[vfio-users] Virtual GPU support in Kernel 4.10

Alex Williamson alex.williamson at redhat.com
Mon Feb 20 15:35:17 UTC 2017


On Mon, 20 Feb 2017 14:41:42 +0000
Ivan Volosyuk <ivan.volosyuk at gmail.com> wrote:

> I see lots of hype for the new feature of the Linux kernel. What does it
> mean for my gaming setup? Can i get better performance with the technology
> once it get supported by Nvidia?

For a desktop user, vGPU is essentially limited to Intel IGD.  The new
VFIO Mediated Device infrastructure allows vendor drivers to expose
virtual devices to userspace using the VFIO API.  It is completely at
the vendor driver's discretion whether to support these mediated
devices for a given hardware implementation.  In the case of NVIDIA,
the vendor driver is of course the proprietary nvidia.ko module.  If
you look at nvidia.com for vGPU information, the deployment guides
include only the product briefs for M-series Tesla cards, you won't
find mention of support for GeForce anywhere.

As for performance, vGPUs are essentially fragments of the physical
GPU, one to perhaps many mediated devices may be created, depending on
the vendor driver implementation.  Deciding which vGPU model to use
depends on the requirements of the individual use case, but in general,
a vGPU can never exceed the performance of the physical GPU.  In other
words, it's a fan-out solution, like virtual SR-IOV, to provide more
virtual devices from a single physical device.  It's not an acceleration
technique with capabilities beyond direct assignment.  Thanks,

Alex




More information about the vfio-users mailing list