Aviz ONES 4.2 details decoupled compute sharing and tenant-bound GPU isolation

Aviz ONES 4.2 separates compute VM sharing from GPU device ownership, letting multiple tenants run on the same server while each tenant’s GPUs stay strictly isolated. For enterprise security and infrastructure teams, the change targets higher utilization and lower cost within tenant boundaries.

Research Overview

The vendor describes GPU fabrics where single-tenant server allocation keeps compute, network, and GPU resources bound together for isolation. It frames ONES 4.2 as a shift away from that rigid coupling by decoupling compute and GPU assignment.

The blog positions the update for environments such as shared AI platforms and PaaS, where workloads can change and tenant mixes may be heterogeneous. It links the model to improved utilization and workload density while maintaining isolation for GPU workloads.

Key Findings

The ONES 4.2 model allows multiple tenants to share compute-only virtual machines on the same physical server. It also states that cross-tenant GPU sharing is not permitted and that GPU ownership is bound to a single tenant.

According to the post, the approach addresses inefficiencies from server-level isolation, including underutilized CPU and memory, partially idle GPUs, limited workload placement flexibility, and increased infrastructure cost driven by low density. The blog connects these issues to dynamic workload placement patterns common in AI and PaaS workloads.

Technical Breakdown

The update describes fine-grained GPU allocation managed at the individual GPU device level, with strict tenant binding. It states that GPUs remain isolated even when the compute layer is shared across tenants.

For networking, the blog says ONES uses EVPN VXLAN and VRF to segment tenant traffic across the fabric. It adds that VLAN tagging is used for tenant traffic segmentation, VLAN behavior is consistent during provisioning, and VLAN termination occurs at the leaf layer with mapping into tenant-specific routing domains.

Operational Impact

The post describes GPU-aware orchestration that validates server eligibility based on GPU allocation before placing workloads. It states the orchestration enforces separation between compute-only and GPU workloads and ensures placement adheres to tenant boundaries.

It also describes the operational outcome as higher resource utilization, increased workload density, reduced costs, and consistent tenant isolation on shared infrastructure. The blog ties cost reduction to sharing overhead from CPU and memory when multiple tenants run compute workloads on the same physical server.

Leadership Perspective

The blog presents ONES 4.2 as a model for scalable GPU infrastructure that removes constraints tied to rigid, server-level isolation. It frames the balance as enabling compute sharing while maintaining strict GPU ownership per tenant.

For enterprise leaders, the post emphasizes that operators can move compute workloads dynamically as demand shifts without changing GPU assignments. It links that capability to predictable performance and the ability to support high-density multi-tenant GPU environments for AI and PaaS workloads.

Aviz ONES 4.2 introduces a decoupled approach where compute VMs can be shared across tenants while GPU devices remain assigned to a single tenant under strict isolation controls. The blog emphasizes higher utilization and workload density through GPU-aware orchestration and EVPN VXLAN with VRF-based tenant segmentation, stating this “Blog Signals brief” is a fact-based summary of the vendor blog.