NVIDIA
Deploying GRID without use of Citrix/VmWare
Hi guys, we are experimenting with a GRID deployment but without the use of Citrix or VMware. Our first aim is to simply passthrough the Tesla Grid M10 card to 4 different Vms. Hypervisor: WS2012 Std Vm: Win 8.1 Ent 64bit Card: Nvidia Tesla M10 Hw: PowerEdge T620 All software dependencies have been checked & installed, Grid driver installed to the Hyper-V, RemoteFX video card has been added to the VM in the Hyper-V Manager, but I cannot install the GRID driver onto the VM. It shows the "Microsoft RemoteFX Grapics Device - WDDM" in Device Manager with the exclamation mark, however if I try to install the proper driver pack "369.71_grid_win8_win7_server2012R2_server2008R2_64bit_international" it tells me that "This graphics driver could not find compatible hardware". What could be the problem here? Is there a solution for deploying without ESXi or Xen? Thanks for the help mates! Regards, Mike
Hi guys,


we are experimenting with a GRID deployment but without the use of Citrix or VMware. Our first aim is to simply passthrough the Tesla Grid M10 card to 4 different Vms.

Hypervisor: WS2012 Std
Vm: Win 8.1 Ent 64bit
Card: Nvidia Tesla M10
Hw: PowerEdge T620

All software dependencies have been checked & installed, Grid driver installed to the Hyper-V, RemoteFX video card has been added to the VM in the Hyper-V Manager, but I cannot install the GRID driver onto the VM.

It shows the "Microsoft RemoteFX Grapics Device - WDDM" in Device Manager with the exclamation mark, however if I try to install the proper driver pack "369.71_grid_win8_win7_server2012R2_server2008R2_64bit_international" it tells me that "This graphics driver could not find compatible hardware".

What could be the problem here? Is there a solution for deploying without ESXi or Xen?
Thanks for the help mates!

Regards,
Mike

#1
Posted 12/17/2016 03:28 PM   
[list] [.]The remoteFX solution is not passthrough solution. "Microsoft RemoteFX Grapics Device - WDDM" is not NVidia device and NVidia drivers does not attach.[/.] [.]The Grid VM driver is for vGPU (eg. different "emulated" pci-id supported only in VMware/Xenserver hypervisor VM solution).[/.] [.]There should not be needed driver in hypervisor for passthrough (only configure GPU as passthrough to VM and GPU "disappears" from any hypervisor). [/.] [/list] You are probably looking for "[b]Discrete Device Assignment[/b]" for GPU that is supported from WS2016 ([i]you should search the web[/i]). If you are able to successfully passthrough device then you should use standard driver for passthrough or direct use from standard download location [url]http://www.nvidia.com/Download/index.aspx?lang=en-us[/url] -> "GRID" -> "GRID Series" -> ... -> 376.33-quadro-grid-desktop-notebook-win8-win7-64bit-international-whql.exe and you must buy and install NVidia license for Tesla M10 passthrough (see "GRID LICENSING"). ([b]Update:[/b] I look to nv_dispwi.inf but support [b]Tesla M10 is missing in driver package[/b]. The package from licensing portal should be probably used for passthrough.) You can also use Linux KVM or Linux/FreeBSD XEN (besides XenServer) for passthrough. [i]My opinion: There was cheaper solution for [b]Maxwell[/b] generation multiple chip passthrough card - GRID M40 (not Tesla M40!) card. It was the same board as M10 but half of memory (eg. 4GB per GPU, larger RAM for Tesla M10 is about $100 additional chip cost per whole card then GRID M40, see [url]https://gridforums.nvidia.com/default/topic/823/grid-boards/tesla-m10/[/url]). But NVidia deny existence of GRID M40 board (never design, build, sell or support, but check Ebay). There was supported paired board solution (for vGPU and for passthrough) in previous [b]Kepler[/b] generation - K1 ~ K340 (more SM/less memory) and K2 ~ K520. Alternatively you can use middle and higher Quadro models that are passthrough ready ([url]http://www.nvidia.com/object/dedicated-gpus.html[/url] + [url]http://www.nvidia.com/object/sli_multi_os.html[/url] - NVidia does not disclosure updated list for passthrough support but there is HCL for example for XenServer [url]http://hcl.vmd.citrix.com/gpus/[/url] (M10 is not officially certified for passthrough now)).[/i]
  • The remoteFX solution is not passthrough solution. "Microsoft RemoteFX Grapics Device - WDDM" is not NVidia device and NVidia drivers does not attach.
  • The Grid VM driver is for vGPU (eg. different "emulated" pci-id supported only in VMware/Xenserver hypervisor VM solution).
  • There should not be needed driver in hypervisor for passthrough (only configure GPU as passthrough to VM and GPU "disappears" from any hypervisor).

