Oracle adds AI networking upgrades to Acceleron

Oracle has added new networking features to Oracle Acceleron for artificial intelligence workloads, expanding its cloud infrastructure offering for data movement, distributed AI communication and multi-tenant cluster security.

The additions sit within Oracle Cloud Infrastructure and focus on how data moves between compute and storage systems, how AI workloads respond to network faults, and how shared environments are isolated. Oracle said the changes build on its investments in network design, host-level acceleration and security.

Among the new features is Oracle Acceleron High-Performance Fabric, which brings RDMA connectivity to front-end compute for inference, storage and high-performance computing workloads. Oracle said it is intended to improve performance for customers running data-intensive applications.

Oracle has also widened NVMe block storage attachment options across Acceleron bare metal and virtual machine shapes. The change can improve performance by as much as three times for AI, database and other data-heavy workloads, according to Oracle.

Network resilience

A separate set of changes targets distributed AI communication, where large clusters can be affected by a single failed or congested connection. Oracle introduced a Collective Communications Library that spreads traffic across its backend multiplanar network and redirects data to healthier paths when failures occur.

The library is designed to help workloads recover immediately when links, ports or network interface cards fail, the company said, with the aim of improving availability and resilience for AI jobs spread across large numbers of accelerators.

Another feature, Transparent Front-End Multiplanar Networking, extends multiplanar networking to front-end networks. It includes automatic path selection, failover and congestion control for front-end RDMA traffic across compute infrastructure, Oracle said.

These changes come as cloud providers compete to attract demand for large-scale AI training and inference, which place heavy pressure on the networks linking processors, storage and servers. Operators are under growing pressure to reduce downtime and bottlenecks because AI systems often depend on thousands of accelerators working together.

Oracle said Acceleron combines dedicated network fabrics, more direct data paths, and a converged network interface card with host-level zero-trust routing and a multi-planar design. It added that the software architecture is designed to unify networking, storage and security functions across different types of hardware.

Pradeep Vincent, Executive Vice President of Compute and AI Infrastructure at Oracle Cloud Infrastructure, described network reliability as a central issue for large AI systems.

"AI infrastructure can be unforgiving when it comes to network friction," Vincent said.

"When thousands of accelerators are working together, one congested or failed path can have consequences far beyond a single connection. With Oracle Acceleron, we are building a network that can respond by rerouting traffic, recovering from failures and helping customers scale securely," he added.

The announcement reflects a broader shift in cloud infrastructure spending towards specialised networking and storage components that support AI training and inference. As model sizes and data volumes grow, vendors are placing more emphasis on network architecture as a differentiator alongside compute hardware.

For Oracle, the latest Acceleron additions deepen its push to position OCI for customers running complex AI workloads across shared and dedicated infrastructure. The new functions are intended to give customers more flexibility across compute and storage while allowing the network to adapt to changing conditions, according to the company.