[Bug]: Weird issue with multiple VMs

[Tobi-La]

Which OS?

Win11 Pro

Which release?

24H2

Describe the bug

Its seems that there is an issue with the interaction of virtual display and my GPU which is passed through to the VM via VMGpuAdapters. I have currently 6 VMs active because I am working on an Environment for some kind of Splitscreen project to play Games with my friends on my projector.

I have an NVidia RTX 4090 and Windows 11 Pro VMs and connect to them via Moonlight / Sunshine.

It works with nearly no issues up to 4 players, I can start the game without issues. When trying to connect to the fifth instance I get the error from moonlight "Failed to initialize video capture / encoding. Is a display connected and turned on? (Error 503)". It works when VM 1-4 don't have a game running or I stop one game at least on one of them.

After a lot of attempts I can connect, but explorer.exe is restarting sometimes, and I can't start games anymore. Wreckfest returns "D3D11CreateDeviceAndSwapChain() failed" and Fall Guys "There is a problem with your graphics card. Please ensure your card meets the minimum system requirements and that you have the latest drivers installed.". So I guess this VM has issues to access my graphics card somehow.

My host pc isn't struggling, CPU at 50%, RAM at 45% and GPU at 40%. VM gpu is at 40%, cpu at 2, ram at 57%.

When I deactivate VirtualDisplay Moonlight uses the default HyperVMonitor Display, I can start games on all 6 VMs at the same time, no issues, even up to VM6.

Thanks for this amazing tool, appreciate your work!

Steps to reproduce

Have some workload on other VMs on the gpu VirtualDisplay is connected to.

Expected behavior

Games should run like they do with the default HyperVMonitor.

Log File

I activated log+debug log and logging for all gpus, but nothing concerning the game crashes is in the logs:
[2025-01-08 00:32:55] [COMPANION] Logging Enabled
[2025-01-08 00:32:59] [PIPE] Client Connected
[2025-01-08 00:32:59] [PIPE] Client Handling Enabled
[2025-01-08 00:32:59] [PIPE] LOG_DEBUG true
[2025-01-08 00:32:59] [COMPANION] Pipe debugging enabled
[2025-01-08 00:32:59] [DEBUG] Debug Logs Enabled
[2025-01-08 00:33:06] [PIPE] Client Connected
[2025-01-08 00:33:06] [PIPE] Client Handling Enabled
[2025-01-08 00:33:06] [PIPE] GETALLGPUS
[2025-01-08 00:33:06] [COMPANION] Logging all GPUs
[2025-01-08 00:33:06] [INFO] Any GPUs which show twice but you only have one, will most likely be the GPU the driver is attached to
[2025-01-08 00:33:06] [COMPANION] GPU Name: NVIDIA GeForce RTX 4090 Memory: 24142 MB
[2025-01-08 00:33:06] [COMPANION] GPU Name: NVIDIA GeForce RTX 4090 Memory: 24142 MB
[2025-01-08 00:33:06] [COMPANION] GPU Name: NVIDIA GeForce RTX 4090 Memory: 24142 MB
[2025-01-08 00:33:06] [COMPANION] GPU Name: Microsoft Basic Render Driver Memory: 0 MB
[2025-01-08 00:33:06] [COMPANION] Logged all GPUs

Contact Details

[email protected]

[timminator]

I'm running into a similar issue I think. For me the issue is not exactly VirtualDisplayDriver: When I have more than 4 VM's with GPU-Support started my whole system becomes laggy, even the host.
I think this is not an issue related to the VirtualDisplayDriver. One thing i noticed: When setting up my VM's I set the GPUResourceAllocationPercentage to 25 percent for each VM - maybe its an overallocation issue?
What script did you use to set up your VM? Was this parameter in your setup by chance also at 25%?

[Tobi-La]

Jeah, I realized that too. Played around with the VMGPUPartitionAdapter values, but couldn't notice any difference. I am not really sure why it worked for me a few times using the hyperv virtual display adapter, because a few days later it didn't anymore. But it seems to work far more reliably when using the hyperV manager connection instead of sunshine / moonlight like I do. Investigating further if its something maybe on the encoding / decoding site of the setup. Thought for a short time this would be the solution, too, but sadly that didn't help:
https://github.com/keylase/nvidia-patch