Oracle's OCCL plugin lifts NCCL collective bandwidth up to 14%
Oracle released Oracle Cloud Infrastructure Collective Communications Library (OCCL), a network plugin that improved NCCL AllReduce bandwidth by up to 10% and AllToAll by 14% on 16–128 node GB300 clusters, deployed atop OCI's multiplanar RDMA architecture with in-transport congestion control and NIC failure recovery.
Why it matters OCCL provides production-tested collective communications offload for OCI GPU clusters at scale, reducing network-induced delays in multi-node all-reduce patterns; multiplanar redundancy and transport-layer congestion control narrow the performance gap between OCI Spectrum-X Ethernet and proprietary NVIDIA Quantum InfiniBand on NCCL workloads.
Oracle announced Oracle Acceleron enhancements and OCCL at Oracle AI World 2026 (Nov 4), releasing a network plugin designed to run on OCI's multiplanar RDMA infrastructure. OCCL provides application-level congestion control and failure recovery at the collective-communications transport layer, enabling frameworks to continue using standard NCCL or RCCL APIs while OCI's plugin handles multi-plane network intelligence. Benchmarks using NVIDIA GB300 systems across 16–128 nodes showed 10% bandwidth gains for AllReduce (parameter synchronization in distributed training) and 14% for AllToAll (gradient exchange). OCI deployed 850 MW of data center capacity and 300,000+ GPUs in fiscal Q1 2027. For practitioners on OCI, OCCL now provides multi-plane redundancy with in-transport congestion management on collective operations at scale, reducing single-plane saturation from 100-node all-reduces and simplifying tuning of lossless Ethernet (PFC, ECN) across multiple independent network planes.
Read the original at sdxcentral.com ↗