Product Line: VERDE
There are two main types of GPU cards:
Those that support virtual GPU and those that do not.
For the virtual GPU cards - the ones that we have tested with VERDE:
There is a single physical GPU card that is plugged in to the server motherboard.
The card has enough hardware resources to support more than one user.
I guess CAD software likes to claim exclusive use of GPU resources. Without vGPU then one card can only be used by one user at a time.
So, NVIDIA developed vGPU - GRID technology - within their own firmware. They basically virtualized the GPU hardware resources so that more than one user can have access at one time.
It's kind of similar to VDI or Terminal Sessions with Windows. When you connect to a Windows Server using RDP terminal services it looks like you have complete control of the machine. But, you are sharing the resources with other users. Same principle.
With NVIDIA it seems the common setup is to support a maximum of 4 vGPU connections. This is probably the limit before the internal hardware on the card becomes overwhelmed.
So, with 4 vGPU "slots" (not physical slots) the single NVIDIA GPU card will look like 4 individual GPU cards to the end user. Kind of like a terminal services session connection to Windows Server.
Now - how to access those slots...
The more recent KVM and Linux kernels come with drivers that understand these GPU cards.
We have to run that special setup command noted in the KB so that VERDE becomes aware of the GPU card and the vGPU slots.
Each vGPU slot is assigned a unique identifier - kind of like a MAC address but not really. But, this allows the drivers to know which vGPU slot to use.
This vGPU unique identifier is what is selected when you configure the Gold Image - the Video Card Passthrough option:
If the setup script was run properly then you will see values in the Video Card Pass Through dropdown. Each value is the unique identifier for each vGPU slot.
You then configure the Gold Image to use a vGPU "slot" by selecting one of the options in the Video Card Passthrough.
The Gold Image must also include special drivers provided by the GPU card vendor (NVIDIA).
Finally - when a guest session starts using the Gold Image it will try to gain control of the vGPU slot defined above. If successful - then the session thinks it has exclusive access to a GPU card.
If another session has already grabbed the vGPU slot - then subsequent sessions won't have access to GPU.
Therefore, you have to define 4 separate Gold Images - each configured to use one of the available 4 vGPU slots or identifiers.
So careful planning is required when setting this up - you have to get the Gold Images configured properly and also Desktop Policy so that the users requiring GPU will all have their own Gold Image.