Windows 2016 Server with Grid K2
Hi All, I am currently testing Windows 2016 server with Grid K2. The latest drivers for K2 were provided by Nvidia support team, thank you. I have two Win10 Ent VM's with RemoteFX enabled. The server is cloud based with 1GB speeds up and down. The overall user experience is good when using the VM's but there seems to always be issues with video playback no matter what the resolution quality, sometimes audio is out of sync with video. I have played around with various VM settings on Hyper V with Remote FX set to different memory for the GPU and RAM with multiple vCPU's. I was told upto 32 users are able to share a K2 card in VM environment. What is the expectation for video playback in Windows 2016 and K2? Also my understanding is that when using RemoteFX, 32 users approx can be supported through the card. However for GPU pass through it is dedicated only to one VM, is that correct. I am trying to understand if both GPU pass through also be used on various VM's at the same time? Thanks
Hi All,

I am currently testing Windows 2016 server with Grid K2. The latest drivers for K2 were provided by Nvidia support team, thank you.
I have two Win10 Ent VM's with RemoteFX enabled.
The server is cloud based with 1GB speeds up and down.
The overall user experience is good when using the VM's but there seems to always be issues with video playback no matter what the resolution quality, sometimes audio is out of sync with video. I have played around with various VM settings on Hyper V with Remote FX set to different memory for the GPU and RAM with multiple vCPU's.
I was told upto 32 users are able to share a K2 card in VM environment. What is the expectation for video playback in Windows 2016 and K2?
Also my understanding is that when using RemoteFX, 32 users approx can be supported through the card. However for GPU pass through it is dedicated only to one VM, is that correct.
I am trying to understand if both GPU pass through also be used on various VM's at the same time?

Thanks

#1
Posted 04/06/2017 04:25 PM   
Hi I'm not the right person to help you with Remote FX (Never used it). However, Page 4 of this guide will help you with K2 profiles: http://images.nvidia.com/content/grid/pdf/GRID-vGPU-User-Guide.pdf Each K2 will support a maximum of 16 [u]individual[/u] sessions (8 sessions per GPU of which there are 2 per board), but each session will only have 512MB of frame buffer. So for Windows 10 / Server 2016, this won't be really be good enough. For Windows 10, ideally nothing less than a K240Q profile (1GB). With 2x K2 boards in your server (4x GPUs) you'll get 16x Windows 10 sessions. If you want more than that, then you're best off using RDS with the GPUs in Passthrough. The term "Passthrough" is where you pass the physical device (in this case a GPU) through the Hypervisor and into a VM. When you do this, no other VMs can use that piece of hardware. It is a 1:1 ratio. With the Kepler GPUs, Passthrough does offer a little extra performance, access to CUDA and a higher resolution (up to 4K). You also do not need to install a driver in the Hypervisor. With newer architectures, the benefits of Passthrough are slightly less. As an example, with a single K2, you could run 1x GPU in Passthrough and assign it to a RDS host, and the other GPU you could slice in to 4x K240Q Profiles. Don't forget, you still need to balance out CPU, RAM and Disk IO between all your VMs, despite the use of a GPU. Regards
Hi

I'm not the right person to help you with Remote FX (Never used it). However, Page 4 of this guide will help you with K2 profiles: http://images.nvidia.com/content/grid/pdf/GRID-vGPU-User-Guide.pdf


Each K2 will support a maximum of 16 individual sessions (8 sessions per GPU of which there are 2 per board), but each session will only have 512MB of frame buffer. So for Windows 10 / Server 2016, this won't be really be good enough.

For Windows 10, ideally nothing less than a K240Q profile (1GB). With 2x K2 boards in your server (4x GPUs) you'll get 16x Windows 10 sessions. If you want more than that, then you're best off using RDS with the GPUs in Passthrough.

The term "Passthrough" is where you pass the physical device (in this case a GPU) through the Hypervisor and into a VM. When you do this, no other VMs can use that piece of hardware. It is a 1:1 ratio.

With the Kepler GPUs, Passthrough does offer a little extra performance, access to CUDA and a higher resolution (up to 4K). You also do not need to install a driver in the Hypervisor. With newer architectures, the benefits of Passthrough are slightly less.

As an example, with a single K2, you could run 1x GPU in Passthrough and assign it to a RDS host, and the other GPU you could slice in to 4x K240Q Profiles.

Don't forget, you still need to balance out CPU, RAM and Disk IO between all your VMs, despite the use of a GPU.

Regards

#2
Posted 04/07/2017 04:09 PM   
Hi Thank you for the reply, appreciate the help. I do not really see the point of GPU pass through if it can only be used on one VM, expensive solution. Maybe there are some use cases for it. My aim is to deliver win10 VM's using K2 Cards on Win2016 server using Hyper V. Overall experience is good but i find video playback can sometimes be bit hit or miss and that is the bit I am trying to fix if possible. I know video playback of any sort over RDP has always has issues but Windows 2016 RemoteFx has improved. Thanks
Hi

Thank you for the reply, appreciate the help. I do not really see the point of GPU pass through if it can only be used on one VM, expensive solution. Maybe there are some use cases for it. My aim is to deliver win10 VM's using K2 Cards on Win2016 server using Hyper V. Overall experience is good but i find video playback can sometimes be bit hit or miss and that is the bit I am trying to fix if possible.
I know video playback of any sort over RDP has always has issues but Windows 2016 RemoteFx has improved.

Thanks

#3
Posted 04/09/2017 12:18 PM   
Scroll To Top