How to install Nvidia Tesla K10 in Vmware ESXI 6

[Suvajit]

Hello

I have purchased Cisco C240 Server with Nvidia Tesla K10 and install Vmware ESXI 6. How to install Nvidia Tesla K10 and share n number host?

Please help me about this.

[virtualg_uk]

What's your plan for the Tesla K10's as I do not believe that they are supported by VMware yet: VMware Compatibility Guide - Shared Pass-Through Graphics

[adamwiso]

Cannot comment on the Tesla cards - but if they are anything like the K1/K2 cards we have pretty straight forward.  Install those bad boys in the server themselves.  Then pump the VIB into the hypervisor to expose the acceleration to the hypervisor (you can then see the graphics profiles in the web client once installed).  This will also expose the different vGPU profiles which affect your frame buffer and VM per GPU (once again your mileage may vary on those cards).    Finally install the guest drivers to the guests themselves. 

Hope it helps.

[ArrowSIVAC]

I have a similar issue here and did not see this addressed but figured I would post in this thread as it may address the original posting.

I have a customer who wants to accelerate a software package (IBM DLE – Deep Learning Engine)  that can take advantage of the Tesla GPU.  It does not need a display, just the GPU functions.

I have a Tesla K10. I have upgraded to 6.5

After doing some more research I am not sure if the K10 will work. The Tesla K10 is a Kepler chip family.

The support and driver option for nVidia are mostly focused on Display rendering VDI acceleration vs rendering (such as Tesla).  The Tesla card which does have a driver is for vSphere 6.5 and is the Maxwell chip family.

http://www.vmware.com/resources/compatibility/search.php?deviceCategory=sptg&productid=41690&deviceC...

When I upgrade the host from 6.0U2 to 6.5 which has a Tesla driver. It still shows to host as a GK107GL (GRID K1)

vSphere Device List

[root@nxch101:~] lspci -v | grep NVIDIA 0000:0d:00.0 VGA compatible controller Display controller: NVIDIA Corporation NVIDIAGRID K1 [vmgfx0] 0000:0e:00.0 VGA compatible controller Display controller: NVIDIA Corporation NVIDIAGRID K1 [vmgfx1] 0000:0f:00.0 VGA compatible controller Display controller: NVIDIA Corporation NVIDIAGRID K1 [vmgfx2] 0000:10:00.0 VGA compatible controller Display controller: NVIDIA Corporation NVIDIAGRID K1 [vmgfx3] [root@nxch101:~] esxcli software vib update -v /vmfs/volumes/LabFiles/Lenovo/dx360/NVIDIA-kepler-VMware_ESXi_6.5_Host_Driver_367.64/NVIDIA-kepler-VMware_ESXi_6.5_Host_D river_367.64-1OEM.650.0.0.4240417.vib

Based on this: https://en.wikipedia.org/wiki/Nvidia_Quadro#Quadro the GK107GL is one of the Quadro models K420/K600/K2000/K2000D.

Its almost like their are several issues here. 

1) vSphere cannot pass through the Tesla card via a PCI-E slot direct to a VM, until the hypervisor itself has a driver.

2) Their is not a Keplar compiled driver for vSphere of any version

3) The card it can "emulate" is a GK107GL (K1) card as that is the closest driver it can find and pass through. As such, for using this for computational rendering, ...

               a) Will this even work for GPU rendering?

               b) Do I pass through to the VM all the virtual cards, or for best performance just the first one (which is what I currently have set)

               c) What other acceleration tuning can I do for this system?