NVIDIA vGPU with the GRID 15.0 driver
Two days ago, NVIDIA released their latest enterprise GRID driver. I created a patch that allows the use of most consumer GPUs for vGPU. One notable exception from that list is every officially unsupported Ampere GPU and GPUs from the Ada Lovelace generation.
This guide and all my tests were done on a RTX 2080 Ti which is based on the Turing architechture.
This tutorial assumes you are using a clean install of Proxmox 7.3, or ymmv when using an existing installation. Make sure to always have backups!
This guide should work for other linux systems with a recent kernel (5.15 to 5.19) but I have only tested it on the current proxmox version. If you are not using proxmox, you have to adapt some parts of this tutorial to work for your distribution.
Are you upgrading from a previous version of this guide?
If you are upgrading from a previous version of this guide, you should uninstall the old driver first:
nvidia-uninstallThen you also have to make sure that you are using the latest version of
vgpu_unlock-rs, otherwise it won't work with the latest driver.Either delete the folder
/opt/vgpu_unlock-rsor enter the folder and rungit pulland then recompile the library again usingcargo build --release
Packages
Make sure to add the community pve repo and get rid of the enterprise repo (you can skip this step if you have a valid enterprise subscription)
echo "deb http://download.proxmox.com/debian/pve bullseye pve-no-subscription" >> /etc/apt/sources.list
rm /etc/apt/sources.list.d/pve-enterprise.list
Update and upgrade
apt update
apt dist-upgrade
We need to install a few more packages like git, a compiler and some other tools.
apt install -y git build-essential dkms pve-headers mdevctl
Git repos and Rust compiler
First, clone this repo to your home folder (in this case /root/)
git clone https://gitlab.com/polloloco/vgpu-proxmox.git
You also need the vgpu_unlock-rs repo
cd /opt
git clone https://github.com/mbilker/vgpu_unlock-rs.git
After that, install the rust compiler
curl https://sh.rustup.rs -sSf | sh -s -- -y
Now make the rust binaries available in your $PATH (you only have to do it the first time after installing rust)
source $HOME/.cargo/env
Enter the vgpu_unlock-rs directory and compile the library. Depending on your hardware and internet connection that may take a while
cd vgpu_unlock-rs/
cargo build --release
Create files for vGPU unlock
The vgpu_unlock-rs library requires a few files and folders in order to work properly, lets create those
First create the folder for your vgpu unlock config and create an empty config file
mkdir /etc/vgpu_unlock
touch /etc/vgpu_unlock/profile_override.toml
Then, create folders and files for systemd to load the vgpu_unlock-rs library when starting the nvidia vgpu services
mkdir /etc/systemd/system/{nvidia-vgpud.service.d,nvidia-vgpu-mgr.service.d}
echo -e "[Service]\nEnvironment=LD_PRELOAD=/opt/vgpu_unlock-rs/target/release/libvgpu_unlock_rs.so" > /etc/systemd/system/nvidia-vgpud.service.d/vgpu_unlock.conf
echo -e "[Service]\nEnvironment=LD_PRELOAD=/opt/vgpu_unlock-rs/target/release/libvgpu_unlock_rs.so" > /etc/systemd/system/nvidia-vgpu-mgr.service.d/vgpu_unlock.conf
Have a vgpu supported card? Read here!
If you don't have a card like the Tesla P4, or any other gpu from this list, please continue reading at Enabling IOMMU
Disable the unlock part as doing this on a gpu that already supports vgpu, could break things as it introduces unnecessary complexity and more points of possible failure:
echo "unlock = false" > /etc/vgpu_unlock/config.toml
Enabling IOMMU
Note: Usually this isn't required for vGPU to work, but it doesn't hurt to enable it. You can skip this section, but if you run into problems later on, make sure to enable IOMMU.
To enable IOMMU you have to enable it in your BIOS/UEFI first. Due to it being vendor specific, I am unable to provide instructions for that, but usually for Intel systems the option you are looking for is called something like "Vt-d", AMD systems tend to call it "IOMMU".
After enabling it in your BIOS/UEFI, you also have to enable it in your kernel. Depending on how your system is booting, there are two ways to do that.
If you installed your system with ZFS-on-root and in UEFI mode, then you are using systemd-boot, everything else is GRUB. GRUB is way more common so if you are unsure, you are probably using that.
