Proxmox - Virtualisation: Advanced GPU Passthrough with IOMMU Isolation

The Holy Grail of Virtualisation: Near-Native GPU Performance

GPU Passthrough allows a Virtual Machine (Windows for gaming, or Linux for AI/CUDA) to have direct, exclusive access to a physical PCIe graphics card. In a professional setting, this is essential for VDI (Virtual Desktop Infrastructure), hardware-accelerated video encoding, or local LLM (Large Language Model) training.

However, the path to a working passthrough is littered with “Code 43” errors and system hangs. A experienced administrator knows that success depends on two things: IOMMU Isolation and Vendor-Hidden Virtualisation.

Phase 1: Enabling the IOMMU Foundation

IOMMU (Intel VT-d or AMD-Vi) is the hardware feature that allows the kernel to isolate a PCIe device from the rest of the system’s memory.

1. Kernel Parameters

Edit /etc/default/grub and add the isolation flags:

# For Intel
GRUB_CMDLINE_LINUX_DEFAULT="quiet intel_iommu=on iommu=pt"
# For AMD
GRUB_CMDLINE_LINUX_DEFAULT="quiet amd_iommu=on iommu=pt"

Note: iommu=pt (Pass Through) prevents the kernel from trying to manage every device, which improves performance and stability.

2. Loading the VFIO Modules

You must tell Proxmox to use the “Virtual Function I/O” drivers for your GPU instead of the standard NVIDIA/AMD drivers.

# /etc/modules
vfio
vfio_iommu_type1
vfio_pci
vfio_virqfd

Phase 2: Resolving the IOMMU Group Trap

A PCIe device cannot be passed through alone if it shares an IOMMU Group with other devices (like a USB controller or a SATA port). If you pass the GPU, you must pass everything in that group, or the system will crash.

The Diagnostic:

# Check groups
for d in /sys/kernel/iommu_groups/*/devices/*; do n=${d#*/iommu_groups/*}; n=${n%%/*}; printf 'Group %s ' "$n"; lspci -nns "${d##*/}"; done

If your GPU is in a “dirty” group with critical system components, you have two choices:

1. Move the card: Try a different physical PCIe slot (often the primary slot has better isolation).
2. The ‘ACS Override’ Patch: A risky kernel patch that forces the kernel to ignore hardware grouping.

pcie_acs_override=downstream,multifunction

Phase 3: Bypassing ‘Code 43’ (Vendor Workarounds)

NVIDIA and AMD drivers often check if they are running inside a VM. If they detect a hypervisor, they may disable themselves.

The Advanced VM Config: In the Proxmox .conf file for your VM (/etc/pve/qemu-server/XXX.conf), you must hide the “KVM” signature and spoof the PCIe headers.

cpu: host,hidden=1,flags=+pdpe1gb
hostpci0: 0000:01:00,x-vga=1,pcie=1
args: -cpu 'host,hv_ips,hv_relaxed,hv_reset,hv_runtime,hv_spinlocks=0x1fff,hv_stimer,hv_synic,hv_vapic,hv_vpindex,kvm=off'

Strategy: Handling Audio over HDMI

A GPU passthrough usually includes the HDMI Audio controller. You must pass both the Video and Audio IDs together.

# Blacklist the default drivers to prevent the host from grabbing the card
echo "blacklist nouveau" >> /etc/modprobe.d/blacklist.conf
echo "options vfio-pci ids=10de:1b81,10de:10f0" >> /etc/modprobe.d/vfio.conf

Summary: Testing for Success

1. The ‘Black Screen’ Fix: Ensure you have disabled the “VGA” output in the VM hardware settings. Once passthrough is active, the GPU is the only output.
2. Power Management: Ensure the GPU is getting enough power. Some high-end cards fail to initialize if the PCIe slot doesn’t provide enough “negotiated” wattage before the guest driver takes over.

GPU Passthrough transforms a Proxmox host into a high-performance workstation. By mastering the nuances of IOMMU grouping and hypervisor masking, you can achieve 98-99% of bare-metal performance inside a virtual environment. This is the hallmark of a optimal virtualisation strategy.