Building the Perfect Edge AI Supercomputer – Adding an Edge Virtualization Layer with Proxmox and GPU Passthrough

  • Blog
Views: 877

I built on my edge AI hardware by adding Proxmox VE as the virtualization layer. After prepping BIOS, using Rufus with the nomodeset trick, and installing Proxmox, I enabled IOMMU, configured VFIO, and passed through 2× RTX 3090 GPUs to a single Ubuntu VM. This setup lets me run private AI workloads at near bare-metal speed, while keeping Windows and native Ubuntu for special use cases.

In my previous post, Building the Perfect Edge AI Supercomputer – Cost Effective Hardware, I focused on the hardware. In this follow‑up I’ll show the exact steps I used to layer Proxmox VE on top and pass two RTX 3090 GPUs through to a single Ubuntu VM. The goal is a repeatable journey anyone can follow.

What You’ll Build

Host OS: Proxmox VE on a dedicated NVMe (ext4)

Triple‑boot plan: Windows 11 (video editing & the odd game with my boys), native Ubuntu (raw performance), and Proxmox (daily driver for labs).

AI VM: One big Ubuntu VM with 2× RTX 3090 via PCIe passthrough

Use case: Private, offline edge AI with maximum flexibility for lab VMs alongside

Tip: I prepared the installer USB with Rufus. During boot I used the nomodeset trick (press e at the boot menu, add nomodeset to the kernel line) so the machine wouldn’t hang on the GPU while booting the installer from USB.

Why Proxmox (and not Hyper-V/WSL2 or vSphere)?

I already run Proxmox on another server, and for 115 EUR/year the subscription feels like a fair deal for a bare-metal hypervisor that rivals vSphere (formerly ESXi, back when it was still free for home use).

Unfortunately, Broadcom’s new direction means VMware enthusiasts like me can’t really continue with vSphere the way we used to, and a VMUG membership still costs more.

A note to Broadcom/VMware: thanks for the amazing five years of employment and for shaping much of my understanding of enterprise virtualization.

Hyper-V? No thanks. I had enough of it during my time at Microsoft. And while WSL2* is technically an option, it’s built on Hyper-V and GPU passthrough might work, but my gut says performance won’t be on par. Maybe one day I’ll run benchmarks to confirm, but for now, Proxmox is the clear choice.

* What is Windows Subsystem for Linux | Microsoft Learn

A note to Microsoft: I’m equally grateful for my seven years there – five wonderful years with hosting and service providers, followed by two more. Even today, I still keep an eye on where it’s heading and occasionally use it in Windows 11 and Server 2025. Tools like Windows Admin Center and Azure ARC remain part of my toolbox.

1) Install Proxmox VE

  1. Create a USB installer with Rufus (GPT + UEFI).
  2. Boot the installer; if you get a black screen, press e and add nomodeset to the kernel args.
  3. Choose ext4 (simple & robust) and install Proxmox on a dedicated NVMe (e.g., /dev/nvme1n1).
  4. Give Proxmox a static management IP.
  5. Reboot into Proxmox.
Using Rufus to create boot USB stick for Proxmox VE
The first time I tried booting from the USB installer, the system froze. No panic – there’s always a solution. Kudos to AI for pointing out the root cause: the NVIDIA GPUs were loading drivers too early, which conflicts with the installer. Adding the classic nomodeset option to the boot parameters bypassed that issue and let the installation continue smoothly.
As mentioned earlier, just press e at the boot menu, append nomodeset to the kernel arguments, and the system will boot cleanly.
Install summary
Here’s my BIOS setup: I’m keeping it simple and using the UEFI boot priority list to start Proxmox for now. Later, I’ll add Proxmox to Ubuntu’s GRUB menu for a smoother boot experience.

