Oracle Acceleron adds multiplanar RDMA networking for AI clusters

What is changing?

Oracle announced new OCI Acceleron capabilities on October 27 aimed at front-end networking and storage connectivity for AI clusters. High-Performance Fabric now supports front-end RDMA, while expanded NVMe attachment options are available across bare-metal systems and virtual machines. Oracle says the new option can deliver up to three times higher storage performance in some configurations; that figure is a provider claim.

The new Oracle Collective Communications Library (OCCL) distributes collective-communication traffic across multiple queue pairs on a multiplanar network. It can favor healthier, higher-throughput paths and is designed to shift traffic after a link or network-interface failure without restarting an AI job. Transparent front-end multiplanar networking also reduces the need for applications to manage that topology directly.

Hierarchical token-bucket controls on the SmartNIC provide fair sharing between RDMA and TCP traffic, while one-way-delay congestion control detects busy paths. For production teams, the decisive measures will be independent workload tests, regional availability and actual recovery time during failures.