AMD Pensando Pollara 400 AI NIC

The AMD Pensando Pollara 400 is a purpose-built AI networking interface card designed to meet the extreme bandwidth and latency demands of large-scale GPU cluster interconnects. Leveraging AMD's Pensando DPU architecture, the Pollara 400 delivers 400Gbps of aggregate throughput per adapter while supporting the emerging Ultra Ethernet Consortium (UEC) specification, positioning it as a next-generation alternative to InfiniBand for AI fabric deployments. The card implements hardware-accelerated RDMA over Converged Ethernet (RoCEv2) with advanced congestion control, adaptive routing, and packet spray capabilities natively in silicon, enabling near wire-rate performance at scale without burdening host CPUs.

Ideal for

• Large-scale distributed deep learning training across thousands of GPUs in hyperscale AI clusters requiring low-latency, high-bandwidth all-to-all collective communication
• Generative AI and large language model inference serving farms where GPU-to-GPU memory transfers must sustain sub-microsecond latency to meet service-level objectives
• High-performance computing environments running MPI-based scientific simulations that benefit from hardware-offloaded RDMA and adaptive routing over Ethernet fabric
• Cloud service provider GPU-as-a-service infrastructure deployments seeking an open, Ultra Ethernet Consortium-compliant fabric to reduce dependency on proprietary interconnect ecosystems
• AI supercomputing clusters requiring congestion-free packet delivery with hardware-based multipath and packet spray to maximize bisection bandwidth utilization
• Enterprise AI model training pipelines on converged Ethernet infrastructure where co-location of storage, compute, and AI traffic demands a single high-performance network fabric

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.