[vfio-users] New kernel feature: how close is this to what we are doing here?

[Alex Williamson]

On Fri, Oct 23, 2015 at 9:27 AM, ALG Bass <olorin12 at gmail.com> wrote:

> http://phoronix.com/scan.php?page=news_item&px=VirtIO-VirGL-3D-In-Mesa
>
> Saw this story on Phoronix. How close is this to what we are doing here
> with vga passthrough?
>

Virtio GPU is to direct assigned GPU as virtio net is to direct assigned
NIC.  Yes, it will add 3D support to the guest with offload to hardware as
well as multiplexing a single GPU to multiple VMs.  It's going to be a
really nice feature.  Is it the holy grail?  Only to the same extent that
virtio net is to networking.  It's going to require paravirtualized drivers
in the guest, those will come quickly for Linux, likely less so for
Windows.  Gaming compatibility with those drivers (DX10/11/??) is currently
unknown I expect.  There's also a certain amount of overhead, commands to
the GPU use a common intermediate format, which like virtio-net limit the
peak performance.  I'm sure there are going to be games for which this is
more than sufficient and for VDI solutions, it's hopefully exactly what
they need.  However, for peak performance, just like for peak networking,
direct assignment is going to be the benchmark.

There's also work underway that allow more direct access to the GPU, and
native guest drivers, while still allowing it to be split between VMs.
This is what both NVIDIA vGPU and Intel GT (KVM-GT) technology is meant to
do.  These will hopefully both offer very near the performance of direct
assignment and the native drivers in the guest, while offering the option
to share the device among VMs.

In the next few years, we could see some really interesting VM graphics
capabilities.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/vfio-users/attachments/20151023/f3b0414f/attachment.htm>