First Principles: Oracle Acceleron Networking for AI at Scale

Oracle Acceleron networking innovations help Oracle Cloud Infrastructure support high performance AI workloads with RDMA congestion control, path awareness, and resilient multiplanar networking. See how these capabilities address congestion, load balancing, and network failures while maintaining application compatibility at https://social.ora.cl/6003BGFqON Oracle Acceleron networking innovations address a fundamental challenge for AI training and other high performance workloads: keeping network traffic moving efficiently when congestion, failures, and large-scale deployments collide. In this episode of First Principles, Pradeep Vincent is joined by OCI Networking architects Jag Brar and Joy Ganguly for a technical discussion of how Oracle Cloud Infrastructure approaches network performance, reliability, resilience, and scale. The conversation explores how OCI extends remote direct memory access (RDMA) beyond dedicated cluster networks to a shared front-end fabric that carries both RDMA and non-RDMA traffic. Learn how hierarchical virtual queues help maintain fairness between traffic classes and respond to incast, and how one-way delay measurements help identify fabric congestion. The speakers also explain how Oracle Acceleron SmartNICs can probe alternate network paths and signal senders to slow down when congestion persists. The episode also examines multiplanar networking and the Oracle collective communication library plugin, including how intelligence in the collective communication layer can distribute traffic across available paths. See how the plugin detects congestion, shifts traffic toward less-congested paths, handles port, NIC, and plane failures, and supports recovery while remaining transparent to applications. Together, these networking capabilities are designed to support performance and resilience for AI, database, storage, and cloud workloads at scale. 00:00 Introduction: Networking at AI Scale 01:16 Two Key Networking Innovations 02:39 Bringing RDMA to a Shared Front-End Network 03:53 Fairness Between RDMA and TCP Traffic 05:47 Preventing Incast Congestion 06:33 Congestion Control Using One-Way Delay 07:37 SmartNIC Path Selection and Congestion Response 08:49 Application Compatibility and RDMA Use Cases 10:07 Why Multiplanar Networks Matter 11:02 Introducing Oracle Collective Communication Library (OCCL) 11:55 Load Balancing Across Network Planes 13:22 Simplifying Multiplanar Networking for Customers 15:15 Why Put Intelligence in the CCL Plugin? 16:02 Handling Link, Port, and NIC Failures 17:34 Transparent Plane Failover and Recovery 19:04 Congestion-Aware Traffic Balancing 20:13 Future Performance and Reliability Improvements 21:03 Bringing the Innovations Together 22:02 Key Takeaways Like what you learned? Subscribe now https://social.ora.cl/60025k87j Join us at Oracle AI World Vegas: https://social.ora.cl/60035eVvJ Contact OCI sales: https://social.ora.cl/6007cxqVc Learn more about OCI: https://social.ora.cl/60055GhEx Try OCI for free: https://social.ora.cl/6004CEo6q Follow us here: LinkedIn: https://social.ora.cl/60035eXuF X: https://social.ora.cl/6008abVgG Instagram: https://social.ora.cl/60095eXDN Facebook: https://social.ora.cl/60065enQp