Aviz Networks outlines ONES 3.0 for AI fabric management
Aviz Networks introduced ONES 3.0, a centralized management platform that consolidates switch, Network Interface Controller (NIC) and Graphics Processing Unit (GPU) telemetry and controls for multi-vendor Artificial Intelligence (AI) data centers to support Remote Direct Memory Access (DMA) (RDMA) traffic and AI-oriented topologies.
Research overview
Recent vendor analysis highlights growing demand for networking switches and GPUs as workloads shift toward generative and other AI models, increasing interest in open network operating systems and multi-vendor hardware stacks.
The material frames this trend as a driver for new operational models in data centers, where operators seek vendor-agnostic control and greater visibility across compute and network elements.
Technical breakdown
The vendor brief describes GPU-centric deployments that prefer topologies such as fat-tree, dragonfly and butterfly to meet latency and throughput requirements, and identifies RDMA over Converged Ethernet (RoCE) as a common method to handle high-bandwidth flows.
It also highlights requirements for lossless transport, low entropy traffic patterns, Quality of Service (QoS) profiles and mechanisms such as priority flow control to maintain stable Remote Direct Memory Access (RDMA) sessions across large GPU clusters.
Product update
ONES 3.0 is presented as a platform that inventories and visualizes switches, NICs and GPUs, orchestrates topology-specific configurations, and centralizes telemetry for multi-vendor fabrics.
The documentation notes support for RoCE/RDMA traffic patterns, automation of QoS profiles, Power Factor Correction (PFC) monitoring and deeper compute-to-network correlation to aid troubleshooting and policy enforcement.
Operational impact
The brief states centralized management can reduce manual configuration work by providing unified policy enforcement and consistent device settings across vendors, which may shorten time to diagnose RDMA packet drops and other fabric faults.
It also describes benefits from correlated observability across GPUs and network elements, enabling faster identification of hotspots, microbursts and QoS misconfigurations in AI clusters.
Implementation challenges
The vendor identifies interoperability, scalability, configuration simplicity and comprehensive monitoring as core obstacles when deploying a single-pane management tool in heterogeneous AI fabrics.
Addressing those areas requires vendor-agnostic interfaces, scalable telemetry pipelines and orchestration capabilities that map traffic classes and policies consistently across diverse topologies.
The brief concludes that centralized management addresses orchestration, telemetry and RDMA requirements for multi-vendor AI fabrics and is relevant for enterprise infrastructure and security leaders evaluating AI Data Center Operations (DCO). This “Blog Signals brief” is a fact-based summary of the vendor blog.