[vfio-users] Tesla K40 passthrough

Alex Williamson alex.williamson at redhat.com
Thu Jan 7 17:20:35 UTC 2016


On Thu, 2016-01-07 at 14:28 +0800, Eddie Yen wrote:
> Hello, guys.
> 
> Few days ago I got Tesla K40 and try to passthrough it into VM.
> 
> Using Windows 7 as guest, and the virtual VGA is QXL.
> Every installation works fine, I don't have to add kvm=off and x-vga=on.
> Tested on CUDA and OpenCL, both works.
> 
> But now is a problem, how can I make 3D acceleration works on K40?
> I knew that K40 didn't have any VGA output, so I got slow motion when playing video and rotate 3D models in some 3D program like Autodesk.
> 
> My guess is, K40's 3D acceleration only works at 3D rendering, it didn't work in viewing 3D models.
> Does anyone have a idea?
> 
> Also I found that virtio-gpu is available since QEMU 2.5+. But not work on Windows guest yet.
> Is that can make a simple graphic accleration on Windows?

How are you connecting to the VM, just the QXL display?  Rendering into
the emulated graphics framebuffer is possible, but you lose most of the
performance.  I've never used a Tesla personally, but with GRID, which
is also a headless GPU, you'll generally want to disable the emulated
graphics (happens automatically with win7) and use guest-based remote
desktop solutions to access the accelerated desktop.  Minimally a vnc
server running inside the guest, though that burns a lot of CPU.  There
are more performant commercial solutions.

virtio-gpu is a different beast.  Once we have windows drivers there
you'll be able to get GPU offload from any host GPU supporting gallium,
but it's not clear when that driver will come or how much that will
enable games.  You certainly wouldn't have access to CUDA or PhysX or
GPU specific SDKs in the guest using virtio-gpu.  Thanks,

Alex




More information about the vfio-users mailing list