AMD details Pensando networking stack for Helios racks

AMD positioned networking as the third pillar of its Advancing AI 2026 rack-scale strategy, detailing the Pensando silicon and fabric technology behind the Helios rack. The system is designed to handle the demands of connecting many GPUs, memory pools and high-speed data paths within a single rack while also supporting scale-out links across multiple racks in the data center.

The company tied the push to rapid AI model growth, from the original 65-million-parameter Transformer in 2017 to trillion-parameter GPT-4-class systems and 10-trillion-plus parameter agentic and reasoning models. Running a model with more than 10 trillion parameters requires dozens of GPUs, high-speed VRAM and a low-latency interconnect to split model weights effectively, with Ethernet serving as the preferred networking foundation.

Helios uses three layers: a third-generation Pensando “Salina” DPU on the front end, a first-generation UALink-over-Ethernet (UALoE) scale-up fabric inside the rack and a second-generation Pensando “Vulcano” AI NIC for scale-out. AMD’s Helios management software and Fabric Manager oversee the stack, as bandwidth needs are doubling in less than two years and data fabric intelligence becomes increasingly important alongside compute performance.