How production fabrics balance RoCE today
AI training traffic is a handful of huge RDMA flows between GPUs, and ECMP was built for millions of small ones. Every operator running RoCE at scale has had to deal with that mismatch. The fixes in real use today come down to one question: who gets to choose the path? Read them left to right and the decision moves out of the switch's hash and into the NIC that sends the traffic.