Post‑install (shell)

apt update && apt -y full-upgrade

(Optional) enable the no‑subscription repo for updates if you don’t use a sub:

sed -i 's/^deb/#deb/g' /etc/apt/sources.list.d/pve-enterprise.list
cat >/etc/apt/sources.list.d/pve-no-subscription.list <<'EOF'
deb http://download.proxmox.com/debian/pve bookworm pve-no-subscription
EOF
apt update && apt -y upgrade

2) Enable IOMMU on the Host

Edit /etc/default/grub and append the right flag to GRUB_CMDLINE_LINUX_DEFAULT:

  • AMD: amd_iommu=on
  • Intel: intel_iommu=on

Example (AMD):

GRUB_CMDLINE_LINUX_DEFAULT="quiet amd_iommu=on"

update-grub

reboot

Verify after reboot:

root@pve-01:~# dmesg | egrep -i "iommu|amd-vi|intel-iommu"
[    0.000000] Command line: BOOT_IMAGE=/boot/vmlinuz-6.14.8-2-pve root=/dev/mapper/pve-root ro nomodeset quiet amd_iommu=on
[    0.000000] Kernel command line: BOOT_IMAGE=/boot/vmlinuz-6.14.8-2-pve root=/dev/mapper/pve-root ro nomodeset quiet amd_iommu=on
[    0.000000] AMD-Vi: Unknown option - 'on'
[    0.000000] AMD-Vi: Using global IVHD EFR:0x58f77ef22294ade, EFR2:0x0
[    0.267301] iommu: Default domain type: Translated
[    0.267301] iommu: DMA domain TLB invalidation policy: lazy mode
[    0.300624] pci 0000:00:00.2: AMD-Vi: IOMMU performance counters supported
[    0.300723] pci 0000:00:00.0: Adding to iommu group 0
[    0.300766] pci 0000:00:01.0: Adding to iommu group 1

... BLA BLA BLA ...

[    0.303223] pci 0000:35:00.0: Adding to iommu group 45
[    0.303249] pci 0000:36:00.0: Adding to iommu group 46
[    0.309178] AMD-Vi: Extended features (0x58f77ef22294ade, 0x0): PPR X2APIC NX GT IA GA PC GA_vAPIC
[    0.309195] AMD-Vi: Interrupt remapping enabled
[    0.309197] AMD-Vi: X2APIC enabled
[    0.309444] AMD-Vi: Virtual APIC enabled
[    0.310018] perf/amd_iommu: Detected AMD IOMMU #0 (2 banks, 4 counters/bank).
[    0.349456] iommu ivhd0: AMD-Vi: Event logged [INVALID_DEVICE_REQUEST device=0000:00:00.0 pasid=0x00000 address=0xfffffffdf8000000 flags=0x0a00]
root@pve-01:~#

You should see lines confirming IOMMU is enabled.

3) Load VFIO Modules & Keep NVIDIA Off the Host

Load VFIO at boot:

cat >/etc/modules <<'EOF'
vfio
vfio_pci
vfio_iommu_type1
vfio_virqfd
EOF

Blacklist host GPU drivers (so Proxmox won’t claim the cards):

cat >/etc/modprobe.d/blacklist-nvidia.conf <<'EOF'
blacklist nouveau
blacklist nvidia
blacklist nvidia_drm
blacklist nvidia_modeset
blacklist nvidiafb
EOF

Find your GPU devices and IDs:

root@pve-01:~# lspci | grep -i nvidia
31:00.0 VGA compatible controller: NVIDIA Corporation GA102 [GeForce RTX 3090] (rev a1)
31:00.1 Audio device: NVIDIA Corporation GA102 High Definition Audio Controller (rev a1)
32:00.0 VGA compatible controller: NVIDIA Corporation GA102 [GeForce RTX 3090] (rev a1)
32:00.1 Audio device: NVIDIA Corporation GA102 High Definition Audio Controller (rev a1)
root@pve-01:~# 
lspci -n -s 31:00.0
lspci -n -s 31:00.1
lspci -n -s 32:00.0
lspci -n -s 32:00.1
31:00.0 0300: 10de:2204 (rev a1)
31:00.1 0403: 10de:1aef (rev a1)
32:00.0 0300: 10de:2204 (rev a1)
32:00.1 0403: 10de:1aef (rev a1)