Depending on which system you are using to boot, you have to chose from the following two options:
GRUB
Open the file /etc/default/grub in your favorite editor
nano /etc/default/grub
The kernel parameters have to be appended to the variable GRUB_CMDLINE_LINUX_DEFAULT. On a clean installation that line should look like this
GRUB_CMDLINE_LINUX_DEFAULT="quiet"
If you are using an Intel system, append this after quiet:
intel_iommu=on iommu=pt
On AMD systems, append this after quiet:
amd_iommu=on iommu=pt
The result should look like this (for intel systems):
GRUB_CMDLINE_LINUX_DEFAULT="quiet intel_iommu=on iommu=pt"
Now, save and exit from the editor using Ctrl+O and then Ctrl+X and then apply your changes:
update-grub
systemd-boot
The kernel parameters have to be appended to the commandline in the file /etc/kernel/cmdline, so open that in your favorite editor:
nano /etc/kernel/cmdline
On a clean installation the file might look similar to this:
root=ZFS=rpool/ROOT/pve-1 boot=zfs
On Intel systems, append this at the end
intel_iommu=on iommu=pt
For AMD, use this
amd_iommu=on iommu=pt
After editing the file, it should look similar to this
root=ZFS=rpool/ROOT/pve-1 boot=zfs intel_iommu=on iommu=pt
Now, save and exit from the editor using Ctrl+O and then Ctrl+X and then apply your changes:
proxmox-boot-tool refresh
Loading required kernel modules and blacklisting the open source nvidia driver
We have to load the vfio, vfio_iommu_type1, vfio_pci and vfio_virqfd kernel modules to get vGPU working
echo -e "vfio\nvfio_iommu_type1\nvfio_pci\nvfio_virqfd" >> /etc/modules
Proxmox comes with the open source nouveau driver for nvidia gpus, however we have to use our patched nvidia driver to enable vGPU. The next line will prevent the nouveau driver from loading
echo "blacklist nouveau" >> /etc/modprobe.d/blacklist.conf
Applying our kernel configuration
I'm not sure if this is needed, but it doesn't hurt :)
update-initramfs -u -k all
...and reboot
reboot
Check if IOMMU is enabled
Note: See section "Enabling IOMMU", this is optional
Wait for your server to restart, then type this into a root shell
dmesg | grep -e DMAR -e IOMMU
On my Intel system the output looks like this
[ 0.007235] ACPI: DMAR 0x000000009CC98B68 0000B8 (v01 INTEL BDW 00000001 INTL 00000001)
[ 0.007255] ACPI: Reserving DMAR table memory at [mem 0x9cc98b68-0x9cc98c1f]
[ 0.020766] DMAR: IOMMU enabled
[ 0.062294] DMAR: Host address width 39
[ 0.062296] DMAR: DRHD base: 0x000000fed90000 flags: 0x0
[ 0.062300] DMAR: dmar0: reg_base_addr fed90000 ver 1:0 cap c0000020660462 ecap f0101a
[ 0.062302] DMAR: DRHD base: 0x000000fed91000 flags: 0x1
[ 0.062305] DMAR: dmar1: reg_base_addr fed91000 ver 1:0 cap d2008c20660462 ecap f010da
[ 0.062307] DMAR: RMRR base: 0x0000009cc18000 end: 0x0000009cc25fff
[ 0.062309] DMAR: RMRR base: 0x0000009f000000 end: 0x000000af1fffff
[ 0.062312] DMAR-IR: IOAPIC id 8 under DRHD base 0xfed91000 IOMMU 1
[ 0.062314] DMAR-IR: HPET id 0 under DRHD base 0xfed91000
[ 0.062315] DMAR-IR: x2apic is disabled because BIOS sets x2apic opt out bit.
[ 0.062316] DMAR-IR: Use 'intremap=no_x2apic_optout' to override the BIOS setting.
[ 0.062797] DMAR-IR: Enabled IRQ remapping in xapic mode
[ 0.302431] DMAR: No ATSR found
[ 0.302432] DMAR: No SATC found
[ 0.302433] DMAR: IOMMU feature pgsel_inv inconsistent
[ 0.302435] DMAR: IOMMU feature sc_support inconsistent
[ 0.302436] DMAR: IOMMU feature pass_through inconsistent
[ 0.302437] DMAR: dmar0: Using Queued invalidation
[ 0.302443] DMAR: dmar1: Using Queued invalidation
[ 0.333474] DMAR: Intel(R) Virtualization Technology for Directed I/O
[ 3.990175] i915 0000:00:02.0: [drm] DMAR active, disabling use of stolen memory
Depending on your mainboard and cpu, the output will be different, in my output the important line is the third one: DMAR: IOMMU enabled. If you see something like that, IOMMU is enabled.
