Hi, with Server 2025 out, has anyone tried partitioning consumer nvidia GPU’s to a VM using powershell/Hyper-V?
My card shows up as being “partition-able”, but when I run Get-VMHostPartitionableGpu, I get:
Name : \?\PCI#VEN_10DE&DEV_25B6&SUBSYS…….
ValidPartitionCounts : {32}
PartitionCount : 32
I tried doing something research on what that means, but didn’t have much luck. Does it mean the only valid partition count is 32 and I’m already using 32 on the host and hence there are no excess resources to pass to a VM?
I’ve not tried assigning the GPU to a VM on the fear that if I did, it will crash my system and I won’t be able to undo the change (eg. no resources to run the host if the settings are saved on reboot)
Has anyone tried this with success? My goal is to pass 50% of the cards capacity to a VM and keep 50% on the host.
TL;DR - If you’re not an enterprise customer, it’s going to cost you an arm and a leg. I think you’re better off with pass-through if you want to handover the GPU to a VM. Read below!
A sidetrack to your answer
A comment on complete pass-through: If you have an APU - you can handover the GPU to the VM with no issues. BUT ONLY if you have an APU, otherwise you’d have no means to see the screen. You’d need to connect your display to your APU port (The motherboard’s HDMI/DP port)
Not the answer you want
Hyper-V does support GPU partitioning, and there are means and guides on how to do this on a local Windows Server. You’d need the Nvidia vGPU drivers. But the big catch is the licensing… partitioning an Nvidia GPU costs money (Refer to this pricing/licensing guide) and requires specific GPUs (as far as I know of):
This situation gives me flashbacks to SQL Server CPU licensing… Wendell did a video on how Intel and their Flex 170 GPU is actually better suited for partitioning as it does not have any licensing requirements compared to Nvidia.
Thank you…if it makes any difference, I’m only interested in having access to some of the CUDA cores. I don’t need video output. I run a software call HQPlayer for music and it greatly benefits with the CUDA processing power. Now, I can run HQPlayer on the host as well and it works okay, but every once a while a get dropouts in the music because other software on the host is pushing the video card hard and so I figured if I can limit the use of the card by the other software, and “reserve” it by way of passing x% to the vm, I will avoid the dropouts.
Unfortunately it makes no difference, because of what you want (Partitioning). Nvidia REQUIRES you to license the drivers for partitioning. Sadly, I don’t think it’s feasible for you to pursue partitioning at no cost.
I’d been thinking about this overnight, and took a read about what HQPlayer can do. There are a couple of scenarios that you can do:
The Virtual CPU method: HQPlayer can support CPU processing by default, you may want to try to set up a VM and run it on CPU mode. That way you can actually assign it dedicated resources to prevent it from eatingup other cores/RAM, which may prevent some stuttering
The pre-computed method: Since HQPlayer is meant for upsampling, why don’t you do the upsample of all your songs in advance? (i.e. you convert them to a higher bitrate through HQPlayer or other means) This would be very dependent on how much storage you have, and what bitrate you’re comfy with (I am basic and MP3 320Kbps is fine for me).
Another GPU!: You can add a lesser GPU to your system, and perform a complete passthrough of this GPU to a VM. This would require another Nvidia GPU (costly and HQPlayer only does CUDA offload), and an HDMI dummy adapter/plug and potentially a beefier power supply.
This is all I can think of that may resolve your issues. However, this all starts from the premise that you know that it’s the GPU that’s being impacted and it’s the culprit of the stuttering. You can run HQPlayer with Furmark (GPU) or Prime95 (CPU) as a test to see if the stuttering happens with either to narrow down exactly what’s causing the stuttering.
Thank you. Unfortunately, HQPlayer needs CUDA offload (vs. CPU) for the higher upsampling options and I don’t have space in my box for another GPU and then there’s the expense of buying another 4070 Ti Super.
Yes, HQPlayer can pre-compute and save, put that not very feasible in practice since storing the files will be difficult given a 10k+ song library