Special thanks #

https://github.com/QaidVoid/Complete-Single-GPU-Passthrough

https://youtu.be/BUSrdUoedTo?si=RSYHm7Jp37OSIKDh

Single GPU Passthrough #

I want to start by saying I love Linux and will never switch back to Windows again; however, simracing on Linux is still a bit rough. Sadly, I don’t expect things to get better anytime soon. Sure, with tools such as Lutris/Proton you can make things work, but it’s still not the full “just works” factor. At times I simply want to hop on the wheel, drive, and go about my day. So sadly, I have to install Windows, but obviously I’m not doing it the traditional way; let’s virtualize it!

For this guide, you need quite a good PC; otherwise, the performance lost will be quite impactful.

Written with the help of Local LLM gemma4 running on 3070ti.

To start, this is not a guide. Please do not consider it as one. I’m creating this solely for my personal use. If you want to use it, please modify or fork this project.

This is v2 — updated after actually getting it working, with the real bugs and their fixes included so future-me doesn’t have to rediscover them.

Getting started & Specs #

           .-------------------------:                    root@overlord
          .+=========================.                    -------------
         :++===++==================-       :++-           OS: CachyOS x86_64
        :*++====+++++=============-        .==:           Kernel: Linux 7.1.4-1-cachyos
       -*+++=====+***++==========:                        Uptime: 33 mins
      =*++++========------------:                         Packages: 1353 (pacman)
     =*+++++=====-                     ...                Shell: bash 5.3.15
   .+*+++++=-===:                    .=+++=:              Display (MSI MP341CQ): 3440x1440 in 34", 100 Hz [External]
  :++++=====-==:                     -*****+              Display (LG ULTRAWIDE): 2560x1080 in 29", 100 Hz [External]
 :++========-=.                      .=+**+.              Terminal: /dev/pts/0 10.4p1
.+==========-.                          .                 CPU: 11th Gen Intel(R) Core(TM) i7-11700K (16) @ 5.00 GHz
 :+++++++====-                                .--==-.     GPU: NVIDIA GeForce RTX 3070 Ti [Discrete]
  :++==========.                             :+++++++:    Memory: 1.93 GiB / 31.13 GiB (6%)
   .-===========.                            =*****+*+    Swap: 4.01 MiB / 31.13 GiB (0%)
    .-===========:                           .+*****+:    Disk (/): 21.91 GiB / 465.76 GiB (5%) - btrfs
      -=======++++:::::::::::::::::::::::::-:  .---:      Disk (/mnt/Game_1): 335.83 GiB / 915.82 GiB (37%) - ext4
       :======++++====+++******************=.             Disk (/mnt/Game_2): 471.80 GiB / 915.82 GiB (52%) - ext4
        :=====+++==========++++++++++++++*-               Local IP (enp4s0): 192.168.100.72/24
         .====++==============++++++++++*-                Locale: en_US.UTF-8
          .===+==================+++++++:
           .-=======================+++:
             ..........................

1. Install packages #

sudo pacman -S tpm2-tools qemu libvirt edk2-ovmf virt-manager dnsmasq ebtables qemu-hw-usb-host swtpm

swtpm is required if you use emulated TPM (recommended — see TPM section below).

2. Enable IOMMU #

vim /etc/default/grub

Add to GRUB_CMDLINE_LINUX_DEFAULT:

intel_iommu=on iommu=pt

Regenerate:

grub-mkconfig -o /boot/grub/grub.cfg

3. Enable libvirt + networking #

systemctl enable --now libvirtd
virsh net-start default
virsh net-autostart default
usermod -aG kvm,input,libvirt $(whoami)

Note: the default libvirt network is NAT, reachable only from the host itself. Other devices on your LAN (phone, laptop) can’t reach the VM’s IP directly — see the Networking section below for how to work around this or switch to bridged. You can use tailscale to resolve this issue.

4. Create the VM #

Launch virt-manager, create a new VM, select Customize before install.

