AMD has officially stepped into the next stage of AI networking with its Pensando Pollara 400 AI NIC, unveiled at Hot Chips 2025. 
This card marks a milestone as the industry’s first UEC-ready (Ultra Ethernet Consortium) solution, designed specifically to meet the crushing bandwidth demands of AI and HPC systems.
At its core, the Pollara 400 delivers 400GbE speeds, putting it on par with NVIDIA’s ConnectX-7. While NVIDIA’s ConnectX-8 paired with Blackwell Ultra already doubles that at 800GbE, AMD’s approach isn’t just about raw speed. It’s about rethinking how AI networks operate under extreme loads.
The NIC is built around a P4 programmable architecture, allowing flexibility in how data packets are handled. Its pipeline features specialized units like the Match Processing Unit for efficient field manipulation and new enhancements such as virtual-to-physical address translation and atomic memory operations near SRAM. These tweaks aim to cut job completion times and improve cache coherency while maintaining high availability.
AMD highlights the real-world challenges of AI-scale networks: congestion, packet loss, and wasted bandwidth from uneven load balancing. Unlike general-purpose systems, AI workloads often saturate networks to the max, leading to inefficiencies. By embracing UEC, an open and interoperable communications standard, AMD believes it can offer a 25–40% performance uplift compared to legacy RoCEv2 implementations.
The Pollara 400 slots neatly into AMD’s broader data center ecosystem, complementing EPYC CPUs and Instinct accelerators, and connects directly through a Gen5 x16 PCIe interface without requiring an onboard switch. In short, it’s not just another NIC – it’s a piece of AMD’s bid to build a full-stack AI data center alternative to NVIDIA’s dominance.