Intel Releases New Versions Of KVMGT & XenGT

[phoronix]

Phoronix: Intel Releases New Versions Of KVMGT & XenGT

Intel has released their first new versions of XenGT and KVMGT for 2016 for GPU virtualization solutions for Xen and KVM, respectively...

http://www.phoronix.com/scan.php?page=news_item&px=Intel-2016-XenGT-KVMGT

[dh04000]

Wait, a gpu can be virtualized and two systems can have their screens displayed on one gpu? If I understand this correctly, that is pretty cool. I thought Xen/KVM systems HAD to have TWO gpu by default.

[bridgman]

Yep, this is pretty recent, as is the Virgil3D work that airlied has been doing.

https://virgil3d.github.io/

[anto44]

With XenGT and KVMGT one GPU is shared between multiple virtual machines, it doesn't need two or more GPUs nor two or more monitor like normal GPU passthrough virtualization. You need a integrated Intel GPUs (Haswell or +), Virtual machine and host machine use Windows/linux native graphics driver, the GPU is shared over time intervals when one machine needs it... performance is 85-90% of native speed.

More in spanish:

[Ericg]

Wait, a gpu can be virtualized and two systems can have their screens displayed on one gpu? If I understand this correctly, that is pretty cool. I thought Xen/KVM systems HAD to have TWO gpu by default.

Using standard, normal, PCIE Passthrough, that is true. Because PCIE Passthrough is all or nothing. Apparently Intel has come up with an Intel-specific technology that lets it be shared. Which is very cool, but unfortunately not cross-vendor.

[kobblestown]

Using standard, normal, PCIE Passthrough, that is true. Because PCIE Passthrough is all or nothing. Apparently Intel has come up with an Intel-specific technology that lets it be shared. Which is very cool, but unfortunately not cross-vendor.

I think both AMD and nVidia have solutions for such scenarios but the virtualization support is built in the hardware and only for workstation class boards. I guess it is seen as several PCIe devices (or "functions" or whatever the jargon is; maybe similar to how high-end networking chips do it) that can be passed through independently. Intel's solution does not require special hardware support but only works on Haswell and later GPUs. The advantage is that (1) most processors have one built in and (2) doesn't require special drivers in the guest. As far as I know, Virgil 3D should be able to use any sufficiently capable GPU at the host but will require its own drivers in the guest.