# Aviz Networks outlines RTX PRO AI Factory fabric design

URL: https://decisioninsights.ai/aviz-networks-outlines-rtx-pro-ai-factory-fabric-design/

Aviz Networks details how ONES generates networking for NVIDIA RTX PRO AI Factory servers, pairing a Spectrum-X compute fabric with a BlueField-3 front-end. The design matters to enterprise IT and network leaders deploying GPU infrastructure in air-cooled data centers.RTX PRO Server NetworkingEach RTX PRO server is an NVIDIA-Certified system with up to eight RTX PRO 6000 Blackwell Server Edition GPUs, each carrying 96 GB of GDDR7. The reference design uses four BlueField-3 SuperNICs for GPU-to-GPU traffic and one BlueField-3 DPU for users, orchestration and storage.NVIDIA labels configurations with notation such as 2-8-5-200, which covers CPU sockets, GPUs, NICs and Gbps per GPU. The 2-8-5-200 baseline provides 200 Gbps per GPU, while 400 Gbps configurations are also available, and NVIDIA recommends 400 Gbps for fine-tuning and image or video generation.East-West Compute NetworkThe compute network is a non-blocking leaf-spine fabric built on Spectrum-X Ethernet switches, with a scalable unit of 32 servers and two leaf switches. Each server's four compute NICs split across both leaves, so each leaf uses 64 400G downlinks and 64 400G uplinks at a 1:1 ratio.The article states that if one leaf fails, every server in the unit keeps half of its compute bandwidth. Scaling adds two leaves per unit, with spines added in pairs, and 32 units of 8-GPU servers total 8,192 GPUs.North-South Front-End NetworkEach server has a single BlueField-3 DPU for front-end traffic, so user and storage traffic share one path and are separated logically. The front-end fabric includes compute leaves, storage leaves, border leaves and a pair of front-end spines, with VXLAN used for tenant segmentation by default.Designing a Fabric in Aviz ONESONES models RTX PRO as a GPU node type within the Aviz RA 1.0 fabric and derives the NIC layout, scalable unit size, switch platform and cabling from that selection. Users choose 4 or 8 GPUs per server and set the number of scalable units along with a maximum for growth, which ONES uses to reserve spine capacity.ONES generates device names, addressing and links for both networks, with /31 point-to-point links for each GPU NIC. In the walkthrough's 2-SU example, the design includes 78 devices, 64 servers and 512 GPUs.Pre-Deployment Review and ValidationThe Topology tab presents the generated design per scalable unit and in east-west and north-south views before any switch is configured. Users can modify the YAML definition, and uploads are checked for issues such as missing reverse links, duplicate ports, orphaned devices and YAML syntax errors.Deployment then validates the switches and servers before pushing configuration, with tenant onboarding and Day-2 operations following through the ONES UI, APIs or Terraform.This Blog Summary is a fact-based summary of the vendor blog. The article presents RTX PRO AI Factory networking as a defined reference design that ONES translates into a reviewable, deployable fabric, a point relevant to enterprise decision-makers planning air-cooled GPU infrastructure.
