Meta's custom transport protocol embraces packet chaos to boost AI throughput (www.sdxcentral.com)

Meta has announced the development of MetaRoCE, a custom transport protocol aimed at enhancing throughput for AI workloads over Ethernet-based networks. Unlike traditional high-performance protocols that send packets in a sequential order, MetaRoCE intentionally disperses packets out of order, allowing for more efficient network utilization and higher data throughput. This design enables data to be written directly to its final memory location upon arrival, mitigating delays caused by packet loss, and significantly improves performance during adverse network conditions. The significance of MetaRoCE lies in its potential to transform how data center networks handle AI processing, particularly in scale-out environments. By decentralizing intelligence to the network endpoints and reducing reliance on switches for packet order, the protocol enhances throughput and lowers flow completion times, even in scenarios with high packet loss. Tested on AMD hardware, MetaRoCE demonstrated strong performance across various network topologies, maintaining 86% throughput under conditions that typically challenge existing protocols. Meta plans to share the full protocol specifications with the Open Compute Project, encouraging collaboration across the industry to further refine AI infrastructure and improve system efficiency.