Overview: Chipset = Q35, Firmware = UEFI
CPU: model = host-passthrough, topology = match your physical cores/threads (see CPU Pinning section below for a full pinning config once installed)
Disk: bus = virtio
NIC: model = virtio
TPM: use emulated TPM 2.0, not passthrough (see TPM section — passthrough caused a real hardware lockout during testing)
Add a CDROM with virtio-win.iso for drivers

About the virtio-win.iso: The download speed is f*cking slow, so be patient.

Begin installation. Windows won’t see the virtio disk — Load Driver → select virtio-iso/amd64/w11 (or win11 depending on your ISO) when prompted. After install, run the virtio driver installer from the CDROM, then remove the ISO.

Now add PCI Host devices for your GPU’s VGA function and its HDMI/DP audio function.

5. GPU vbios (required for single-GPU passthrough) #

If your GPU was the primary boot GPU, its ROM was already consumed by host firmware and QEMU can’t reliably re-read it live — you’ll hit vfio-pci: No more image in the PCI ROM without a dumped copy. Dump it before running the passthrough script, while the GPU is still driven by the host driver:

echo 1 > /sys/bus/pci/devices/0000:01:00.0/rom
cat /sys/bus/pci/devices/0000:01:00.0/rom > /root/vbios.rom
echo 0 > /sys/bus/pci/devices/0000:01:00.0/rom

Should start with 55 aa. If the dump is empty, use a matching vbios from TechPowerUp’s VGA BIOS database instead, matched to your exact card + board partner (check subsystem ID via lspci -vvv -s 01:00.0).

Important — where you put the file matters. Libvirt runs QEMU in a private mount namespace; a file under /root/ or /home/<user>/ is invisible to QEMU even as root, causing failed to find romfile. Put it alongside your VM’s disk image instead, since that path is already whitelisted:

cp /root/vbios.rom /mnt/Game_1/KVMS/vbios.rom

Only reference it on the GPU hostdev, not the audio one (You can do it there too, but only the GPU hostdev requires the romfile):

<hostdev mode="subsystem" type="pci" managed="yes">
  <source>
    <address domain="0x0000" bus="0x01" slot="0x00" function="0x0"/>
  </source>
  <rom file="/mnt/Game_1/KVMS/vbios.rom" bar="on"/>
  <address type="pci" domain="0x0000" bus="0x04" slot="0x00" function="0x0"/>
</hostdev>

6. First driver install — temporary virtual display #

Use a SSH tunnel and use RDP client.

ssh -L 3389:192.168.122.162:3389 root@<Linux_host_IP>

7. Hide the hypervisor from NVIDIA (avoid Code 43) #

<features>
  <acpi/>
  <apic/>
  <hyperv mode="custom">
    <relaxed state="on"/>
    <vapic state="on"/>
    <spinlocks state="on" retries="8191"/>
    <vpindex state="on"/>
    <runtime state="on"/>
    <synic state="on"/>
    <stimer state="on"/>
    <vendor_id state="on" value="fuckwindows"/>
    <frequencies state="on"/>
    <tlbflush state="on"/>
    <ipi state="on"/>
    <evmcs state="on"/>
    <avic state="on"/>
  </hyperv>
  <kvm>
    <hidden state="on"/>
  </kvm>
  <smm state="on"/>
</features>

Consumer NVIDIA GPUs refuse to initialize properly if they detect they’re in a VM — this is the intended fix.

8. TPM — use emulated, not passthrough (If windows 11) #

<tpm model="tpm-crb">
  <backend type="emulator" version="2.0"/>
</tpm>

This satisfies Windows 11’s TPM requirement via swtpm (a per-VM software TPM libvirt manages itself) — no /dev/tpm0 permission issues, no exclusivity conflicts.

Avoid real hardware TPM passthrough (<backend type="passthrough"><device path="/dev/tpm0"/></backend>) unless you specifically need hardware-backed TPM state. Repeated/failed passthrough attempts on this hardware put the physical fTPM into a bad state (Tss2_Tcti_Device_Init() ... errno 5: Input/output error), which cascaded into libvirtd itself failing to start (systemd’s credential-import step also talks to /dev/tpm0). Recovery required a full host reboot. Emulated TPM avoids this whole class of problem.

