vGPU Configuration Guide

Overview

vGPU (virtual GPU) allows partitioning a physical GPU into multiple virtual GPUs that can be assigned to different VMs. Essential for graphics-intensive workloads and Virtual Desktop Infrastructure (VDI).

Types of vGPU

NVIDIA vGPU

• Tesla: Data center GPUs (T4, A10, A100)
• Quadro: Professional GPUs (RTX 6000, 8000)
• GRID: Virtualization-specific GPUs

Intel GVT-g

• Integrated GPUs: Intel HD/UHD Graphics
• Available: Intel 6th gen+ CPUs

AMD MxGPU

• FirePro: Professional GPUs
• S7150, S7100: Series

Requirements

Hardware

• GPU with vGPU support
• Enough GPU memory for partitions
• VT-d/IOMMU enabled

Software

• Proxmox VE 8.0+
• vGPU driver (NVIDIA/Intel/AMD)
• License (NVIDIA GRID)

NVIDIA vGPU Setup

Install vGPU Manager

# Download from NVIDIA portal
wget NVIDIA-vGPU-16-13.8.0-16.13.8.x86_64.run
 
# Make executable
chmod +x NVIDIA-vGPU-*.run
 
# Install
./NVIDIA-vGPU-*.run
 
# Or extract and install
./NVIDIA-vGPU-*.run --extract-only
cd /tmp/NVIDIA-vGPU-*/DriverLink
./nv-carrier --dry-run
./nv-carrier --install
 
# Reboot
reboot

Verify Installation

# Check driver
nvidia-smi
 
# Check vGPU capabilities
ls /sys/class/mdev_bus/
 
# List available profiles
cat /sys/class/mdev_bus/mdev0/device/mdev_supported_types/*/name

Create vGPU

# Check available GPUs
lspci | grep -i nvidia
 
# Create vGPU (example)
echo "nvidia-grid-p40-4q" > /sys/class/mdev_bus/mdev0/device/mdev_supported_types/nvidia-grid-p40-4q/create
 
# Verify
ls -la /sys/class/mdev_bus/mdev0/
 
# Get UUID
cat /sys/class/mdev_bus/mdev0/device/mdev_supported_types/nvidia-grid-p40-4q/available_instances

vGPU Profiles

Profile	GPU Memory	vCPUs	Usability
nvidia-grid-p40-4q	4GB	4	Light work
nvidia-grid-p40-8q	8GB	8	Workstation
nvidia-grid-p40-16q	16GB	16	Heavy load
nvidia-grid-p40-24q	24GB	24	Full GPU

Assign to VM

# Get vGPU UUID
ls -la /sys/class/mdev_bus/mdev0/device/
 
# Assign to VM
qm set 100 --hostpci0 01:00.0,pcie=1,vgpu=uuid
 
# Using pcie format
qm set 100 --hostpci0 01:00,pcie=1,x-vga=1
 
# Alternative using mdev
qm set 100 --hostpci0 01:00.0,mdev=nvidia-grid-p40-4q,pcie=1

Intel GVT-g Setup

Enable GVT-g

# Add kernel parameter
nano /etc/default/grub
 
GRUB_CMDLINE_LINUX_DEFAULT="quiet intel_iommu=on iommu=pt i915.enable_gvt=1"
 
# Update GRUB
update-grub
 
# Reboot
reboot

Check Support

# Verify GVT-g enabled
dmesg | grep -i gvt
 
# List available types
ls /sys/class/mdev_bus/*/device/mdev_supported_types/
 
# Available profiles
find /sys/devices -name i915-*

Create Intel vGPU

# Check GPU
lspci | grep -i vga
 
# List available
ls /sys/class/mdev_bus/
 
# Create vGPU
echo "i915-GVTg_V5_4" > /sys/class/mdev_bus/mdev0/device/mdev_supported_types/i915-GVTg_V5_4/create
 
# Verify
ls /sys/class/mdev_bus/mdev0/

Intel vGPU Profiles

Profile	Memory	Use Case
i915-GVTg_V5_1	1GB	Basic
i915-GVTg_V5_2	2GB	Office
i915-GVTg_V5_4	4GB	Workstation
i915-GVTg_V5_8	8GB	Heavy

Assign to VM

# Assign to VM
qm set 100 --hostpci0 00:02.0,pcie=1,mdev=i915-GVTg_V5_4

AMD MxGPU Setup

Install MxGPU

# Download from AMD support portal
# Install using amdgpu-pro installer
./amdgpu-pro-install -- virtualization
 
# Configure
modprobe amdgpu-vgpu
 
# Reboot
reboot

Create MxGPU

# Check
ls /sys/class/mdev_bus/
 
# Create
echo "amd-mxgpu-4" > /sys/class/mdev_bus/mdev0/device/mdev_supported_types/amd-mxgpu-4/create

VM Configuration

Windows Guest

# Install NVIDIA/Intel driver in VM
# Download from vendor portal
 
# For NVIDIA:
# Install GRID driver
# Activate license (automatic or manual)
 
# Verify
nvidia-smi in VM

Linux Guest

# Install driver
apt install nvidia-driver-525
 
# Or for vGPU
apt install nvidia-vgpu-guest
 
# Verify
nvidia-smi

Performance Tuning

Memory Allocation

# Choose appropriate profile based on workload
# - Design/CAD: 8GB+ profiles
# - Office: 2GB profile
# - Light: 1GB profile

Licensing

# NVIDIA GRID license
# Configure license server
nvfliplib-cli -s license-server
 
# Manual activation
nvfliplib-cli -a -c "XXXX-XXXX-XXXX"

Use Cases

VDI (Virtual Desktop)

# Profile for office workers
# i915-GVTg_V5_2 or nvidia-grid-p40-4q
 
# Create pool
for i in 200..210; do
  qm create $i --name "vdi-$i" \
    --memory 4096 --cores 2 \
    --hostpci0 00:02.0,mdev=i915-GVTg_V5_2
done

CAD/Design

# High-performance profile
# nvidia-grid-p40-16q or nvidia-grid-p40-24q
 
qm create 300 --name "cad-workstation" \
  --memory 32768 --cores 8 \
  --hostpci0 01:00,pcie=1,x-vga=1,mdev=nvidia-grid-P40-16Q

Machine Learning

# Large GPU allocation
qm create 400 --name "ml-workstation" \
  --memory 65536 --cores 16 \
  --hostpci0 01:00,pcie=1,mdev=nvidia-grid-P40-24Q

Troubleshooting

vGPU Not Available

# Check driver
lsmod | grep -E "kvm|nvidia"
 
# Check IOMMU
dmesg | grep -i iommu
 
# Check license
nvidia-smi -q | grep -i license

Performance Issues

# Monitor GPU usage
nvidia-smi
 
# Check for throttling
nvidia-smi -q
 
# Profile memory
nvidia-smi -q -d MEMORY

Keywords

vgpu nvidia intel-gvt amd-mxgpu gpu-virtualization vdi desktop-virtualization grid

Related Articles

• Advanced
• PCI-Passthrough
• SR-IOV
• Performance

---

Back to Proxmox VE