# AMD Positions Programmable AI NICs for Multiple Ethernet Transports

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/31377
Published: 2026-10-13T17:14:26.000Z
Updated: 2026-10-13T17:14:26.000Z
Section: Technology

> The Pensando Pollara 400 supports up to 400 Gbps, while the Vulcano 800 reaches up to 800 Gbps of Ethernet connectivity.

AMD is positioning its Pensando Pollara 400 and Vulcano 800 AI network interface cards as programmable platforms for adapting Ethernet transports as large AI clusters demand faster, more resilient accelerator-to-accelerator communication.

The Pollara 400 uses a fully programmable hardware pipeline based on P4. That design supports RoCEv2, Ultra Ethernet Consortium (UEC) features and custom transport protocols without requiring a replacement networking platform. AMD lists speeds of up to 400 Gbps for the card.

The architecture also includes packet spraying, selective retransmission, path-aware congestion control and rapid fault detection. These functions target AI training systems in which accelerators exchange data continuously across multiple network paths.

AMD frames conventional Ethernet transports such as RoCEv2 as poorly suited to the congestion management, failure recovery and multipath traffic requirements of large AI clusters. The company’s work on Multipath Reliable Connection, or MRC, includes packet-spray load balancing, path-failure detection, traffic rerouting and recovery.

MRC development involved OpenAI, Microsoft and other industry participants. The Pollara 400 was validated in AMD laboratories with AMD Instinct MI350 and MI355 GPU clusters and co-validated with OpenAI.

MetaRoCE was first implemented on the Pollara 400 before the transport was deployed on the higher-bandwidth Pensando Vulcano 800 AI NIC. MetaRoCE uses multiple network paths, moves congestion-control functions to the endpoint and retransmits only missing packets.

The Vulcano 800 provides up to 800 Gbps of Ethernet connectivity. A configuration using three NICs per GPU can deliver up to 2.4 Tbps of scale-out bandwidth per GPU.

AMD is positioning the architecture as a way for customers to adapt networking functions as AI transport standards develop. The company’s stated path is to develop and validate transport logic on the Pollara 400 and carry that work forward to the Vulcano 800 platform.

The networking approach complements AMD’s broader effort to build integrated AI infrastructure around its accelerators and Ethernet-based systems. The products are focused on the network layer, where congestion control, path recovery and bandwidth determine how efficiently large clusters operate.