GPU Server/Virtualization

[kno3]

In 2 gamers 1 CPU, 2 (different) graphic cards were used to serve 2 VM's. Being the owner of a (single) Titan X, I would like to add a VM (in addition to my 'personal' OS) to share the GPU with friends/colleagues, to allow them to utilize the GPU for CUDA programming. Basically share the GPU between 2 'systems'. I'm getting the impression that Nvidia GRID can do this using Citrix XenDesktop or vmware (although not sure if Titan X would be supported). Are there any other alternatives/solutions to have access to a TitanX on both systems? Having SLI TitanX I believe would not solve the 'issue' to use a similar setup as linus and luke used in the video (i.e. unraid/KVM) since the hardware needs to be different; correct?

[Electronics Wizardy]

You can't split a gpu using passtrhough. The grid gpu's have multiple gpu's on each graphics card. You can have multiusers with hyper-v, but it lower performance. You can have 2 of the same gpu's, but it makes it harder as you don't know which one is which.

[spartaman64]

4 minutes ago, Electronics Wizardy said:

You can't split a gpu using passtrhough. The grid gpu's have multiple gpu's on each graphics card. You can have multiusers with hyper-v, but it lower performance. You can have 2 of the same gpu's, but it makes it harder as you don't know which one is which.

wouldnt the software assign a hardware label to each gpu

[Electronics Wizardy]

1 minute ago, spartaman64 said:

wouldnt the software assign a hardware label to each gpu

Yes, you can't physically tell which is which, which linus wanted to do in 2 gamers.

[spartaman64]

6 minutes ago, Electronics Wizardy said:

Yes, you can't physically tell which is which, which linus wanted to do in 2 gamers.

you can probably figure it out by the software telling you which PCI slot its connected to and you can write a label

[kno3]

Thanks for the information. If they are identical, why is it important to physically know which is which? Is intel's KVMGT only for CPU-integrated graphics (attached)? Linux KVMGT (intel)

[Electronics Wizardy]

2 hours ago, kno3 said:

Thanks for the information. If they are identical, why is it important to physically know which is which? Is intel's KVMGT only for CPU-integrated graphics (attached)? Linux KVMGT (intel)

Linus wanted that so he knows which monitor to plug into which gpu.

[Electronics Wizardy]

2 hours ago, spartaman64 said:

you can probably figure it out by the software telling you which PCI slot its connected to and you can write a label

It doesn't tell you the slot, only the id, which isn't the same.

[spartaman64]

4 minutes ago, Electronics Wizardy said:

It doesn't tell you the slot, only the id, which isn't the same.

do you need to know though as long as you can distinguish between them and can assign them to different vms

[Electronics Wizardy]

1 minute ago, spartaman64 said:

do you need to know though as long as you can distinguish between them and can assign them to different vms

No, unless you have a monitor connected like linus did in multi gamers.

[spartaman64]

4 minutes ago, Electronics Wizardy said:

No, unless you have a monitor connected like linus did in multi gamers.

i guess but theres always trial and error

[kno3]

The virtual machine (Ubuntu) will be used remotely (so locally will only be using one as a desktop). The former will act as a server, hosting apache database and so, while I can use the desktop at the same time. Does this make any sense? New to the technical terms  

[leadeater]

3 hours ago, Electronics Wizardy said:

You can't split a gpu using passtrhough. The grid gpu's have multiple gpu's on each graphics card. You can have multiusers with hyper-v, but it lower performance. You can have 2 of the same gpu's, but it makes it harder as you don't know which one is which.

Actually the GRID GPUs also work much like SR-IOV network cards and can spawn multiple virtual GPUs at the host driver level that can be used on VMs. AMD also has the same ability in their FirePro S series. Nvidia is not using SR-IOV however just easy to compare to it.

 

Quote

AMD’s Multiuser GPU uses the SR-IOV (Single Root I/O Virtualization) standard developed by PCI SIG, and the company claims it’s explicitly designed for both OpenCL and graphics performance, with OCL 2.0 support, OpenGL 4.4, and DirectX 12 all supported as well.

 

http://www.extremetech.com/extreme/213278-amds-new-multiuser-gpu-will-slug-it-out-with-nvidias-geforce-grid

http://www.tomsitpro.com/articles/amd-firepro-s1750-firepro-s7150x2-hardware-virtualization,1-3129.html