NVIDIA Driver
As of the time of this writing (December 2022), the latest available GRID driver is 15.0 with vGPU driver version 525.60.12. You can check for the latest version here. I cannot guarantee that newer versions would work without additional patches, the patch in this guide works ONLY on 15.0 (525.60.12).
Obtaining the driver
NVIDIA doesn't let you freely download vGPU drivers like they do with GeForce or normal Quadro drivers, instead you have to download them through the NVIDIA Licensing Portal (see: https://www.nvidia.com/en-us/drivers/vgpu-software-driver/). You can sign up for a free evaluation to get access to the download page.
NB: When applying for an eval license, do NOT use your personal email or other email at a free email provider like gmail.com. You will probably have to go through manual review if you use such emails. I have very good experience using a custom domain for my email address, that way the automatic verification usually lets me in after about five minutes.
The file you are looking for is called NVIDIA-GRID-Linux-KVM-525.60.12-525.60.13-527.41.zip, you can get it from the download portal by downloading version 15.0 for Linux KVM.
For those who want to find the file somewhere else, here are some checksums :)
sha1: e4147e1dcebfc5459759ea013b56bca1d30f3578
md5: 0e2be7de643b99a62a1cca6ca37fd1ee
After downloading, extract that and copy the file NVIDIA-Linux-x86_64-525.60.12-vgpu-kvm.run to your Proxmox host into the /root/ folder
scp NVIDIA-Linux-x86_64-525.60.12-vgpu-kvm.run root@pve:/root/
Have a vgpu supported card? Read here!
If you don't have a card like the Tesla P4, or any other gpu from this list, please continue reading at Patching the driver
With a supported gpu, patching the driver is not needed, so you should skip the next section. You can simply install the driver package like this:
chmod +x NVIDIA-Linux-x86_64-525.60.12-vgpu-kvm.run ./NVIDIA-Linux-x86_64-525.60.12-vgpu-kvm.run --dkmsTo finish the installation, reboot the system
rebootNow, skip the following two sections and continue at Finishing touches
Patching the driver
Now, on the proxmox host, make the driver executable
chmod +x NVIDIA-Linux-x86_64-525.60.12-vgpu-kvm.run
And then patch it
./NVIDIA-Linux-x86_64-525.60.12-vgpu-kvm.run --apply-patch ~/vgpu-proxmox/525.60.12.patch
That should output a lot of lines ending with
Self-extractible archive "NVIDIA-Linux-x86_64-525.60.12-vgpu-kvm-custom.run" successfully created.
You should now have a file called NVIDIA-Linux-x86_64-525.60.12-vgpu-kvm-custom.run, that is your patched driver.
Installing the driver
Now that the required patch is applied, you can install the driver
./NVIDIA-Linux-x86_64-525.60.12-vgpu-kvm-custom.run --dkms
The installer will ask you Would you like to register the kernel module sources with DKMS? This will allow DKMS to automatically build a new module, if you install a different kernel later., answer with Yes.
Depending on your hardware, the installation could take a minute or two.
If everything went right, you will be presented with this message.
Installation of the NVIDIA Accelerated Graphics Driver for Linux-x86_64 (version: 525.60.12) is now complete.
Click Ok to exit the installer.
To finish the installation, reboot.
reboot
Finishing touches
Wait for your server to reboot, then type this into the shell to check if the driver install worked
nvidia-smi
You should get an output similar to this one
Sun Dec 4 12:54:59 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.60.12 Driver Version: 525.60.12 CUDA Version: N/A |
|-------------------------------+----------------------+----------------------+
| 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 ... On | 00000000:01:00.0 Off | N/A |
| 26% 33C P8 43W / 260W | 85MiB / 11264MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
To verify if the vGPU unlock worked, type this command
mdevctl types
The output will be similar to this
0000:01:00.0
nvidia-256
Available instances: 24
Device API: vfio-pci
Name: GRID RTX6000-1Q
Description: num_heads=4, frl_config=60, framebuffer=1024M, max_resolution=5120x2880, max_instance=24
nvidia-257
Available instances: 12
Device API: vfio-pci
Name: GRID RTX6000-2Q
Description: num_heads=4, frl_config=60, framebuffer=2048M, max_resolution=7680x4320, max_instance=12
nvidia-258
Available instances: 8
Device API: vfio-pci
Name: GRID RTX6000-3Q
Description: num_heads=4, frl_config=60, framebuffer=3072M, max_resolution=7680x4320, max_instance=8
---SNIP---
If this command doesn't return any output, vGPU unlock isn't working.
