Oracle releases network plugin to help AI training survive connection failures

Oracle released its OCCL network plugin on October 26, 2026, giving supported cloud customers a new way to route AI training traffic around congestion and connection failures. The generally available software carries no additional charge. Oracle says it works beneath existing training applications, avoiding changes to the code that coordinates work across GPUs.

Distributed training repeatedly exchanges model parameters, gradients and other data among GPUs. These exchanges, called collective communication, can leave a training step waiting for its slowest network path. A congested connection can reduce throughput and leave GPUs idle; a failed link or network interface card can stall a job or force a restart.

Rerouting traffic instead of rewriting training code

OCI can provide multiple independent network planes: separate paths that add capacity and help isolate failures. But having those paths does not ensure training traffic uses them evenly. Oracle says traffic can concentrate on some paths while others remain underused. OCCL monitors transfer behavior and dynamically assigns communication data to less-congested available paths.

For supported failures, Oracle says the plugin can redirect traffic to a surviving link or network interface card, then resume using the original connection after it is restored. That recovery depends on an alternate path being available. The plugin supports both single-plane and multiple-plane network designs.

The software sits beneath NVIDIA’s NCCL and AMD’s RCCL, the libraries that coordinate communication among GPUs. It uses their standard network-plugin interface rather than replacing those libraries. Supported configurations include RoCEv2 and InfiniBand, two network technologies used for remote direct memory access, or RDMA. Oracle also says the plugin does not require a specific network vendor’s hardware stack, within OCI’s supported equipment.

More communication bandwidth, not a training-speed promise

Oracle tested the plugin in September on NVIDIA GB300 systems, comparing the same hosts and network configuration with and without it. The tests covered 16, 32, 64 and 128 nodes, with four GPUs per node. Message sizes ranged from 1 GiB to 16 GiB, with 50 measured iterations per size.

Across the tested configurations, AllReduce bandwidth gains ranged from 2.00% to 10.11%, while AllToAll gains ranged from 7.33% to 14.55%. These are communication benchmark results, not measurements of complete training-job duration. Oracle averaged the in-place and out-of-place nccl-tests results and cautions that the gains depend on the tested hardware, software, network, message sizes and channel settings. They are not a performance commitment for every workload.

Preinstalled images or a custom-image route

Oracle offers two deployment paths. Supported Oracle HPC images include the plugin preinstalled; customers using custom images can download the open-source distribution from Oracle’s OCI GitHub repository and integrate it themselves. Neither route requires a separate control plane, fleet of agents or application-side service, according to Oracle.

export NCCL_NET_PLUGIN=/opt/occl-net/lib/liboccl-net-occl-ib.so
export NCCL_NET=IB

The no-additional-charge offer applies to supported OCI GPU and RDMA-enabled infrastructure, not to the underlying resources. Customers still pay standard GPU compute, networking, storage and applicable support charges. Oracle directs teams to check its supported configuration matrix and deployment guidance before enabling the plugin, including the supported software versions and tuning for their environment.