AMD details networking stack for large-scale AI infrastructure

AMD describes AI networking as the high-performance fabric that connects GPUs, accelerators and systems while supporting the throughput, synchronization and performance consistency required for training, inference and agentic AI workloads. The company emphasizes high bandwidth, low latency, programmability and open systems as core requirements for keeping accelerators utilized as workloads grow.

The portfolio is organized around three infrastructure layers. Front-end networking connects users and data to AI infrastructure while accelerating storage, security and networking services through AMD Pensando™ Salina DPU and AMD Pensando™ Vulcano 800 AI NIC technology. Scale-up networking uses UALoE and tightly coupled software to make many GPUs operate as one compute resource, with AMD saying the architecture unifies 72 GPUs and adds alternate connections, automated fault response and workload isolation. Scale-out networking targets distributed AI across racks, data centers and regions using technologies including Multipath Reliable Connection, MetaRoCE and UEC-ready RDMA.

AMD also highlights the Helios Rackscale solution, which combines AMD Instinct™ GPUs, AMD EPYC™ Server CPUs and AMD Pensando™ networking using open industry standards. Helios Management Software supports fabric health operations through monitoring, telemetry, RAS and automation, while the broader software stack includes tools for cluster operations, GPU partitioning, isolation and switch reliability.