Another command you can try to see if your card is recognized as being vgpu enabled is this one:
nvidia-smi vgpu
If everything worked right with the unlock, the output should be similar to this:
Sun Dec 4 12:55:09 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.60.12 Driver Version: 525.60.12 |
|---------------------------------+------------------------------+------------+
| GPU Name | Bus-Id | GPU-Util |
| vGPU ID Name | VM ID VM Name | vGPU-Util |
|=================================+==============================+============|
| 0 NVIDIA GeForce RTX 208... | 00000000:01:00.0 | 0% |
+---------------------------------+------------------------------+------------+
However, if you get this output, then something went wrong
No supported devices in vGPU mode
vGPU overrides
Further up we have created the file /etc/vgpu_unlock/profile_override.toml and I didn't explain what it was for yet. Using that file you can override lots of parameters for your vGPU instances: For example you can change the maximum resolution, enable/disable the frame rate limiter, enable/disable support for CUDA or change the vram size of your virtual gpus.
If we take a look at the output of mdevctl types we see lots of different types that we can choose from. However, if we for example chose GRID RTX6000-4Q which gives us 4GB of vram in a VM, we are locked to that type for all of our VMs. Meaning we can only have 4GB VMs, its not possible to mix different types to have one 4GB VM, and two 2GB VMs.
Important note
C profiles (for example
GRID RTX6000-4C) only work on Linux, don't try using those on Windows, it will not work - at all.
All of that changes with the override config file. Technically we are still locked to only using one profile, but now its possible to change the vram of the profile on a VM basis so even though we have three GRID RTX6000-4Q instances, one VM can have 4GB or vram but we can override the vram size for the other two VMs to only 2GB.
Lets take a look at this example config override file (its in TOML format)
[profile.nvidia-259]
num_displays = 1 # Max number of virtual displays. Usually 1 if you want a simple remote gaming VM
display_width = 1920 # Maximum display width in the VM
display_height = 1080 # Maximum display height in the VM
max_pixels = 2073600 # This is the product of display_width and display_height so 1920 * 1080 = 2073600
cuda_enabled = 1 # Enables CUDA support. Either 1 or 0 for enabled/disabled
frl_enabled = 1 # This controls the frame rate limiter, if you enable it your fps in the VM get locked to 60fps. Either 1 or 0 for enabled/disabled
framebuffer = 0x74000000
framebuffer_reservation = 0xC000000 # In combination with the framebuffer size
# above, these two lines will give you a VM
# with 2GB of VRAM (framebuffer + framebuffer_reservation = VRAM size in bytes).
# See below for some other sizes
[mdev.00000000-0000-0000-0000-000000000100]
frl_enabled = 0
# You can override all the options from above here too. If you want to add more overrides for a new VM, just copy this block and change the UUID
There are two blocks here, the first being [profile.nvidia-259] and the second [mdev.00000000-0000-0000-0000-000000000100].
The first one applies the overrides to all VM instances of the nvidia-259 type (thats GRID RTX6000-4Q) and the second one applies its overrides only to one specific VM, that one with the uuid 00000000-0000-0000-0000-000000000100.
You don't have to specify all parameters, only the ones you need/want. There are some more that I didn't mention here, you can find them by going through the source code of the vgpu_unlock-rs repo.
For a simple 1080p remote gaming VM I recommend going with something like this
[profile.nvidia-259] # choose the profile you want here
num_displays = 1
display_width = 1920
display_height = 1080
max_pixels = 2073600
Common VRAM sizes
Here are some common framebuffer sizes that you might want to use:
- 512MB:
framebuffer = 0x1A000000 framebuffer_reservation = 0x6000000 - 1GB:
framebuffer = 0x38000000 framebuffer_reservation = 0x8000000 - 2GB:
framebuffer = 0x74000000 framebuffer_reservation = 0xC000000 - 3GB:
framebuffer = 0xB0000000 framebuffer_reservation = 0x10000000 - 4GB:
framebuffer = 0xEC000000 framebuffer_reservation = 0x14000000 - 5GB:
framebuffer = 0x128000000 framebuffer_reservation = 0x18000000 - 6GB:
framebuffer = 0x164000000 framebuffer_reservation = 0x1C000000 - 8GB:
framebuffer = 0x1DC000000 framebuffer_reservation = 0x24000000 - 10GB:
framebuffer = 0x254000000 framebuffer_reservation = 0x2C000000 - 12GB:
framebuffer = 0x2CC000000 framebuffer_reservation = 0x34000000 - 16GB:
framebuffer = 0x3BC000000 framebuffer_reservation = 0x44000000 - 20GB:
framebuffer = 0x4AC000000 framebuffer_reservation = 0x54000000 - 24GB:
framebuffer = 0x59C000000 framebuffer_reservation = 0x64000000 - 32GB:
framebuffer = 0x77C000000 framebuffer_reservation = 0x84000000 - 48GB:
framebuffer = 0xB2D200000 framebuffer_reservation = 0xD2E00000
framebuffer and framebuffer_reservation will always equal the VRAM size in bytes when added together.