Bind those IDs to VFIO (both GPUs share the same IDs, so list them once):

cat >/etc/modprobe.d/vfio.conf <<'EOF'
options vfio-pci ids=10de:2204,10de:1aef
EOF

Update initramfs and reboot:

update-initramfs -u
reboot

Verify all four functions are using vfio-pci:

root@pve-01:~# 
lspci -k -s 31:00.0
lspci -k -s 31:00.1
lspci -k -s 32:00.0
lspci -k -s 32:00.1
31:00.0 VGA compatible controller: NVIDIA Corporation GA102 [GeForce RTX 3090] (rev a1)
        Subsystem: Micro-Star International Co., Ltd. [MSI] Device 3881
        Kernel driver in use: vfio-pci
        Kernel modules: nvidiafb, nouveau
31:00.1 Audio device: NVIDIA Corporation GA102 High Definition Audio Controller (rev a1)
        Subsystem: Micro-Star International Co., Ltd. [MSI] Device 3881
        Kernel driver in use: vfio-pci
        Kernel modules: snd_hda_intel
32:00.0 VGA compatible controller: NVIDIA Corporation GA102 [GeForce RTX 3090] (rev a1)
        Subsystem: Micro-Star International Co., Ltd. [MSI] Device 3881
        Kernel driver in use: vfio-pci
        Kernel modules: nvidiafb, nouveau
32:00.1 Audio device: NVIDIA Corporation GA102 High Definition Audio Controller (rev a1)
        Subsystem: Micro-Star International Co., Ltd. [MSI] Device 3881
        Kernel driver in use: vfio-pci
        Kernel modules: snd_hda_intel
root@pve-01:~#

(Optional) Check IOMMU Groups

If a GPU shares a group with other devices, you may need ACS override (last resort on a lab host):

