Broadcom has introduced the Tomahawk Ultra, a groundbreaking Ethernet switch designed to meet the demands of high-performance computing (HPC) and artificial intelligence (AI) applications. This innovative switch delivers ultra-low latency of 250 nanoseconds, achieving a throughput of 51.2 terabits per second, which drastically enhances communication in tightly-coupled AI clusters.
The Tomahawk Ultra aims to challenge traditional high-performance interconnects like NVIDIA’s NVLink by supporting up to 1024 accelerators, offering four times the connectivity compared to its counterparts. Its enhanced architecture redefines Ethernet performance, allowing it to process up to 77 billion packets per second while reducing header overhead from 46 bytes to just 10 bytes.
Additionally, the switch incorporates advanced features such as Lossless Fabric, which utilizes Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC) to prevent packet loss, ensuring reliable data transfer. Notably, it can also perform in-network collective operations directly within the switch, alleviating the computational burden on accelerator devices.
The Tomahawk Ultra is fully pin-compatible with its predecessor, the Tomahawk 5, easing integration for OEMs while promoting quick deployment in AI and HPC environments. With the emergence of AI workloads demanding supercomputer-level performance, Broadcom’s Tomahawk Ultra is poised to play a pivotal role in defining the network architecture of next-generation data centers.
By adopting this switch, companies can significantly enhance their AI capabilities and optimize their data processing infrastructures, solidifying Ethernet’s position as the interconnect of choice for high-performance applications.more secured and convenient authentication methods, offering our customers a robust solution that meets their needs.”
Share to your social below!
