[SOLVED] GPU confusion

[jjg60]

I have been reading about Promox over the holidays. For my home network, I think it will be great for my Docker and Kubernetes workloads, but I am confused about GPU.

If I do not PCI passthrough the GPU, do desktop operating systems with desktops like Ubuntu and windows have their output only available over the lan via browsers or will the GPU port give me a way to passthrough to the VM?

If I passthrough the GPUs to a VM, then will the GPUs be available for bootup post? I have not seen info about Designair TRX40 being able to post without GPU.

 

[Dunuin]

If I do not PCI passthrough the GPU, do desktop operating systems with desktops like Ubuntu and windows have their output only available over the lan via browsers or will the GPU port give me a way to passthrough to the VM?

If you don't passthrough the GPU no VM can use it. So non of the physical outputs is working (or it will only output the PVE console) and everything needs to be rendered by the CPU in software mode (so stuff like video playback might stutter because no hardware accelerated decoding can be used). Accessing your VMs will only work using remote access protocols like Spice, RDP, VNC and so on.

If I passthrough the GPUs to a VM, then will the GPUs be available for bootup post? I have not seen info about Designair TRX40 being able to post without GPU.

Many GPUs got problems if they were used by the host before. Then they can't be passedthrough later. So best is to have 2 GPUs. One small GPU (like iGPU or a GT710 where PCIe 1x versions exist) that you set in the BIOS to be your primary GPU and another one that you passthrough later. And ideally these GPUs are from a different vendors so you can blacklist the drivers so your PVE host won't try to initialize it at all.

 

[jjg60]

I think I got it. Correct me if I got this wrong.

A small cheap GPU from a different vendor for bios post and proxmox console. Proxmos console is used only during initialization of Proxmox then everything about proxmox is done over the lan. So this cheap GPU port is for configuration and recovery activity only, and to make bios post happy if the bios does not like not having a gpu.

The primary GPU or GPUs are blacklisted based on vendor, These are passthrough to the VM. Before the VM starts, these ports are dead. Once VM is operational, these GPU ports act just like if they were on a bare metal install of windows or ubuntu-desktop.

 

[Dunuin]

I think I got it. Correct me if I got this wrong.

A small cheap GPU from a different vendor for bios post and proxmox console. Proxmos console is used only during initialization of Proxmox then everything about proxmox is done over the lan. So this cheap GPU port is for configuration and recovery activity only, and to make bios post happy if the bios does not like not having a gpu.

Right. Its always useful to be able to access the console using keyboard and a monitor. You don't need it often but if you get network problems this is the only way to access your host to fix stuff.

The primary GPU or GPUs are blacklisted based on vendor, These are passthrough to the VM. Before the VM starts, these ports are dead. Once VM is operational, these GPU ports act just like if they were on a bare metal install of windows or ubuntu-desktop.

Jep. You can try it without a small GPU for boot/host but don't wonder if you need to get one if it doesn't work. Using a different vendor is the easiest thing, for example Nvidia for boot/host and ATI for passthrough. In that case you can just blacklist the ati drivers so PVE won't try to initialize them so its more likely that then can be used by a guest (ATI card for example often got a "reset bug" in firmware). But it is also possible to use all GPUs of the same vendor. But in that case you need to backlist individual devices which is a bit harder.

 

[jjg60]

Thank you. I will start experimenting as soon as my server arrives.