NVIDIA
TESLA M60 Server 2016 Hyper-V VDI (RDVH)
Hi, I have a Tesla M60 installed on MS Server 2016 (Dell R730) utilizing Remote Desktop Virtualization Host VDI with 30 Windows 10 VM's. I was using RemoteFx in a 'para-virtualization' mode but decided to attempt DDA direct pass-thru as the applications we are running are very video processing intense (AutoDesk Suite), and need to use hardware acceleration. I got everything running (https://docs.microsoft.com/en-us/windows-server/virtualization/hyper-v/plan/plan-for-deploying-devices-using-discrete-device-assignment#machine-profile-script) with two VM's but couldn't add anymore, as the Windows host device manager only sees 2 GPU's. Using RemoteFx Hyper-V lets me associate 30 VM's (as the M60 has 2 cores w/ 16 vGPU's each). Am I limited to only using 2 vGPUs in pass-through mode? That can't possibly be correct (I hope). Has anyone else had experience with this? Thank you, Steve
Hi,

I have a Tesla M60 installed on MS Server 2016 (Dell R730) utilizing Remote Desktop Virtualization Host VDI with 30 Windows 10 VM's. I was using RemoteFx in a 'para-virtualization' mode but decided to attempt DDA direct pass-thru as the applications we are running are very video processing intense (AutoDesk Suite), and need to use hardware acceleration.

I got everything running (https://docs.microsoft.com/en-us/windows-server/virtualization/hyper-v/plan/plan-for-deploying-devices-using-discrete-device-assignment#machine-profile-script) with two VM's but couldn't add anymore, as the Windows host device manager only sees 2 GPU's. Using RemoteFx Hyper-V lets me associate 30 VM's (as the M60 has 2 cores w/ 16 vGPU's each).

Am I limited to only using 2 vGPUs in pass-through mode? That can't possibly be correct (I hope).

Has anyone else had experience with this?

Thank you,
Steve

#1
Posted 07/08/2018 07:37 PM   
[quote=""]Am I limited to only using 2 [s]vGPUs[/s] GPUs in pass-through mode? That can't possibly be correct (I hope).[/quote] Bad news, yes. DDA passthrough whole chip ("PCIe physical function"). NVidia vGPU is para-virtualization/mediated-virtualization and it is NOT supported for Windows based virtualization - [url]https://docs.nvidia.com/grid/latest/product-support-matrix/index.html[/url]. You must "wait" or change hypervisor (btw. RemoteFX is deprecated [url]https://docs.microsoft.com/en-us/windows-server/get-started/windows-server-1803-removed-features[/url]). M60 = 2x chip, M10 = 4x chip ([url]https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units[/url]). And do not forget to buy "Quadro vDWS" license for DDA ("GRID vPC" license for RemoteFX is insufficient) (see [url]https://images.nvidia.com/content/grid/pdf/161207-GRID-Packaging-and-Licensing-Guide.pdf[/url]).
said:Am I limited to only using 2 vGPUs GPUs in pass-through mode? That can't possibly be correct (I hope).

Bad news, yes. DDA passthrough whole chip ("PCIe physical function"). NVidia vGPU is para-virtualization/mediated-virtualization and it is NOT supported for Windows based virtualization - https://docs.nvidia.com/grid/latest/product-support-matrix/index.html. You must "wait" or change hypervisor (btw. RemoteFX is deprecated https://docs.microsoft.com/en-us/windows-server/get-started/windows-server-1803-removed-features). M60 = 2x chip, M10 = 4x chip (https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units). And do not forget to buy "Quadro vDWS" license for DDA ("GRID vPC" license for RemoteFX is insufficient) (see https://images.nvidia.com/content/grid/pdf/161207-GRID-Packaging-and-Licensing-Guide.pdf).

#2
Posted 07/08/2018 09:44 PM   
Thanks for your response mcerveny - I have a few questions... 1. I'm a bit confused because you said that NVidia vGPU is NOT supported for Windows based virtualization; however, the link you provided clearly says it supports RemoteFX for Win 08, 12, & 16. I am using Windows RDS in remote desktop virtualization host (RDVH) mode (as opposed to remote desktop session host mode-<RDSH>), and this is most definitely Windows based virtualization. 2. RemoteFX - I had no idea RemoteFX was being deprecated I appreciate the information. That being said, it does also say "We're developing new graphics acceleration options for virtualized environments", so clearly they still have intentions on supporting some type of pass-thru/para-virtualization option (or maybe even making it better). 3. The licensing is also confusing - is the provided link saying that for 30 users it would cost $7500.00 per year (30 x $250) for the annual licensing OR $13,500.00 one-time fee ($450 x 30) plus $100.00 annually for the perpetual licensing (using the Quadro vDWS license)?
Thanks for your response mcerveny - I have a few questions...

1. I'm a bit confused because you said that NVidia vGPU is NOT supported for Windows based virtualization; however, the link you provided clearly says it supports RemoteFX for Win 08, 12, & 16. I am using Windows RDS in remote desktop virtualization host (RDVH) mode (as opposed to remote desktop session host mode-<RDSH>), and this is most definitely Windows based virtualization.

2. RemoteFX - I had no idea RemoteFX was being deprecated I appreciate the information. That being said, it does also say "We're developing new graphics acceleration options for virtualized environments", so clearly they still have intentions on supporting some type of pass-thru/para-virtualization option (or maybe even making it better).

