Within the world of applications for Linux there are a large number of scientific applications that make use of the unique functions of GPUs, but we do not usually have access to it if we run a virtual one, unless we are combining a dedicated graphics card with a integrated.
What is virtualization on a GPU?
The command processors of a GPU can normally only execute a list of graphical commands to display the next frame on the screen, which is a problem under two conditions. The first is when we have mounted a server in the cloud and we have several clients to which we must grant a graphical interface in real time to each one of them, the second is when we run one or more virtual machines on an operating system.
Virtualization requires the use of a custom command processor for this, there are systems such as Microsoft’s Xbox One that run two environments at the same time, the game environment and the console menus, it uses two command processors at the same time. weather. In other cases, special versions of the GPUs are created with the modified command processor, especially with support for SR-IOV, but it must be taken into account that these GPUs are not sold for desktop PCs.
That is why support for virtualization on desktop GPUs is something that is appreciated, especially if you have a system that lacks an integrated GPU and you cannot assign a second GPU to the virtual machine, having to run an emulated VGA.
Limited virtualization on NVIDIA GeForce GPUs
Starting with the R465 driver, the following GeForce series GPUs support virtualization:
- Desk: Kepler (GTX 700 and GTX 800), Maxwell (GTX 900), Pascal (GTX 1000), Turing (RTX 2000 and GTX 1600) and Ampere (RTX 3000).
- Laptops: Maxwell (GTX 800M and GTX 900 M), Pascal (GTX 1000M), Turing (RTX 2000 and GTX 1600) and Ampere (RTX 3000).
However, since it is not a hardware implementation but through the driver we are limited to a virtual machine and the SR-IOV is not active. So if you want to have virtualization for several virtual machines you will have to acquire a card from the professional range of NVIDIA such as the RTX Enterprise, Tesla or Quadro.
So if when you started reading the news you were planning to create a server in the cloud with GeForce graphics cards, forget about it, since this NVIDIA solution through its drivers is limited. In any case, given NVIDIA’s tendency to unify its architectures in a single design along with the growth of Cloud Gaming, especially if we take into account its GeForce Now, which helps NVIDIA to reposition its unsold graphics cards for Cloud Gaming.