Through various recent projects, I had to work through the clutter of information regarding NVIDIA vGPU licensing.
Here is a small summary of this information.
NVIDIA vGPU Architecture
Under the control of the NVIDIA GPU Virtual Manager, running in the hypervisor, the NVIDIA Physical GPU can operate multiple virtual GPU devices (vGPUs), that can be assigned directly to the Guest VM.
The Guest VMs use the NVIDIA virtual vGPUs in the same way as a physical GPU would come from the hypervisor by direct passed through. The NVIDIA Driver loaded into the guest VM provides Direct GPU Access for high-performance operations. The NVIDIA Virtual GPU Manager paravirtualized interface performs the non-performance management operations for the NVIDIA Driver.
Each NVIDIA vGPU is analogous to a conventional GPU, equipped with a fixed amount of GPU frame buffer and one or more virtual display outputs or “heads”. The vGPU frame buffer is allocated from the physical frame buffer of the physical GPU at the time of creation, and this vGPU retains exclusive access to this frame buffer part until it is destroyed.
All vGPUs residing on a physical GPU can share access to the GPU engines, including graphics (3D), video decoding, and video encoding modules.
Each physical GPU can support several different types of vGPU Profiles (Virtual GPU Type) simultaneously. Each vGPU profile has fixed hardware key data, such as the frame buffer size, number of supported displays (Virtual Display Heads) and the maximum resolution (per display head). They are divided into different series, each corresponding to different load classes. Each series is identified by the last letter of the vGPU profile name.
- The Q-Series is designed for Designers and Power Users.
- The B-Series is designed for Power Users.
- The A-Series is designed for Users of virtual applications.
The number after the card type (P4 for Pascal Microarchitecture) in the name of the vGPU profile indicates the size of the frame buffer. For example, a P4-8Q vGPU on a Tesla P4 board is assigned 8192 MB of Frame Buffer.
Because of the different resource requirements of each vGPU profile, the maximum number of vGPUs that can be created simultaneously on a single physical GPU varies. For example, a Tesla P4 can board on its physical GPU up to 4 P4-2Q vGPUs, but only 2 P4-4Q vGPU profiles.
Licensing of a NVIDIA vGPU
When booting with a supported GPU, a vGPU with reduced capacity will be executed until a license is purchased.
The performance of an unlicensed vGPU is limited as follows:
- The frame rate is limited to 3 frames per second.
- The allocation of GPU resources is limited, so some applications can not run properly.
- For vGPUs that support CUDA, CUDA is disabled.
These restrictions are removed when importing a license.
NVIDIA License Terminology
It is important to note that the licenses are not to pay per VM, but in the Concurrent Use Model, per user who works with the vGPUs VM.
You must watch the peak times regarding access (for example 10 users) and calculates the license consumption.
Behind SUMs, NVIDIA hides the Support, Upgrade and Maintenance program that is automatically purchased for each type of license.
The Perpertual License never expire and you just have to renew the SUMs. Initially the Perpetual License are sold with SUMs lengths of 3, 4 or 5 years and can be renewed annually thereafter.
These licenses are valid for a period of time (1, 3 or 5 years) and directly include the SUMs for that period.
At the end of this period, you can not extend the Annual Subscription License, you must purchase it again.
NVIDIA vGPU Software Edition
NVIDIA distinguishes between 3 software editions, with more or less features. A mixed form of the different license models can be provided.
Example of a license mixed form For the NX 3D application, there are 50 users with read access to the files via Citrix Published Apps and 10 users who develop the models in a Citrix Virtual Desktop Windows 10 VM in this program. 50 x vApps License 10 x Quadro vDWS License
NVIDIA grid virtual Applications
This is useful for organizations that have deployed Citrix Virtual Apps solutions in their portfolio, for example, to allow users to view files created in 3D applications.
- No Desktop Virtualization (e.g. Citrix Virtual Desktop Windows 10)
- No Linux OS
- Maximum 1 Display Head
- Maximum Resolution of 1280 x 1024
- No CUDA & OpenCL
NVIDIA GRID Virtual PC
NVIDIA Grid vPC is ideal for users who need a virtual desktop, to run client applications, browsers and HD video with increased graphics performance.
- Maximum 4 Display Heads
- Maximum Resolution per Display Head of 4096 x 2160 (by 2 Heads) or 2560 x 1600 (by 4 Heads)
- No CUDA & OpenCL
- No GPU Pass-through
- No Bare Metal Support
- Maximum Frame Buffer of 2 GB
NVIDIA Quadro Virtual Data Center Workstation
This edition is ideal for mainstream and high-end designers using powerful 3D applications, such as SOLIDWORKS, 3DExcite, Siemens NX or Autodesk Maya.
- Maximum 4 Display Heads
- Maximum Resolution per Display Head of 4096 x 2160
Example License Calculation
Example for 10 simultaneous user Quadro vDWS Perpetual License for 1 CCU over 3 years License cost for 1 user --> 450$ SUMSs for 3 years --> 300$ Total cost for 1 user --> 750$ For 10 user --> 7500$ Renew the SUMs after the 3 years for another year For 10 user --> 1000$ Total cost for 4 years --> 8500$ Total cost for 5 years --> 9500$
Example for 10 simultaneous user Quadro vDWS Annual Subscription for 1 CCU over 3 years License cost for 1 user for 3 years (incl. SUMs) --> 750$ For 10 user --> 7500$ Renew the license after the 3 years for another year For 10 user --> 2500$ Total cost for 4 years --> 10000$ Total cost for 5 years --> 12500$