Spoofing your vGPU instance
Note: This only works on Windows guests, don't bother trying on Linux.
You can very easily spoof your virtual GPU to a different card, so that you could install normal quadro drivers instead of the GRID drivers that require licensing.
For that you just have to add two lines to the override config. In this example I'm spoofing my Turing based card to a normal RTX 6000 Quadro card:
[profile.nvidia-259]
# insert all of your other overrides here too
pci_device_id = 0x1E30
pci_id = 0x1E3012BA
pci_device_id is the pci id from the card you want to spoof to. In my case its 0x1E30 which is the Quadro RTX 6000/8000.
pci_id can be split in two parts: 0x1E30 12BA, the first part 0x1E30 has to be the same as pci_device_id. The second part is the subdevice id. In my case 12BA means its a RTX 6000 card and not RTX 8000.
You can get the IDs from here. Just Ctrl+F and search the card you want to spoof to, then copy the id it shows you on the left and use it for pci_device_id.
After doing that, click the same id, it should open a new page where it lists the subsystems. If there are none listed, you must use 0000 as the second value for pci_id. But if there are some, you have to select the one you want and use its id as the second value for pci_id (see above).
Important note when spoofing
You have to pick a Quadro Driver from the same driver branch, so in this case R525. Using newer drivers will NOT WORK and maybe even make your VM crash.
If you accidentally installed such a driver, its best to either remove the driver completely using DDU or just install a fresh windows VM.
The quadro driver for R525 branch can be found here (for 527.27).
Adding a vGPU to a Proxmox VM
There is only one thing you have to do from the commandline: Open the VM config file and give the VM a uuid.
For that you need your VM ID, in this example I'm using 1000.
nano /etc/pve/qemu-server/<VM-ID>.conf
So with the VM ID 1000, I have to do this:
nano /etc/pve/qemu-server/1000.conf
In that file, you have to add a new line at the end:
args: -uuid 00000000-0000-0000-0000-00000000XXXX
You have to replace XXXX with your VM ID. With my 1000 ID I have to use this line:
args: -uuid 00000000-0000-0000-0000-000000001000
Save and exit from the editor. Thats all you have to do from the terminal.
Now go to the proxmox webinterface, go to your VM, then to Hardware, then to Add and select PCI Device.
You should be able to choose from a list of pci devices. Choose your GPU there, its entry should say Yes in the Mediated Devices column.
Now you should be able to also select the MDev Type. Choose whatever profile you want, if you don't remember which one you want, you can see the list of all available types with mdevctl types.
Finish by clicking Add, start the VM and install the required drivers. After installing the drivers you can shut the VM down and remove the virtual display adapter by selecting Display in the Hardware section and selecting none (none). ONLY do that if you have some other way to access the Virtual Machine like Parsec or Remote Desktop because the Proxmox Console won't work anymore.
Enjoy your new vGPU VM :)
Support
If something isn't working, please create an issue or join the Discord server and ask for help in the #proxmox-support channel.
When asking for help, please describe your problem in detail instead of just saying "vgpu doesn't work". Usually a rough overview over your system (gpu, mainboard, proxmox version, kernel version, ...) and full output of dmesg and/or journalctl --no-pager -b 0 -u nvidia-vgpu-mgr.service (<-- this only after starting the VM that causes trouble) is helpful.
Please also provide the output of uname -a and cat /proc/cmdline
Further reading
Thanks to all these people (in no particular order) for making this project possible
- DualCoder for his original vgpu_unlock repo with the kernel hooks
- mbilker for the rust version, vgpu_unlock-rs
- KrutavShah for the wiki
- HiFiPhile for the C version of vgpu unlock
- rupansh for the original twelve.patch to patch the driver on kernels >= 5.12
- mbuchel#1878 on the GPU Unlocking discord for fourteen.patch to patch the driver on kernels >= 5.14
- erin-allison for the nvidia-smi wrapper script
- LIL'pingu#9069 on the GPU Unlocking discord for his patch to nop out code that NVIDIA added to prevent usage of drivers with a version 460 - 470 with consumer cards
If I forgot to mention someone, please create an issue or let me know otherwise.
Contributing
Pull requests are welcome (factual errors, amendments, grammar/spelling mistakes etc).