3 ms·

> I wonder if we will ever see some kind of GPU sharing for virtualisation Hasn't this been solved for some time now? Or am I misunderstanding what you want? I

> I wonder if we will ever see some kind of GPU sharing for virtualisation Hasn't this been solved for some time now? Or am I misunderstanding what you want? I remember seeing a few articles about using this for gaming on Windows VMs on Linux hosts. I haven't jumped in though so maybe it's not what I think it is. https://wiki.gentoo.org/wiki/GPU_passthrough_with_libvirt_qemu_kvm https://wiki.gentoo.org/wiki/GPU_passthrough_with_libvirt_qe... I think it's also definitely in use for compute stuff on VMs in professional environments. https://docs.microsoft.com/en-us/windows-server/virtualization/hyper-v/deploy/deploy-graphics-devices-using-remotefx-vgpu https://docs.microsoft.com/en-us/windows-server/virtualizati... Edit: Ah, maybe sharing a GPU between a host and VMs isn't possible at the moment. Only pass-through and maybe sharing between the VMs?

apatheticonion 5y ago
I could be wrong but I think you can only pass through an entire GPU as long as it's unused on the host. Sharing a used GPU, in the same way we can share a CPU with no performance loss via Intel Hyper v, is not possible.
- RamRodification 5y ago
  Yeah, just edited my post with that point. I think you're right.
- funcDropShadow 5y ago
  Sharing a CPU with no performance loss isn't possible either. The amount of CPU loss varies dramatically by the exact virtualisation stack used and by the application running on top of it, but it is very clearly observable.
  - apatheticonion 5y ago
    Are you sure? I am not sure if Geekbench is a valid benchmark but I have an AMD 5900x. Running it in Windows (host) I have a lower score than I get in VMWare running Ubuntu (guest).
- joveian 5y ago
  Intel GPUs can be shared, it is called GVT-g (expensive AMD and nVidia have similar abilities). It isn't trivial to set up at this point but it sounds like it works.
NavinF 5y ago
Nope, Nvidia will never let you pass PCIe virtual functions to VMs. They'd rather charge 7x the price for data center GPUs with that capability.
- Fnoord 5y ago
  Again: don't buy Nvidia. This is one of the many reasons why you should consider avoiding this brand (for GPUs). I even gave them the benefit of the doubt with their Nvidia Shield (bought two Pro 2020's) only to have them shove me up ads through my throat via a software update.
  - my123 5y ago
    That's not NV related sadly, but applies for all Android TV devices. It's beyond their control... And for PCIe VFs for virtualized GPUs, AMD does _not_ provide it either outside of their datacenter line. Note that if you just passthrough the whole GPU, that works out of the box on NVIDIA cards. (btw, for some NV cards, there is: https://github.com/DualCoder/vgpu_unlock https://github.com/DualCoder/vgpu_unlock to crack that protection)
  - NavinF 5y ago
    Again: You don't have a choice. Which AMD GPUs support VFs? Which ones support ROCm? The older AMD consumer GPUs did, but the newer ones don't. They prefer to charge extra for the data center GPUs just like Nvidia.
- mycall 5y ago
  I have GPU-P working with Hyper-V on a NVidia 3070. It is strange though, as the guest OS sees the 3070 and some software works well (e.g. FurMark) but you can't install the NVidia Game Ready installer (doesn't see a GPU). Otherwise, you are correct.

apatheticonion 5y ago

I could be wrong but I think you can only pass through an entire GPU as long as it's unused on the host. Sharing a used GPU, in the same way we can share a CPU with no performance loss via Intel Hyper v, is not possible.

RamRodification 5y ago
Yeah, just edited my post with that point. I think you're right.
funcDropShadow 5y ago
Sharing a CPU with no performance loss isn't possible either. The amount of CPU loss varies dramatically by the exact virtualisation stack used and by the application running on top of it, but it is very clearly observable.
- apatheticonion 5y ago
  Are you sure? I am not sure if Geekbench is a valid benchmark but I have an AMD 5900x. Running it in Windows (host) I have a lower score than I get in VMWare running Ubuntu (guest).
joveian 5y ago
Intel GPUs can be shared, it is called GVT-g (expensive AMD and nVidia have similar abilities). It isn't trivial to set up at this point but it sounds like it works.

NavinF 5y ago

Nope, Nvidia will never let you pass PCIe virtual functions to VMs. They'd rather charge 7x the price for data center GPUs with that capability.

Fnoord 5y ago
Again: don't buy Nvidia. This is one of the many reasons why you should consider avoiding this brand (for GPUs). I even gave them the benefit of the doubt with their Nvidia Shield (bought two Pro 2020's) only to have them shove me up ads through my throat via a software update.
- my123 5y ago
  That's not NV related sadly, but applies for all Android TV devices. It's beyond their control... And for PCIe VFs for virtualized GPUs, AMD does _not_ provide it either outside of their datacenter line. Note that if you just passthrough the whole GPU, that works out of the box on NVIDIA cards. (btw, for some NV cards, there is: https://github.com/DualCoder/vgpu_unlock https://github.com/DualCoder/vgpu_unlock to crack that protection)
- NavinF 5y ago
  Again: You don't have a choice. Which AMD GPUs support VFs? Which ones support ROCm? The older AMD consumer GPUs did, but the newer ones don't. They prefer to charge extra for the data center GPUs just like Nvidia.
mycall 5y ago
I have GPU-P working with Hyper-V on a NVidia 3070. It is strange though, as the guest OS sees the 3070 and some software works well (e.g. FurMark) but you can't install the NVidia Game Ready installer (doesn't see a GPU). Otherwise, you are correct.