My main desktop runs Linux, I do have a Windows VM (on qemu/libvirt) that I use primarily for Fusion360 and some gaming.

I had to setup GPU passthrough with a second GPU for this.

It is not that hard, you will however need to make sure your motherboard supports proper IOMMU grouping. This however, is the tricky part since most motherboard manufacturers do not really provide this information.

[1] More info on ArchWiki: https://wiki.archlinux.org/index.php/PCI_passthrough_via_OVM...

So IIUC your approach, you're exposing the actual GPU to the guest OS via PCI-passthrough. And in practice that means you'll want two GPUs: one for the host and one for the guest.

But looking at the online docs for VMWare Workstation, it sounds like they take a different approach: the guest VM has a virtual device driver that lets both host and guest OSs use the same GPU at the same time.

I would think the VMWare Workstation-like approach is preferred, because it avoids the hassle of using two graphics cards. I know that Workstation is expensive ($250) and not OSS. Besides those two reasons, are there still reasons to prefer the PCI-passthrough approach?

> Besides those two reasons, are there still reasons to prefer the PCI-passthrough approach?

The software approach generally has lousy performance due to the extra overhead of forwarding rendering commands between the guest and host. In addition, since this approach requires writing a completely new graphics driver for the guest, there tend to be limitations with rendering API compatibility, as well as bugs. With the hardware approach, the guest can use the GPU vendor’s normal driver.

Just yesterday I finally got to running actual tests of graphical performance between my host (Windows 7 Ultimate, NVIDIA GTX 950 - not top of the range :)) and guest (Windows 10 Pro running inside VMWare Workstation 15). Used "PerformanceTest 10" from Passmark. Host scored 6015, ran all benchmarks nicely. Guest scored 1645, only ran some benchmarks (due to lack of Dx12 and compute capabilities), and even those were penalized.

Virtualbox guest didn't participate in the contest, as I find graphical performance of VBox guests noticeably worse.

I must add though, that for desktop tasks VMWare Workstation performance is absolutely adequate, able to run double 2K monitors wonderfully. For gaming setup I'm personally planning to play with PCI-passthrough. The only problem is not to get lost in the tinkering, before playing any actual games :)

I did not try VMWare Workstation, but I did try VMWare Player and VirtualBox to see if I can get away with running Fusion 360 on a plain VM, and both had extremely horrible performance.

I remember even SketchUp wouldn't start in VirtualBox. I was able to get SketchUp running in VMWare Player, but the performance was very bad.

edit: this was also about 1.5 years ago, not sure how much improvements VMWare and VirtualBox have made in that time in terms of virtualizing GPU calls.

That is why I decided to end up setting up a VM with GPU passthru. It works extremely well for my usecase. I was able to score a cheap GPU for my Linux host (since I do not do anything graphically intensive on Linux, just need to drive 3 monitors), and was able to passthru my more expensive GPU over to the Windows VM. I would assume if you have integrated graphics on your CPU, you could use that for your host OS and get away with having only a single GPU.

With your setup, is there a convenient way to flip back and forth between which GPU is used by the Linux host vs. which GPU is used by the Win10 guest?

I'd still like to use Windows 10 only for those games that I can't make work on Linux.

I use a separate GPU for my linux host and the windows VM.

I am 99.9% sure you can not passthru a GPU to a VM that is being used by the host.

edit: I think I misunderstood your question. The GPU that gets passed thru to the VM, never gets initialized by the host system during boot. So, if you want to change which GPU you want to passthru to the VM, you will have to reboot to allow your host to use the other GPU (and not init the on you want passed thru to the VM).

I would suggest you look at a youtube video where they run thru setting this up, so you get a better idea of what this entails!

Thanks! I'm still deciding if it's worth the effort just to play certain games on Windows 10. The ease with which I can change the OS-GPU association may be the deciding factor.

I haven't tried it myself, but there's a bit of info on the /g/ wiki with a hardware database linked at the bottom.

https://wiki.installgentoo.com/index.php/PCI_passthrough

I've used PCI passthrough with VMWare ESXi. It's good when it works, but when it doesn't, you're out of luck. I doubt GP will have much luck with VMs once it gets to 3d acceleration.

GPU passthrough with qemu/kvm works spectacularly to 3D applications, I've been using it for years to play Windows only games from my linux desktop.

I run Proxmox (Debian) with an AMD GPU as well as Nvidia. I can pass both to Windows and they work fine, the Nvidia one is much more of a hassle though, you can run into Code 43 issues, meaning that Nvidia detects you're in a VM and refuses to start. On AMD this is much less of a problem.