Aviz Networks outlines ONES Fabric Manager for NVIDIA GB200 and GB300

NVIDIA GB200 and GB300 NVL72 systems use NVLink and InfiniBand fabrics to support multi-tenant GPU environments, while ONES Fabric Manager provides a single control plane for partitioning and isolation across single-rack and multi-rack deployments.

Research Overview

The blog describes rack-scale AI factory infrastructure built around 72 Blackwell GPUs in GB200 and GB300 NVL72 systems. It says NVLink handles scale-up traffic inside a rack, while InfiniBand manages scale-out traffic between racks.

It also states that standard VLAN-based isolation is not applicable because NVLink is not an Ethernet network. As a result, isolation has to be applied at the fabric layer.

Key Findings

ONES Fabric Manager is presented as a unified management plane for GPU allocation and fabric isolation. The blog says it identifies topology, applies NVLink partitions within racks, and coordinates cross-rack InfiniBand partitioning when a tenant spans multiple racks.

The post says provisioning is atomic, so requests either complete fully or roll back to the prior state. It also says the platform maintains live inventory and coordinates with NVIDIA UFM for cross-rack use cases.

Technical Breakdown

For tenants that remain within one rack, ONES Fabric Manager applies hardware-level GPU partitions only at the NVLink layer. In that case, the InfiniBand fabric is not involved in isolation.

When a tenant spans more than one rack, the platform adds a private InfiniBand partition through UFM alongside the per-rack NVLink partitions. The blog says this keeps tenant traffic separated across both fabrics.

Operational behavior

The blog says the system is designed so operators do not need to know in advance whether a tenant will occupy one rack or several. They specify the target nodes and GPUs, and the platform determines the needed isolation steps automatically.

It also says failed requests do not leave partial configurations behind. The post describes fault containment, automatic recovery, and live reconciliation between platform state and fabric state.

Product Update

The post says ONES Fabric Manager supports NVIDIA GB200 and GB300 rack-scale topologies, including NVL72, NVL36, dual-plane scale-out fabrics, and tray- and GPU-level allocations. It is described as using one API for both single-rack and multi-rack provisioning.

Figures in the blog show separate tenants placed on separate racks as well as a tenant spanning two racks. In both cases, the platform keeps the configured GPU resources isolated and reports completed provisioning.

Operational Impact

For operators, the main change is that multi-tenant GPU sharing does not require separate manual workflows for rack-contained and cross-rack assignments. The blog says the same control plane handles both cases and automatically adds cross-rack fencing only when needed.

The post also emphasizes self-contained racks, fail-fast request handling, and automatic restoration of failed controller connections. It presents these functions as part of day-to-day management of shared GPU infrastructure.

Overall, the blog explains how ONES Fabric Manager coordinates GPU partitioning and fabric isolation for NVIDIA GB200 and GB300 NVL72 systems across single-rack and multi-rack deployments. This Blog Signals brief is a fact-based summary of the vendor blog.

Blog post, originally published by Nikhil Morey at aviznetworks.com.