MetaRoCE: A New RDMA Transport Built for AI-Scale Ethernet (engineering.fb.com)

Meta has unveiled MetaRoCE, a groundbreaking RDMA transport protocol specifically designed for AI workloads on commodity Ethernet. This new protocol enhances performance and reliability for distributed AI systems that depend on seamless communication between vast networks of GPUs. Meta is releasing the MetaRoCE specification alongside software implementations and compliance testing tools through the Open Compute Project (OCP), encouraging broader industry adoption and innovation. MetaRoCE significantly improves network efficiency by moving intelligence from the Ethernet fabric to endpoints, enabling out-of-order packet delivery and native multipathing capabilities. This architecture simplifies traffic management and congestion control while promoting resilience to packet loss, ensuring high throughput and low latency even under challenging conditions. Additionally, the protocol allows for scalable connections without the complexity of traditional RDMA configurations. As organizations seek to optimize AI infrastructure, MetaRoCE's open design aims to meet growing demands for performance, facilitating collaborative development across diverse hardware environments. With its robust capabilities, MetaRoCE positions itself as a critical advancement for AI infrastructure, promising to streamline operations for training and serving the next generation of AI models.