Skip to content

Omniverse Application Platform on Proxmox VE

TL;DR

  • Goal: Evaluate a Proxmox VE-based platform design for running NVIDIA Omniverse workloads using a hybrid model: Windows vWS for real-time interactive use and Kubernetes for containerized batch rendering.
  • Result: A research-oriented PoC architecture was defined, including two execution pools (Windows Pool and K8s Pool) and a clear vGPU allocation model (GPU 0 -> win11-pro / 96Q, GPU 1 -> k8s-wk-01 / 96Q).
  • Key takeaways:
  • Proxmox VE is a viable platform layer for mixed Omniverse workload patterns (interactive + batch) in a lab/PoC context.
  • Separating platform architecture and vGPU allocation mapping into two diagrams significantly improves readability.
  • Dedicated vGPU profile assignment per workload VM makes resource ownership and troubleshooting clearer during validation.

Status

  • Current status: Internal research / paused follow-up
  • Customer engagement: No recent follow-up from the customer during the observation period (2026-02-04 to 2026-02-21).
  • Document intent: This page is maintained as an internal technical research note and reference for future Omniverse/PVE discussions.

Architecture

Platform Architecture

flowchart TB
    host["Physical Host<br/>Dell PowerEdge R770<br/>(2x RTX PRO 6000 Blackwell)"]
    pve["Platform Layer<br/>Proxmox VE (PVE)<br/>(KVM / QEMU + NVIDIA vGPU Manager)"]

    host --> pve

    subgraph POOLS["Execution Pools"]
        direction TB

        subgraph K8S["K8s Pool"]
            direction TB
            cp["VM: k8s-cp-01<br/>K8s Control Plane"]
            wk["VM: k8s-wk-01<br/>K8s Worker Node"]
        end

        subgraph WIN["Windows Pool"]
            direction TB
            win11["VM: win11-pro<br/>vWS (Virtual Workstation)"]
        end
    end

    pve --> POOLS

vGPU Allocation and Workload Mapping

flowchart LR
    subgraph PHY["Physical GPU Resources"]
        direction TB
        g0["GPU 0<br/>RTX PRO 6000 Blackwell"]
        g1["GPU 1<br/>RTX PRO 6000 Blackwell"]
    end

    subgraph PROF["vGPU Profiles"]
        direction TB
        p0["Profile: 96Q"]
        p1["Profile: 96Q"]
    end

    subgraph VM["Assigned VMs"]
        direction TB
        win11["win11-pro<br/>vWS"]
        wk["k8s-wk-01<br/>K8s Worker Node"]
    end

    subgraph USE["Workload Purpose"]
        direction TB
        winUse["Omniverse Real-time<br/>Interactive Design"]
        wkUse["Omniverse Containers<br/>Batch Rendering"]
    end

    g0 --> p0 --> win11 --> winUse
    g1 --> p1 --> wk --> wkUse

Research Notes

  • Focused on architecture feasibility and resource mapping, rather than production rollout or customer delivery.
  • Windows Pool (vWS) is positioned for interactive Omniverse usage (real-time / design-oriented workflows).
  • K8s Pool is positioned for containerized batch rendering workflows.
  • vGPU mapping model documented in this case:
  • GPU 0 -> win11-pro (vWS), profile 96Q
  • GPU 1 -> k8s-wk-01, profile 96Q

Next Steps (If Re-engaged)

  • Validate runtime behavior and stability under actual Omniverse workload execution.
  • Confirm workload requirements (interactive vs batch priority) and GPU profile sizing.
  • Define production scope (HA, operations model, monitoring, and support boundaries) before implementation.