Skip to content

Proxmox GPU Passthrough Tutorial (AMD Ryzen 9 7950X + NVIDIA RTX 5060 Ti)

This post is a cleaned-up tutorial from a real troubleshooting session. The target setup is:

  • Host: Proxmox VE on AMD platform
  • CPU: Ryzen 9 7950X (with integrated AMD Raphael graphics)
  • Motherboard: MSI PRO B650M-A WIFI
  • Discrete GPU to passthrough: NVIDIA GeForce RTX 5060 Ti
  • Host display stays on the Ryzen iGPU, while the NVIDIA card is dedicated to the VM

If you are on a similar AM5 + NVIDIA setup, this walkthrough should map very closely.

Technology Stack Behind This Tutorial

This guide touches multiple layers of the virtualization stack. Understanding these layers makes troubleshooting much easier:

  1. Firmware layer (UEFI/BIOS)

    • SVM (AMD-V) enables CPU virtualization extensions.
    • IOMMU (AMD-Vi) enables DMA remapping and device isolation.
    • UEFI mode avoids legacy initialization behavior that can interfere with modern GPU passthrough.

  2. Hypervisor layer (Proxmox VE)

    • Proxmox is the management layer; actual VM acceleration is provided by Linux KVM and device emulation by QEMU.
    • GPU passthrough means QEMU exposes a real PCIe device directly to the guest, instead of emulating a virtual GPU.

  3. Isolation layer (Linux IOMMU groups)

    • IOMMU = Input-Output Memory Management Unit.
    • The kernel groups PCI devices by isolation boundaries.
    • Passthrough is safe/reliable only when devices are in proper IOMMU groups.

  4. Driver binding layer (VFIO)

    • VFIO = Virtual Function I/O.
    • vfio (VFIO core framework), vfio_iommu_type1 (Type-1 IOMMU backend), vfio_pci (PCI device binding driver), and vfio_virqfd (virtual interrupt eventfd support) allow userspace VMs to own PCI devices securely.
    • Binding the GPU to vfio-pci prevents host graphics drivers from taking the card first.

  5. Boot and module orchestration layer

    • GRUB injects kernel flags (amd_iommu=on iommu=pt).
    • initramfs ensures required modules and binding logic are available early during boot.
    • /etc/modules-load.d/ controls module auto-load; /etc/modprobe.d/ controls module options and blacklists.

  6. Guest VM platform layer

    • OVMF (UEFI) + q35 gives a modern virtual platform aligned with PCIe passthrough.
    • CPU type: host exposes native CPU features.
    • Disabling memory ballooning avoids DMA-related instability for passthrough workloads.

  7. Verification toolchain

    • dmesg, /proc/cmdline, find /sys/kernel/iommu_groups, and lspci -nnk validate host readiness.
    • nvidia-smi validates the guest-side driver and runtime state.

1. What Success Looks Like

Before details, here is the expected end state:

  1. BIOS has SVM and IOMMU enabled.
  2. Kernel boots with amd_iommu=on iommu=pt.
  3. IOMMU groups exist under /sys/kernel/iommu_groups/.
  4. NVIDIA GPU functions are bound to vfio-pci on the host.
  5. VM is created with OVMF (UEFI) + q35, and GPU is attached as PCI passthrough.

2. BIOS Configuration (MSI Click BIOS 5)

Step 1: Enter Advanced BIOS

Enter BIOS (usually Delete at boot), then switch to Advanced Mode.

img01_msi_bios_advanced_entry

Step 2: Navigate to AMD CBS

Go to:

  • OC -> Advanced CPU Configuration -> AMD CBS

img02_oc_menu_navigation

Step 3: Enable Virtualization Features

Inside AMD CBS, set:

  • SVM Enable -> Enabled
  • IOMMU -> Enabled

Use explicit Enabled instead of Auto for passthrough reliability.

Under the hood:

  • SVM enables hardware-assisted virtualization instructions used by KVM.
  • IOMMU enables DMA remapping, which is the foundation for safe PCIe passthrough.
  • If either is missing, VFIO can load but device assignment will typically fail or be unstable.

img03_amd_cbs_cpu_options

Step 4: Optional Stability Tweak

Optional on Ryzen hosts used as servers:

  • Global C-state Control -> Disabled

Other options like Pre-boot DMA Protection and PCIe ARI Support can remain Auto unless you have a specific need.

img04_amd_cbs_advanced_options.1 img04_amd_cbs_advanced_options.2

Step 5: Confirm Boot Mode

In Settings -> Boot, confirm UEFI boot mode.

  • Boot mode select should be UEFI

img05_boot_mode_uefi

Save and reboot (F10).

3. Verify and Enable IOMMU in Proxmox

Step 6: First Check in Host Shell

dmesg | grep -e DMAR -e IOMMU

On AMD, you may initially see detection lines such as:

AMD-Vi: IOMMU performance counters supported
Detected AMD IOMMU #0

This confirms hardware is visible, but we still need kernel flags.

Step 7: Add Kernel Parameters (GRUB)

Check current cmdline:

cat /proc/cmdline

In this session it showed root=/dev/mapper/pve-root, which indicates GRUB on LVM.