9. CPU pinning & memory #

Example for a 16-thread / 8-core CPU, giving 6 physical cores (12 threads) to the guest and keeping 2 physical cores (4 threads) for the host, with 24GB RAM:

<memory unit="KiB">25165824</memory>
<currentMemory unit="KiB">25165824</currentMemory>

<vcpu placement="static">12</vcpu>
<iothreads>1</iothreads>

<cputune>
  <vcpupin vcpu="0" cpuset="2"/>
  <vcpupin vcpu="1" cpuset="10"/>
  <vcpupin vcpu="2" cpuset="3"/>
  <vcpupin vcpu="3" cpuset="11"/>
  <vcpupin vcpu="4" cpuset="4"/>
  <vcpupin vcpu="5" cpuset="12"/>
  <vcpupin vcpu="6" cpuset="5"/>
  <vcpupin vcpu="7" cpuset="13"/>
  <vcpupin vcpu="8" cpuset="6"/>
  <vcpupin vcpu="9" cpuset="14"/>
  <vcpupin vcpu="10" cpuset="7"/>
  <vcpupin vcpu="11" cpuset="15"/>
  <emulatorpin cpuset="0-1,8-9"/>
  <iothreadpin iothread="1" cpuset="0-1,8-9"/>
</cputune>

<numatune>
  <memory mode="strict" nodeset="0"/>
</numatune>

<cpu mode="host-passthrough" check="none" migratable="off">
  <topology sockets="1" dies="1" clusters="1" cores="6" threads="2"/>
  <numa>
    <cell id="0" cpus="0-11" memory="25165824" unit="KiB"/>
  </numa>
</cpu>

Adjust core/thread numbers, memory, and pinning to match your own lstopo output.

10. Keyboard/mouse #

<input type='tablet' bus='usb'>
<input type='mouse' bus='ps2'/>
<input type='keyboard' bus='ps2'/>

11. Hooks #

mkdir /etc/libvirt/hooks
touch /etc/libvirt/hooks/qemu
chmod +x /etc/libvirt/hooks/qemu

vim /etc/libvirt/hooks/qemu

#!/bin/bash

GUEST_NAME="$1"
HOOK_NAME="$2"
STATE_NAME="$3"
MISC="${@:4}"

BASEDIR="$(dirname $0)"

HOOKPATH="$BASEDIR/qemu.d/$GUEST_NAME/$HOOK_NAME/$STATE_NAME"
set -e

if [ -f "$HOOKPATH" ]; then
eval \""$HOOKPATH"\" "$@"
elif [ -d "$HOOKPATH" ]; then
while read file; do
  eval \""$file"\" "$@"
done <<< "$(find -L "$HOOKPATH" -maxdepth 1 -type f -executable -print;)"
fi

# Script directory structure
[root@overlord hooks]# tree
.
├── qemu
└── qemu.d
    └── win11
        ├── prepare
        │   └── begin
        │       └── start.sh
        └── release
            └── end
                └── stop.sh

mkdir -p /etc/libvirt/hooks/qemu.d/win11/prepare/begin
mkdir -p /etc/libvirt/hooks/qemu.d/win11/release/end
chmod +x /etc/libvirt/hooks/qemu.d/win11/prepare/begin/start.sh
chmod +x /etc/libvirt/hooks/qemu.d/win11/release/end/stop.sh

12. Networking — reaching the VM from other devices #

The default libvirt network is NAT’d behind virbr0 — only the host can reach the VM’s IP directly. To RDP in from another device (phone, laptop):

ssh -L 3389:<vm-ip>:3389 root@<host-ip>

then RDP to localhost:3389 on the client device.

Once you RDP onto the server, you will have to install Nvidia drivers/GPU drivers for us to be able to actual use our monitors.

Get the VM’s current IP:

virsh domifaddr win11
# or
cat /var/lib/libvirt/dnsmasq/virbr0.status

https://github.com/rc0j/single-gpu-passthrough