Could anyone recommend a setup for delivering 6 to 10 concurrent graphics-accelerated RDSH Desktop Sessions (Windows Server 2019) which are intended to watch 720p videos with media player from the WAN. No cinema quality is needed; the users should be able to understand what’s happening. Videos show 2 people sitting and talking in a room, no big movements.
- One Tesla M10 assigned (pass-through) to 4 RDSH VMs on Hyper-V (DDA) with vApps Licenses?
- One Tesla T4 assigned (pass-through) to one RDSH VM on Hyper-V (DDA) with vApps Licenses?
Are the vApps licenses the right choice, even if desktops have > 1280×1024 resolution?
3. Bare-Metal RDSH with Quadro P4000 or M4000 (no licenses?)
Your help is very much appreciated as I didn’t quite get the answer from the other topics.
Yes, vApps licenses are the correct choice for the M10 and T4.
If it were me, I’d use the T4 in Passthrough. It’s a much more modern GPU with better encoding / decoding capabilities, and has plenty of Framebuffer for a single VM with multiple users.
If you wanted to skip the vApps licensing (not that it’s expensive anyway), you could use an RTX4000 (instead of the M4000 or P4000). It has the same benefits of being a more modern GPU with better capabilities than the earlier architectures. No need to do that bare-metal, you can still run that in a VM as Passthrough. The downside to those GPUs, is they only have 8GB of Framebuffer, whereas the T4 has 16GB if you ever needed it.
Framebuffer is typically the limiting factor when running RDSH when maximum user density is reached.
thank you very much!
I have one more question concerning CPU requirements when using GPU in this case:
Is it better to use more cores with lower speed (e.g. Xeon-Gold 6240, 2.6 GHz) or less cores with higher speed (e.g. Xeon-Gold 6234, 3.3Ghz)?
We would prefer lower speed / more cores
How many vCPUs would you assign to the T4-assigned VM with 10 video users?
(I know the question does not directly concern the GPU but this might be the best place to ask anyway…)
Typically you’d want a higher Clock as that helps the entire platform, not just the VM, but you also need to take into account what else is running on the VM and also Hypervisor, so it’s always a trade off. If your system isn’t fully loaded, you could always populate just a single CPU Socket and use a Xeon Gold 6254 which has 18 Cores @ 3.1Ghz. Then if you need more CPU at some stage, buy another and scale up.
Start with 8 vCPUs, 16GB RAM and a T4 and scale resources up / down accordingly. Storage wise, you should be running Flash, not spinning disk. If you wanted a cheap, fast, storage solution, you can run M.2 drives on a PCIe carrier board.
Could you be so pleasure to explain how to divide one T4 onto 8 vGPU under Windows Datacenter Server 2019 with Hyper-V role installed to passthrough them into VMs under such server?
License server is installed.
Hyper-V doesn’t support vGPU. Passthrough (DDA) is your only option.
Depending on how many and also the type of graphics workloads you may want to run in the future, you would do well to look at changing your Hypervisor for those workloads. If you only have RDSH workloads, then sticking with Hyper-V and Passthrough may be sufficient, but if you need to do anything more than that, then you’ll need to change Hypervisors.
Hyper-V is currently the least favourable Hypervisor for graphics workloads. vSphere, XenServer, Acropolis and even KVM would all be a more suitable choice at the moment should you wish to look at changing.
When enabled on Tesla GPUs, licensed editions of GRID Virtual Workstation or Virtual PC are activated by obtaining a license over the network from an NVIDIA GRID License Server. The license is “checked out” or “borrowed” at the time the Virtual Machine (VM) is booted, and returned when the VM is shut down.