> UI is sluggish

You absolutely need to pass through a GPU so that DWM.exe is properly accelerated; otherwise, it falls back to the software-accelerated WARP and the performance tanks to ~15 FPS.

It doesn't need to be anything powerful; if you have an idle integrated card that you aren't using on the Linux host because you only interact with it through a Web server or SSH (for instance, Proxmox), then pass that through. It's what I do on my home lab which runs a 9950X.

Before people raise pitchforks against Linux, this applies there, too, for the record: at work I have a Linux instance just to myself that by any other metric is ridiculously powerful: 64-core Epyc, 96 GB memory, but no iGPU, so remote desktop works very poorly.

To pass through a GPU - you'd need an extra GPU then..?

Also, the last time I checked, many GPUs explicitly detect + block this because they want you to pay for more expensive datacenter versions of the hardware.

Did something change?

I think you are confusing PCI passthrough with enterprise IOMMU GPU support that's nowhere to be found in consumer GPU:s.

when did you check last time? I've been using gpu passthrough for more than the last ten years with different gpus from amd to nvidia to onboard intels. last few years I went back to native windows, because some games refused to run in a vm.

A number of intel consumer CPUs support SR-IOV. The iGPU splits out to 7 "virtual functions" or pci devices to map to a VM. On latest Core Ultra's you need a 2x5 model.

- https://www.intel.com/content/www/us/en/support/articles/000... - https://github.com/strongtz/i915-sriov-dkms

My understanding is that it's trivially defeatable by configuring the hypervisor to hide itself (passthrough the real CPUID strings from the processor, ACPI/SMBIOS data/etc), and graphics drivers didn't really put any more effort into detecting beyond that. It's been years since I've been on this scene though, so my info may be out of date.

yes

or a GPU that supports virtualization

It's pretty good. They use XfreeRDP to remote into the container and display individual windows. This somehow performs a lot better than the GPU emulations of VirtualBox or VMware. I guess Microsoft put some effort into optimizing RDP for Terminal Server applications.

It's definitely the way to go. Been using this setup for years now. Windows rdp server almost never goes down. The occasional "please wait" error when starting a session can be fixed remotely by logging into a 2nd backup user account to unstuck the main account. Gives you windows on mac and linux, lets you choose whatever type hardware for your remote host. Connection outside LAN always wrapped in a tunnel or tailscale

You have minimal to zero leverage of the native Windows debugging, logging, or instrumentation. At best an opaque box with one knob and hopefully it doesn't fail or you will be roaming the countryside learning how to perform correlated packet captures at various levels of crappy obtuse networking. Could be useful for concealing non compliant vulnerable applications from pesky security vulnerability assessment teams. Combine that with the price is right and it is a solid 97% win exceeding performance metrics bonus pool refreshed.