You are probably looking for "Discrete Device Assignment" for GPU that is supported from WS2016 (you should search the web). If you are able to successfully passthrough device then you should use standard driver for passthrough or direct use from standard download location http://www.nvidia.com/Download/index.aspx?lang=en-us -> "GRID" -> "GRID Series" -> ... -> 376.33-quadro-grid-desktop-notebook-win8-win7-64bit-international-whql.exe and you must buy and install NVidia license for Tesla M10 passthrough (see "GRID LICENSING"). (Update: I look to nv_dispwi.inf but support Tesla M10 is missing in driver package. The package from licensing portal should be probably used for passthrough.)


You can also use Linux KVM or Linux/FreeBSD XEN (besides XenServer) for passthrough.

My opinion: There was cheaper solution for Maxwell generation multiple chip passthrough card - GRID M40 (not Tesla M40!) card. It was the same board as M10 but half of memory (eg. 4GB per GPU, larger RAM for Tesla M10 is about $100 additional chip cost per whole card then GRID M40, see https://gridforums.nvidia.com/default/topic/823/grid-boards/tesla-m10/). But NVidia deny existence of GRID M40 board (never design, build, sell or support, but check Ebay). There was supported paired board solution (for vGPU and for passthrough) in previous Kepler generation - K1 ~ K340 (more SM/less memory) and K2 ~ K520.
Alternatively you can use middle and higher Quadro models that are passthrough ready (http://www.nvidia.com/object/dedicated-gpus.html + http://www.nvidia.com/object/sli_multi_os.html - NVidia does not disclosure updated list for passthrough support but there is HCL for example for XenServer http://hcl.vmd.citrix.com/gpus/ (M10 is not officially certified for passthrough now)).

#2
Posted 12/17/2016 03:56 PM   
You will need an evaluation license for the M10, there is a hyper-V/server 2016 one available within the portal. Licensing would then depend on how you delivered it RemoteFX is a vPC license and RDSH vApps if no quadro needs. This article contains details of where to get M10 drivers: http://nvidia.custhelp.com/app/answers/detail/a_id/4120/~/where-to-find-the-drivers-for-nvidia-grid-2.0-vgpu-and-up,-including-m60,-m10
You will need an evaluation license for the M10, there is a hyper-V/server 2016 one available within the portal. Licensing would then depend on how you delivered it RemoteFX is a vPC license and RDSH vApps if no quadro needs.

This article contains details of where to get M10 drivers: http://nvidia.custhelp.com/app/answers/detail/a_id/4120/~/where-to-find-the-drivers-for-nvidia-grid-2.0-vgpu-and-up,-including-m60,-m10

#3
Posted 12/19/2016 11:17 AM   
There do seem to be some community set up guides around: https://blog.workinghardinit.work/2016/04/11/discrete-device-assignment-in-windows-server-2016-hyper-v/

#4
Posted 12/19/2016 11:20 AM   
Hi Mcerveny! Thanks for the quick reply mate! We've installed a WS2016 and assigned (DDA) one M10 instance to the VM (Win10Ent, RDP 10.2) and it works. However the performance isn't that great, I'll have to find out why. We have logged in with a Windows 8 (RDP 8.1) onto the Win10 but the rendering it's somewhat laggy. I'll have to start reading blogs and posts guess so:) As per your M40 theory: you're probably right, the card identifies itself as a M40 in the DELL iDRAC when listing the PCI devices. Regards, MIke
Hi Mcerveny!

Thanks for the quick reply mate! We've installed a WS2016 and assigned (DDA) one M10 instance to the VM (Win10Ent, RDP 10.2) and it works. However the performance isn't that great, I'll have to find out why.

We have logged in with a Windows 8 (RDP 8.1) onto the Win10 but the rendering it's somewhat laggy. I'll have to start reading blogs and posts guess so:)

As per your M40 theory: you're probably right, the card identifies itself as a M40 in the DELL iDRAC when listing the PCI devices.

Regards,
MIke

#5
Posted 12/19/2016 11:28 AM   
The performance problem can be "license" problem. Unlicensed M10/M60/M6 (in "GRID Virtual Workstation" mode) drop performance (~3FPS) and limit maximum resolution and number of attached "monitors" in [b]previous[/b] versions of software. It is now unclear how it works with new drivers in passthrough mode. You can try "eval license" as RachelBerry wrote and study [url]http://images.nvidia.com/content/pdf/grid/guides/GRID-Licensing-Guide.pdf[/url].
The performance problem can be "license" problem. Unlicensed M10/M60/M6 (in "GRID Virtual Workstation" mode) drop performance (~3FPS) and limit maximum resolution and number of attached "monitors" in previous versions of software. It is now unclear how it works with new drivers in passthrough mode. You can try "eval license" as RachelBerry wrote and study http://images.nvidia.com/content/pdf/grid/guides/GRID-Licensing-Guide.pdf.

#6
Posted 12/19/2016 01:22 PM   
Scroll To Top

Add Reply