AMD is advancing its AI strategy with the introduction of the Pensando "Salina" DPU and "Vulcano" AI NIC for its Helios networking solution. Networking is a key pillar of AMD's Advancing AI 2026 rack-scale strategy. The company detailed its Pensando silicon and fabric technology designed to interconnect numerous GPUs, memories, and manage data movement speeds within a single system and across multiple racks.

AI model sizes have grown exponentially, requiring high-speed interconnects to split model weights effectively and minimize latency. AMD's solution employs a three-layer approach: the third-generation Pensando "Salina" DPU for front-end networking at 400G, a first-generation UALink-over-Ethernet (UALo E) fabric for scale-up within the rack, and the second-generation Pensando "Vulcano" AI NIC for scale-out connections. The Salina DPU offers twice the performance of its predecessor and is reportedly used by hyperscalers like Oracle and Microsoft.

The UALo E fabric connects 72 GPUs into a single domain, providing 260 TB/s of bandwidth and a unified memory pool. For scale-out, the "Vulcano" AI NIC operates at 800G, delivering 50% higher bandwidth than competing solutions and reducing costs by a third. This custom solution aims to improve LLM training efficiency and allow reallocation of funds towards more compute infrastructure.