Tomahawk Ultra redefines Ethernet for AI, HPC performance

Broadcom’s new Tomahawk Ultra Ethernet switch is rewriting the rules for AI and high-performance computing. Designed from scratch, it slashes latency to 250ns at full 51.2 Tbps throughput while ensuring lossless data flow—crucial for large-scale AI training. With in-switch collective operations and a drastic cut in header overhead, it’s not just a faster chip—it’s a strategic blueprint for Ethernet’s evolving role in the AI arms race.