Source
MetaRoCE: A New RDMA Transport Built for AI-Scale Ethernet
Meta has developed MetaRoCE, an open RDMA transport protocol purpose-built for AI workloads on commodity Ethernet, and is releasing its specification through the Open Compute Project.
Why it matters
Improving data movement between GPUs reduces AI response times for millions of users and prevents expensive compute capacity from being stranded during training.
The details
- Meta is releasing the specification, reference implementation, and compliance suite via OCP.
- Tests on AMD GPU clusters showed ~86% throughput at 1% packet loss.
- The protocol enables native out-of-order delivery and native multipathing.
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.