root@pve-01:~# find /sys/kernel/iommu_groups/ -type l | sort
/sys/kernel/iommu_groups/0/devices/0000:00:00.0
/sys/kernel/iommu_groups/10/devices/0000:00:07.0
/sys/kernel/iommu_groups/11/devices/0000:00:07.1
/sys/kernel/iommu_groups/12/devices/0000:00:08.0
/sys/kernel/iommu_groups/13/devices/0000:00:08.1
/sys/kernel/iommu_groups/14/devices/0000:00:08.2
/sys/kernel/iommu_groups/15/devices/0000:00:08.3
/sys/kernel/iommu_groups/16/devices/0000:00:14.0
/sys/kernel/iommu_groups/16/devices/0000:00:14.3
/sys/kernel/iommu_groups/17/devices/0000:00:18.0
/sys/kernel/iommu_groups/17/devices/0000:00:18.1
/sys/kernel/iommu_groups/17/devices/0000:00:18.2
/sys/kernel/iommu_groups/17/devices/0000:00:18.3
/sys/kernel/iommu_groups/17/devices/0000:00:18.4
/sys/kernel/iommu_groups/17/devices/0000:00:18.5
/sys/kernel/iommu_groups/17/devices/0000:00:18.6
/sys/kernel/iommu_groups/17/devices/0000:00:18.7
/sys/kernel/iommu_groups/18/devices/0000:01:00.0
/sys/kernel/iommu_groups/19/devices/0000:20:00.0
/sys/kernel/iommu_groups/1/devices/0000:00:01.0
/sys/kernel/iommu_groups/20/devices/0000:21:01.0
/sys/kernel/iommu_groups/21/devices/0000:21:02.0
/sys/kernel/iommu_groups/22/devices/0000:21:03.0
/sys/kernel/iommu_groups/23/devices/0000:21:04.0
/sys/kernel/iommu_groups/24/devices/0000:21:05.0
/sys/kernel/iommu_groups/25/devices/0000:21:06.0
/sys/kernel/iommu_groups/26/devices/0000:21:08.0
/sys/kernel/iommu_groups/26/devices/0000:2e:00.0
/sys/kernel/iommu_groups/26/devices/0000:2e:00.1
/sys/kernel/iommu_groups/26/devices/0000:2e:00.3
/sys/kernel/iommu_groups/27/devices/0000:21:09.0
/sys/kernel/iommu_groups/27/devices/0000:2f:00.0
/sys/kernel/iommu_groups/28/devices/0000:21:0a.0
/sys/kernel/iommu_groups/28/devices/0000:30:00.0
/sys/kernel/iommu_groups/29/devices/0000:23:00.0
/sys/kernel/iommu_groups/2/devices/0000:00:01.1
/sys/kernel/iommu_groups/30/devices/0000:24:00.0
/sys/kernel/iommu_groups/31/devices/0000:25:00.0
/sys/kernel/iommu_groups/32/devices/0000:26:01.0
/sys/kernel/iommu_groups/33/devices/0000:26:03.0
/sys/kernel/iommu_groups/34/devices/0000:26:05.0
/sys/kernel/iommu_groups/35/devices/0000:26:07.0
/sys/kernel/iommu_groups/36/devices/0000:2b:00.0
/sys/kernel/iommu_groups/37/devices/0000:2c:00.0
/sys/kernel/iommu_groups/38/devices/0000:2d:00.0
/sys/kernel/iommu_groups/39/devices/0000:31:00.0
/sys/kernel/iommu_groups/39/devices/0000:31:00.1
/sys/kernel/iommu_groups/3/devices/0000:00:01.2
/sys/kernel/iommu_groups/40/devices/0000:32:00.0
/sys/kernel/iommu_groups/40/devices/0000:32:00.1
/sys/kernel/iommu_groups/41/devices/0000:33:00.0
/sys/kernel/iommu_groups/42/devices/0000:34:00.0
/sys/kernel/iommu_groups/43/devices/0000:34:00.3
/sys/kernel/iommu_groups/44/devices/0000:34:00.4
/sys/kernel/iommu_groups/45/devices/0000:35:00.0
/sys/kernel/iommu_groups/46/devices/0000:36:00.0
/sys/kernel/iommu_groups/4/devices/0000:00:02.0
/sys/kernel/iommu_groups/5/devices/0000:00:03.0
/sys/kernel/iommu_groups/6/devices/0000:00:03.1
/sys/kernel/iommu_groups/7/devices/0000:00:03.2
/sys/kernel/iommu_groups/8/devices/0000:00:04.0
/sys/kernel/iommu_groups/9/devices/0000:00:05.0

If the GPUs are not isolated, add to GRUB: pcie_acs_override=downstream,multifunction, then update-grub && reboot.

Security note: ACS override weakens isolation. Use it only on non‑production lab machines.

4) Create the AI VM (Ubuntu)

Create a new VM in the Proxmox UI:

  • BIOS: OVMF (UEFI)
  • Machine: q35
  • Display: None (or keep VirtIO‑GPU until drivers are installed)
  • SCSI Controller: VirtIO SCSI single
  • Disk: VirtIO SCSI on NVMe storage
  • CPU Type: host (later add hidden=1 if you plan Windows guests)
  • Cores/RAM: as needed (I give the Ubuntu AI VM most of the machine)
  • Network: VirtIO paravirtualized
Next time, I’ll automate this using the Terraform Proxmox provider, deploying new VMs straight from a GitHub Actions workflow running on private runners – just like I already do at work.
After a long workday, I’m fine with a bit of click-ops at 11 PM. Tomorrow starts early anyway – swimming at 6 AM, school drop-off with the boys, then biking to work by 8:15. Busy? Sure. Fun? Absolutely.
Clone the template…

Attach the Two GPUs after the VM is created:

  1. VM → Hardware → Add → PCI Device
  2. Select 31:00.0 (first GPU), check All Functions and PCI‑Express
  3. Add another PCI device 32:00.0 (second GPU), check All Functions and PCI‑Express
  4. (Optional) If this VM will drive a monitor, tick Primary GPU

This automatically pulls the .1 audio functions for each GPU.

