GPU Virtualization with vGPU
vGPU technology uses a special driver to share the resources of a physical GPU between multiple VMs with Cloud GPU support.
This technology benefits users who do not need full physical GPU allocation.
Feature | Physical GPU | vGPU |
|---|---|---|
Cost efficiency | The entire physical GPU is allocated and charged for, and cost reduction by resource limitation is not available | Lower-cost configurations are available via partial allocation of GPU resources, and the configuration can be changed later if needed |
GPU allocation per VM | 100% of the total physical GPU resource | Up to 100%, varies depending on internal resource allocation policies for physical GPUs |
Available VRAM | 100% of total physical GPU memory | Up to 100%, depending on the profile used |
OS configuration steps | Only the installation of the graphics driver is required | In addition to driver installation, the OS also needs to be configured to connect to the license server. |
In addition to the standard characteristics (CPU, RAM, etc.), the name of the flavor with vGPU also includes the vGPU profile.
The vGPU profile contains two parts:
- Profile type, defined by a specific letter. Determines the type of tasks that will be processed by the GPU. Cloud GPU provides universal Q-series profiles. They are suitable for rendering the interface and advanced 3D projects, as well as for complex calculations (AI, ML, HPC) based on CUDA.
- Available video memory (VRAM) in GB.
For example, a flavor with 2 CPUs, 8 GB RAM, an L4 GPU, and 1 GB VRAM would be named vGPU-2-8-L4-1Q.
Cloud GPU provides flavors with vGPU based on NVIDIA® L4.
Flavor | CPU | RAM | GPU | Profile | VRAM, MB |
|---|---|---|---|---|---|
vGPU-2-8-L4-1Q | 2 | 8 | L4 | 1Q | 1024 |
vGPU-2-8-L4-2Q | 2 | 8 | L4 | 2Q | 2048 |
vGPU-4-8-L4-2Q | 4 | 8 | L4 | 2Q | 2048 |
vGPU-4-16-L4-4Q | 4 | 16 | L4 | 4Q | 4096 |
vGPU-8-16-L4-8Q | 8 | 16 | L4 | 8Q | 8192 |
vGPU-16-32-L4-12Q | 16 | 32 | L4 | 12Q | 12288 |
vGPU-32-64-L4-24Q | 32 | 64 | L4 | 24Q | 24576 |
Cloud GPU uses NVIDIA® vGPU Virtual Workstation license for Q-series profiles.
The license check in VK Cloud is performed by the local NVIDIA® License Server. Depending on which OS image is used, additional configuration may be required:
- OS image provided by VK Cloud: no additional actions are required, as the licensing server is already configured.
- Custom VM image: additional system configuration is required to use vGPU. See Installing the Windows vGPU driver and configuring the licensing token or Installing the Linux vGPU driver and configuring the licensing token.
If the check fails, the VM operates in trial mode for up to 20 minutes.