How are GPU cores allocated for each vGPU Profile?

Virtual GPU Forums More vGPU Forums NVIDIA Virtual GPU Technology
1

Hi All,

I have been searching for some official documentation that describes how the GPU cores are allocated/shared in vGPU profiles. I can easily find confirmation that the video RAM is dedicated for each virtual desktop relative to its vGPU profile.

How are the GPU cores allocated for each vGPU profile? Are the cores divided up and then dedicated to each virtual desktop? Or, is access to the GPU cores scheduled like access to physical CPU cores on a hypervisor?

Also, has anyone experienced some of the vGPU profiles missing? I have servers with K1 and K2 cards. Our servers are lacking the K220Q and K120Q profiles. All the other vGPU profiles listed on this page are present. Virtual GPU (vGPU) Monitoring and Management | NVIDIA

Thank You!

Richard

2

The GPU cores are not specifically allocated like the memory is. You are correct in that the GPU cycles are scheduled just like the CPU cycles are scheduled on a hypervisor. The Nvidia vGPU Manager is responsible for this scheduling.

I can’t comment on the vGPU profiles being missing though.

-Mike

3

Thanks very much, Mike. So if an individual GRID GPU had 100 cores (for example), one VM would get access to all 100 cores for a time-slice, then the next VM, and so on?

Best,

Richard

4

Hi Richard,

Mike’s description is correct, resources are effectively "timesliced" by the scheduler. The vGPU profile gives you a guaranteed level of resource i.e. 25% for a 240Q, 50% for 260Q etc and if additional resources are available because the other VM’s aren’t using the GPU, that can also be utilised by the running VM.

As for not being able to see the x20Q profiles, are you using the latest vGPU manager and have you patched XenServer too? There’s a couple of patches needed to make the new profile visible even if you’ve installed the latest vGPU manager.

5

Thanks Jason.

Yes, I am fully patched on XenServer and the GRID manager. I was speaking with a technical rep from NVIDIA at VMWorld and he said that the missing profiles had been deprecated. We looked at this page together, Virtual GPU (vGPU) Monitoring and Management | NVIDIA. I’ll list what I see and what is missing below.

K1:
-passthrough
-K140Q
-K120Q (missing)
-K100

K2:
-passthrough
-K260Q
-K240Q
-K220Q (missing)
-K200

Does this look correct to you?

Thanks again!

Richard

6

I double checked the XenServer patch level and it did in fact need patch XS62ESP1009. I thought we were fully up to date. After updating the hosts, the additional K120Q and K220Q vGPU profiles were present. Thanks for suggesting that, Jason. That also answered my question that these profiles are not deprecated but are actually, as you noted, new profiles.

7

Glad you got it resolved, and thanks for letting us know.