3. The licensing is also confusing - is the provided link saying that for 30 users it would cost $7500.00 per year (30 x $250) for the annual licensing OR $13,500.00 one-time fee ($450 x 30) plus $100.00 annually for the perpetual licensing (using the Quadro vDWS license)?

#3
Posted 07/11/2018 05:47 PM   
[quote=""]1. I'm a bit confused because you said that NVidia vGPU is NOT supported for Windows based virtualization; however, the link you provided clearly says it supports RemoteFX (eg. Microsoft RemoteFX vGPU) for Win 08, 12, & 16. I am using Windows RDS in remote desktop virtualization host (RDVH) mode (as opposed to remote desktop session host mode-<RDSH>), and this is most definitely Windows based virtualization. [/quote] NVidia vGPU is not Microsoft RemoteFX vGPU. NVidia vGPU is not DDA (passthrough to guest) or baremetal (use directly in host) (check "SW Product Deployment" in row "Hypervisor" == "Windows*"). You can use RemoteFX with baremetal and RDVH/RDHS/XenApp with DDA. You can study (new but useless marketing jabbering): [url]https://www.nvidia.com/en-us/design-visualization/technologies/virtual-gpu/[/url] [url]https://blogs.nvidia.com/blog/2018/06/11/what-is-a-virtual-gpu/[/url] Docs: [url]https://docs.nvidia.com/grid/index.html[/url] Presentation from GTC (the older the better): [url]http://on-demand-gtc.gputechconf.com/gtc-quicklink/9DCYSUn[/url] Check presentation/blogs/tweets/tools/guides from NGCA people (external people carefully selected by NVidia) that can enlighten the technology for you: [url]http://www.nvidia.com/object/vgpu-advisory-council.html[/url] [i]My opinion: Welcome to Microsoft, hypervisor vendors and NVidia licensing $$$ hell - check this forum from GRID2 [url]https://gridforums.nvidia.com/default/topic/543/general-discussion/grid-2-0/[/url]. All must be very happy that NVidia discounted license price upto 90% from GRID2 to GRID3 software release ([url]https://gridforums.nvidia.com/default/topic/543/general-discussion/grid-2-0/post/2201/#2201[/url] vs. [url]https://gridforums.nvidia.com/default/topic/543/general-discussion/grid-2-0/post/2718/#2718[/url]). NVidia say that it is not about price but about UX (but you get very fragile and crappy drivers and APIs like "vGPU scheduler", "capture sdk" and "video codec sdk" !) and about "embedded security" (data "should" reside in datacenter, but you must open connection to guest for remote protocol !). You can try to use Quadro >= x000 (like M2000, M4000, ... P2000, ... P5000) "should" support DDA but check other discussion that it is usually untested and very problematic. You can try to leave NVidia GPU and use AMD (MxGPU/SR-IOV). ... and do not forget to computed price per seat (HW+SW, CAPEX+OPEX) to be surprised ! [/i] That all from me for this topic.
said:1. I'm a bit confused because you said that NVidia vGPU is NOT supported for Windows based virtualization; however, the link you provided clearly says it supports RemoteFX (eg. Microsoft RemoteFX vGPU) for Win 08, 12, & 16. I am using Windows RDS in remote desktop virtualization host (RDVH) mode (as opposed to remote desktop session host mode-<RDSH>), and this is most definitely Windows based virtualization.

NVidia vGPU is not Microsoft RemoteFX vGPU. NVidia vGPU is not DDA (passthrough to guest) or baremetal (use directly in host) (check "SW Product Deployment" in row "Hypervisor" == "Windows*"). You can use RemoteFX with baremetal and RDVH/RDHS/XenApp with DDA.

You can study (new but useless marketing jabbering):
https://www.nvidia.com/en-us/design-visualization/technologies/virtual-gpu/
https://blogs.nvidia.com/blog/2018/06/11/what-is-a-virtual-gpu/

Docs: https://docs.nvidia.com/grid/index.html

Presentation from GTC (the older the better): http://on-demand-gtc.gputechconf.com/gtc-quicklink/9DCYSUn

Check presentation/blogs/tweets/tools/guides from NGCA people (external people carefully selected by NVidia) that can enlighten the technology for you: http://www.nvidia.com/object/vgpu-advisory-council.html

My opinion:
Welcome to Microsoft, hypervisor vendors and NVidia licensing $$$ hell - check this forum from GRID2 https://gridforums.nvidia.com/default/topic/543/general-discussion/grid-2-0/. All must be very happy that NVidia discounted license price upto 90% from GRID2 to GRID3 software release (https://gridforums.nvidia.com/default/topic/543/general-discussion/grid-2-0/post/2201/#2201 vs. https://gridforums.nvidia.com/default/topic/543/general-discussion/grid-2-0/post/2718/#2718).
NVidia say that it is not about price but about UX (but you get very fragile and crappy drivers and APIs like "vGPU scheduler", "capture sdk" and "video codec sdk" !) and about "embedded security" (data "should" reside in datacenter, but you must open connection to guest for remote protocol !).
You can try to use Quadro >= x000 (like M2000, M4000, ... P2000, ... P5000) "should" support DDA but check other discussion that it is usually untested and very problematic.
You can try to leave NVidia GPU and use AMD (MxGPU/SR-IOV).
... and do not forget to computed price per seat (HW+SW, CAPEX+OPEX) to be surprised !


That all from me for this topic.

#4
Posted 07/11/2018 08:06 PM   
Scroll To Top

Add Reply