk 6 series provides 102.4 terabits per second of switching capacity. Together with new features in the Tomahawk Ultra, these advancements aim to support AI clusters with routing capabilities and interconnect options suited for large-scale deployment.
