AMD and Meta Unveil MetaRoCE for AI-Scale Ethernet

Meta and AMD have unveiled MetaRoCE, a new RDMA transport protocol designed to tackle the unique networking challenges posed by large-scale AI training and inference. Announced on August 24, 2026, MetaRoCE introduces a multipath data transport method, which optimizes network resource utilization and reduces congestion in AI environments spanning thousands of GPUs. The protocol leverages AMD’s programmable AI NICs, a key enabler for its rapid development and deployment.

Traditional Ethernet protocols, such as RoCEv2, often struggle to meet the demands of massive AI clusters. These systems require consistent, high-throughput communication across an ever-growing number of accelerators. MetaRoCE addresses these issues by moving transport intelligence to the network endpoints. The protocol dynamically distributes traffic across multiple paths, adapts to real-time network conditions, and implements selective retransmission for lost packets, ensuring greater resiliency and efficiency.

How MetaRoCE Changes the Game

MetaRoCE’s multipath transport capability is a significant departure from conventional single-path approaches. By allowing data to flow across multiple network paths simultaneously, it minimizes congestion and boosts bandwidth utilization. This is particularly impactful in AI clusters, where job completion times are sensitive to network performance bottlenecks.

The protocol also introduces a receiver-driven mechanism to manage congestion. Instead of relying solely on the network fabric, the receiving NIC signals a target data rate to the sender, which caps transmission and schedules traffic accordingly. This method improves flow control and reduces the likelihood of network-wide disruptions.

Meta’s tests on a 64-node AMD GPU cluster showed MetaRoCE outperforming RoCEv2, achieving 86% throughput even under 1% packet loss conditions. The protocol also scaled linearly across 4-plane and 8-plane topologies with up to 4,000 concurrent connections, highlighting its robustness in handling large-scale AI workloads.

AMD’s Role: Accelerating Innovation

AMD’s programmable Pensando AI NICs have been instrumental in the development of MetaRoCE. Initial testing was conducted on the Pensando Pollara 400 AI NIC before transitioning to the higher-bandwidth Vulcano 800 AI NIC. The programmable nature of these NICs allowed Meta to iterate on protocol design in software, avoiding the delays typically associated with hardware refresh cycles.

This flexibility not only expedited the transition from proof of concept to production but also provides a path for future enhancements. Updates to MetaRoCE or new transport protocols can be deployed via software updates, preserving existing hardware investments while keeping pace with evolving AI demands.

Broader Implications for AI Networking

Meta plans to release the MetaRoCE specification, a software reference implementation, and a compliance framework through the Open Compute Project at the OCP Global Summit in October 2026. This move aims to foster collaboration across the industry and accelerate adoption of the protocol.

The introduction of MetaRoCE underscores the growing need for adaptive, programmable networking solutions in AI infrastructure. As AI workloads and deployment models continue to evolve, infrastructure providers must embrace flexible designs that can accommodate rapid changes. AMD’s programmable NICs and MetaRoCE represent a forward-looking approach, enabling advancements without waiting for new hardware generations.

For the AI industry, this development highlights the importance of efficient, scalable networking as a cornerstone for future growth. With MetaRoCE, AMD and Meta are setting a precedent for how intelligent networking can drive progress in AI at scale.

Image source: Shutterstock