Wow – both RTX 3090s are showing up! 🎉 All that (boring) time spent wrestling with Linux configs over SSH has paid off. Sure, it’s not as seamless as vSphere, where things just work without much tinkering, but this setup has its own kind of satisfaction.
Nice!

Direct Config File (alternative)

Edit /etc/pve/qemu-server/<VMID>.conf (example VMID 101):

machine: q35
bios: ovmf
cpu: host
hostpci0: 31:00,pcie=1
hostpci1: 32:00,pcie=1
# If you prefer explicit functions:
# hostpci0: 31:00.0,pcie=1
# hostpci1: 31:00.1,pcie=1
# hostpci2: 32:00.0,pcie=1
# hostpci3: 32:00.1,pcie=1

Here’s my current config (ignore the CPU and memory settings for now) – the only thing I care about at this stage is seeing the GPUs inside the VM. I’ll fine-tune the resources later this week when I start loading open-source AI models.

root@pve-01:~# cat /etc/pve/qemu-server/101.conf
boot: order=scsi0;ide2;net0
cores: 4
cpu: host
hostpci0: 0000:31:00,pcie=1
hostpci1: 0000:32:00,pcie=1
ide2: none,media=cdrom
machine: q35
memory: 24576
meta: creation-qemu=10.0.2,ctime=1757455146
name: ubuntu-edge-ai-001
net0: virtio=BC:24:11:1A:A9:32,bridge=vmbr0
numa: 0
ostype: l26
scsi0: local-lvm:vm-101-disk-0,iothread=1,size=32G
scsihw: virtio-scsi-single
smbios1: uuid=095969e7-3ad5-47fe-a2c2-3874c2da38af
sockets: 2
vmgenid: b49b626a-50c6-44e6-b6a8-fc3e21ef1697

5) Install NVIDIA Drivers Inside the Ubuntu VM

  1. Boot the VM from an Ubuntu ISO (22.04/24.04 LTS).
  2. After install, add NVIDIA drivers:
sudo apt update
sudo ubuntu-drivers autoinstall
sudo reboot
  1. Confirm both GPUs are visible:
attila@tpl-ubuntu-24:~$ nvidia-smi
Tue Sep  9 22:32:40 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 575.64.03              Driver Version: 575.64.03      CUDA Version: 12.9     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 3090        Off |   00000000:01:00.0 Off |                  N/A |
| 44%   52C    P8             25W /  350W |       9MiB /  24576MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA GeForce RTX 3090        Off |   00000000:02:00.0 Off |                  N/A |
|  0%   48C    P8             17W /  350W |       1MiB /  24576MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+
attila@tpl-ubuntu-24:~$ date

You should see two RTX 3090 entries with 24 GB VRAM each.

That’s it – job done, and before 2 AM. Perfect timing. Now, off to bed.

(Optional) CUDA & Containers – I’ll save this for my next blog article, not for today.

Performance Tips (Nice‑to‑Have)

  • CPU pinning: Pin vCPUs to physical cores for the AI VM and isolate those cores from the host (tune with isolcpus & Proxmox CPU pinning).
  • Hugepages: Enable 1G/2M hugepages for lower overhead on memory‑intensive training.
  • NUMA awareness: If you’re on a multi‑socket board, align vCPUs and memory to a single NUMA node.
  • Disk layout: Keep the AI dataset on fast NVMe; put lab VMs on separate storage to avoid contention.

The Triple‑Boot Plan (Why I Keep It)

  • Windows 11: video editing and occasional co‑op games with my boys
  • Native Ubuntu: raw, bare‑metal runs for maximum performance experiments
  • Proxmox VE: the main daily environment because I need the rest of the cores and RAM for lab VMs; only one giant Ubuntu VM gets both GPUs for AI

This gives me the best of all worlds without compromising on privacy or flexibility.