Edit:

nano /etc/default/grub

Set:

GRUB_CMDLINE_LINUX_DEFAULT="quiet amd_iommu=on iommu=pt"

Why these flags matter:

  • amd_iommu=on: explicitly enables AMD IOMMU support in kernel boot path.
  • iommu=pt: keeps host-side overhead lower for non-assigned devices by using pass-through mappings where possible.
  • Without correct boot flags, BIOS settings alone may not produce usable runtime isolation.

Apply and reboot:

update-grub
reboot

After reboot, verify:

cat /proc/cmdline

You should now see amd_iommu=on iommu=pt.

Step 8: Verify IOMMU Groups Exist

Quick check:

find /sys/kernel/iommu_groups/ -maxdepth 0

Better check (shows actual group entries):

find /sys/kernel/iommu_groups/ -type l | head

If /sys/kernel/iommu_groups/ exists and contains entries, IOMMU is active.

This is a key architectural checkpoint: IOMMU groups are the kernel's isolation model, and VFIO relies on this model to guarantee a device can be safely handed to a VM.

4. Load VFIO Modules at Boot

Historically people edited /etc/modules, but on systemd-based systems the cleaner way is a dedicated file in /etc/modules-load.d/.

Stack detail:

  • /etc/modules-load.d/*.conf controls which kernel modules are loaded at boot.
  • /etc/modprobe.d/*.conf controls module parameters (for example, VFIO device IDs) and blacklists.
  • Keeping loading and policy in separate files makes maintenance and debugging clearer.

Create:

nano /etc/modules-load.d/vfio.conf

Add:

vfio
vfio_iommu_type1
vfio_pci
vfio_virqfd

Update initramfs:

update-initramfs -u -k all

If you see:

No /etc/kernel/proxmox-boot-uuids found, skipping ESP sync.

that is normal for a GRUB/LVM layout.

Reboot:

reboot

5. Bind the NVIDIA GPU to vfio-pci

Step 9: Identify GPU IDs

lspci -nnk | grep -A 3 "VGA"

In this setup:

  • NVIDIA GPU: 01:00.0 -> [10de:2d04]
  • NVIDIA audio: 01:00.1 -> [10de:22eb]
  • Host iGPU (keep for Proxmox): AMD Raphael 11:00.0 (amdgpu)

Step 10: Configure vfio-pci IDs

Create:

nano /etc/modprobe.d/vfio.conf

Add:

options vfio-pci ids=10de:2d04,10de:22eb disable_vga=1

What this does:

  • ids=... tells VFIO exactly which PCI functions to claim (GPU + HDMI/DP audio).
  • disable_vga=1 helps avoid legacy VGA routing conflicts on some boards/firmware combinations.

Step 11: Blacklist Host NVIDIA Drivers

Create:

nano /etc/modprobe.d/blacklist.conf

Add:

blacklist nvidia
blacklist nouveau

Notes:

  • Do not blacklist amdgpu (host iGPU needs it).
  • blacklist radeon is optional and usually unnecessary on modern Ryzen iGPU setups.

The binding order is critical: if nouveau/nvidia binds first, VFIO cannot cleanly claim the device later without manual unbind/rebind.

Apply and reboot:

update-initramfs -u -k all
reboot

Step 12: Final Host-Side Verification

lspci -nnk -d 10de:

Success should look like:

01:00.0 ... Kernel driver in use: vfio-pci
01:00.1 ... Kernel driver in use: vfio-pci

The chat session reached this exact success state.

6. Create the Ubuntu VM (Proxmox UI)

Once host-side binding is done, create a new VM instead of reusing old SeaBIOS templates.

Recommended wizard settings:

  1. OS: your Ubuntu ISO
  2. System:
    • BIOS: OVMF (UEFI)
    • Machine: q35
    • Graphics: keep Default or VirtIO-GPU during install
  3. CPU:
    • Type: host
  4. Memory:
    • Disable ballooning
  5. Network:
    • VirtIO

Why this VM stack is recommended:

  • OVMF matches modern GPU firmware expectations and guest driver behavior.
  • q35 models a PCIe-first chipset, which is a better fit for passthrough than legacy i440fx.
  • VirtIO keeps paravirtualized devices efficient so CPU/GPU resources are focused on real workloads.

Then add PCI device:

  • Hardware -> Add -> PCI Device
  • Select 01:00.0 (NVIDIA VGA)
  • Check:
  • All Functions
  • ROM-Bar
  • PCI-Express

Why only 01:00.0?

  • Because All Functions also includes 01:00.1 audio automatically.

Install Ubuntu, then inside guest install NVIDIA driver and verify with:

nvidia-smi

7. Troubleshooting Notes

  1. vfio-pci not in use after reboot:
    • Recheck vfio.conf, blacklist file, and rerun update-initramfs -u -k all.
  2. No IOMMU groups:
    • Recheck BIOS SVM/IOMMU and kernel cmdline flags.
  3. Black screen when starting VM installer:
    • Keep a virtual display enabled during initial install, then tune display later.
  4. Confused by Kernel modules: output in lspci:
    • It only means modules exist; Kernel driver in use is what matters.

8. Summary

The host side is done when both 01:00.0 and 01:00.1 show Kernel driver in use: vfio-pci.

At that point, Proxmox is no longer using the NVIDIA card, and the VM can claim it directly for near-native performance.