How does Horizon View share the GPU between VMs?

[kappyninewood]

Say I have three hosts, and only two have the Nvidia GPU, what happens when I vmotion a VM to the host without the GPU?  Does that VM lose the capability to use the GPU?

[LMUISHuntsville · Accepted solution]

Short answer is yes (hardware GPU).

If you have the "3D renderer" option in the view pool settings set to "Automatic" the VM will still be able to power up, but it will be using a software rendered GPU (not hardware like your other hosts with a GPU installed).  If the "3D renderer" setting is set to "Hardware" the VM will not power on when trying to run off of the host with no GPU and I believe that it may not even vmotion over to the host with no GPU installed (there should be a check to make sure if the VM is compatible with the host its trying to vmotion to).

Regards,

Nathan

[waynej]

How does one capacity plan with the new GPU functionality?  For example, say I buy a new Dell PowerEdge R720 w/ nVidia GRID K1.  The K1 has 4 GPUs and 16GB RAM.  The R720 itself can have 384GB RAM in a reasonable price config using 16GB dimms.  If I deploy Windows 7 VMs with 4GB RAM each for AutoCAD (or whatever), would that much system RAM be a complete waste as the GRID K1 could only support 32 (see below) GPU-enabled VMs?

For my example, I assumed the number of VMs would be limited by the 16GB GPU memory split between VMs configured with 512MB video memory.

[RSharma20111014]

that's a great question Waynej. I hope some one can answer. But i think besides running 32 VMs which requires GPU. You can run more VM's for regular users who just need a simple VM without any GPU on the same host.

[Linjo]

Good questions and answers.

Only half of the configured vRAM will be allocated from the GPU, so if you have 1 K1 with 16GB memory then you can have max 64 users. (256MB x 64 = 16)

But with CAD users you probably saturate the GPU much earlier then that, see this recent performance study:

http://www.vmware.com/files/pdf/view/vmware-horizon-view-hardware-accelerated-3Dgraphics-performance...

Since there is no overcommit and GPU utilization is not part of DRS (for now) I usually recommend to build clusters specifically for these workloads since they tend to be more demanding then regular users when it comes to vCPU:s and IOPS anyway.

// Linjo