Meta AI Introduces MetaRoCE: A Clean-Sheet RDMA Transport Built for AI-Scale Ethernet
Training and serving frontier models is now a networking problem as much as a compute problem. Collective operations like all-reduce and all-to-all synchronize thousands of accelerators during training, and the slowest transfer sets the pace for the entire job. Even small amounts of network friction directly strand significant compute capacity. This week, Meta introduced MetaRoCE. It is described as a clean-sheet RDMA transport protocol purpose-built for AI workloads on commodity Ethernet.
Make your prediction
Will Meta publish a technical paper or blog post on MetaRoCE by September 5, 2026?
Resolves by Sep 5, 2026
Your prediction
50% · 50/50 coin flip
NOYES
Get smart on it
Training large AI models requires synchronizing thousands of accelerators across networks, and network performance directly limits training speed. Meta introduced MetaRoCE, a new network transport protocol designed for AI workloads that treats Ethernet as lossy and moves ordering and recovery intelligence to network interface cards rather than relying on network switches to deliver packets in order. The protocol was tested on a GPU cluster and maintained approximately 86% throughput at 1% packet loss, compared to standard approaches that collapse under such conditions. Meta is releasing the specification, software implementation, and compliance tests through the Open Compute Project, with hardware support beginning on programmable network interface cards.
Related news
DOE Keeps Eddystone Power Plant Online Amid Data Center Demand Surge
The 760 MW Eddystone coal plant will remain online longer as PJM faces rising data center load, generator retirements, and delayed power resources.
Liquid Cooling Options: RDHx, Direct-to-Chip, Immersion
AI workloads push data center rack densities beyond the limits of air cooling. Three liquid cooling approaches offer different trade-offs and benefits.
