Oracle adds Acceleron networking controls to tackle AI traffic jams and failures

Oracle is adding networking controls designed to stop data movement from holding up AI workloads. Announced at Oracle AI World in Las Vegas on October 27, 2026, the Acceleron updates address two different problems: sharing network capacity with other cloud services and keeping distributed training running when connections become congested or fail.

Fast transfers without a private network

The first change concerns OCI’s front-end network, which carries a mix of database, storage, compute, internet and other traffic. Oracle’s High-Performance Fabric brings remote direct memory access, or RDMA, to that shared environment. RDMA moves data directly between machines’ memory with little CPU involvement, bypassing much of the operating-system processing involved in conventional transfers.

Oracle says the update extends front-end RDMA to inference—the work of running a model—storage and high-performance computing. Front-end links can also connect separate GPU clusters while their backend networks handle local training. That puts the emphasis on moving data into and between computing resources, rather than only exchanging it inside one training cluster.

Sharing is the difficult part. RDMA is sensitive to packet loss and congestion, while ordinary TCP traffic uses a different method to control its sending rate. Oracle’s technical explanation describes how one can overwhelm the other. Acceleron’s network cards track bandwidth through a shared allowance and separate allowances for RDMA and non-RDMA traffic. Either class can borrow unused capacity.

As those allowances run low, the cards signal congestion by marking compatible packets or dropping others. They also measure delay in the direction data is traveling, rather than using a round trip that could mistake congestion on the return path for trouble ahead. When delay rises, a card probes three alternative paths and switches if it finds a better one.

Training traffic gets a software escape route

The second layer addresses communication among GPUs working together. Synchronized training can wait on its slowest connection, leaving other processors idle. Oracle says conventional equal-cost multipath routing can distribute traffic unevenly because it assigns flows to paths without knowing how much bandwidth each flow needs. Multiple available routes do not necessarily mean balanced use.

Oracle’s Collective Communications Library plugin, OCCL, instead distributes messages and chunks of data according to observed throughput, congestion and connection health. It sends more traffic over better-performing connections and moves traffic away from congested paths. Oracle says it can also redirect communication to another available port or network interface card after a failure, allowing training to continue without restarting.

The distinction is where that intelligence lives. OCCL puts recovery and traffic balancing in the communication software layer, reducing dependence on specialized network hardware and firmware. Oracle says it works with both single-plane networks and designs with multiple independent fabrics. Its networking overview identifies support for NVIDIA’s NCCL and AMD’s RCCL plugin interfaces, which connect GPU communication libraries to the network.

Compatibility promises, but no rollout timetable

Oracle says front-end RDMA supports existing applications without changes, while OCCL is transparent to workloads using supported communication libraries. The announcement also broadens NVMe block-storage attachment options across Acceleron bare-metal servers and virtual machines. Oracle claims storage performance improvements of up to three times; that is a storage claim, not a measured threefold gain in AI training speed.

The October 27 announcement does not specify availability dates for the new capabilities. Its future-product disclaimer says release timing, pricing and functionality may change and should not be relied on for purchasing decisions. For customers evaluating the update, the architecture is described more concretely than the rollout: Oracle explains how the controls work, without setting a deployment schedule.