Troubleshooting Cheatsheet

  • Installer hangs at boot: add nomodeset at the boot menu (both Proxmox installer and first boots if needed).
  • GPUs not isolated in IOMMU groups: move cards to different slots; if still grouped, use pcie_acs_override=downstream,multifunction (lab only).
  • VM won’t start after attaching GPUs: ensure OVMF + q35, tick PCI‑Express, and check journalctl -xe for VFIO errors.
  • Windows guests show Code 43: set cpu: host,hidden=1, keep OVMF/UEFI, use recent NVIDIA drivers.
  • Black screen in guest: keep a temporary VirtIO-GPU display until NVIDIA drivers are installed, or attach via SSH.

What’s Next

With Proxmox in place and 2× RTX 3090 successfully passed through to an Ubuntu VM, I can run private edge models at home with near bare‑metal performance.

In the next article, I’ll walk through how I installed and configured AI models (LLMs, vision, and embeddings) in both virtualized and bare‑metal Ubuntu — and how I had unlimited fun with my private edge AI without relying on the cloud.

Picture of Attila Macskasy

Attila Macskasy

Multi-Cloud Architect
All articles
Search
Categories
Tags
Archives

Related posts

Fixing Proxmox Boot Hangs When Passing Through 2× RTX 3090 GPUs: Step-by-Step Troubleshooting Guide

2025.09.13.

Running multiple NVIDIA GPUs for AI workloads in Proxmox VE can cause early boot hangs if the host OS tries to load conflicting drivers. In this guide I document how my Proxmox host with 2× RTX 3090 was stuck at systemd-modules-load, how I debugged it, which files to inspect (/etc/default/grub, /etc/modprobe.d/, /etc/modules-load.d/), and the final stable configuration for rock-solid GPU passthrough to an Ubuntu VM.

Building the Perfect Edge AI Supercomputer – Adding an Edge Virtualization Layer with Proxmox and GPU Passthrough

2025.09.10.

I built on my edge AI hardware by adding Proxmox VE as the virtualization layer. After prepping BIOS, using Rufus with the nomodeset trick, and installing Proxmox, I enabled IOMMU, configured VFIO, and passed through 2× RTX 3090 GPUs to a single Ubuntu VM. This setup lets me run private AI workloads at near bare-metal speed, while keeping Windows and native Ubuntu for special use cases.

Budget AI Supercomputers: Dell Server vs. Threadripper Build vs. Next-Gen AI Desktop

2025.09.08.

Exploring three budget AI supercomputer paths: a Dell R740xd for enterprise labs with big storage but limited GPU flexibility, a TRX50 + Threadripper 7970X workstation offering fast DDR5, Gen5 NVMe, and dual RTX GPU power, and the futuristic GB10 AI desktop with unified CPU/GPU memory. Dell is lab-friendly, GB10 is AI-only, but the TRX50 build strikes the best balance today.

Building the Perfect Edge AI Supercomputer – Cost Effective Hardware

2025.09.07.

Keeping up with today’s technology is both exciting and demanding. My passion for home labs started many years ago, and while my family often jokes about the time and money I spend on self-education, they understand the value of staying ahead in such a fast-moving field. What started as curiosity has grown into a journey of building cost-effective supercomputers for edge AI and virtualization.

Fix VMware Workstation Performance Issues on Windows 11: Disable Hyper-V and VBS

2025.09.07.

This blog explains why VMware Workstation runs slower on Windows 11 compared to Windows 10, focusing on changes like Hyper-V, VBS, and HVCI being enabled by default on modern CPUs. It explores why sharing hypervisors with native hardware causes performance issues, and why disabling Hyper-V restores full VMware performance. Step-by-step PowerShell scripts are provided to toggle Hyper-V on or off safely.

Terraform deployment for FortiGate Next-Generation Firewall in Microsoft Azure

2025.01.28.

This blog explores deploying FortiGate VM in Azure, tackling challenges like license restrictions, Terraform API changes, and Marketplace agreements. It offers insights, troubleshooting tips, and lessons learned for successful single VM deployment in Azure. Using an evaluation license combined with B-series Azure VMs running FortiGate is primarily intended for experimentation and is not recommended for production environments.

Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.