More of an enterprise topic, but seems like the A-100 is SR-IOV compliant. Would that mean it would be compatible with Looking Glass?
Most Nvidia GPUs are “SR-IOV compliant” in that they contain the functionality to support an SR-IOV driver implementation in theory. In reality the functionality has been locked out behind Nvidia GRID. Even the consumer cards support SR-IOV in the silicon, they just don’t expose the GRID APIs. If you don’t mind paying for VMWare/Critrix with GRID then perhaps it may support Looking Glass, although I’m not entirely sure. You may be better off in a consumer device getting two graphics cards (one for your host OS and one for your guest OS) then using that to do Looking Glass. Alternatively you could play wth a device with more standard APIs for this kind of use case - for example the FirePro W7100.
From what I understand, SR-IOV != Grid
Grid is it’s own thing, and it is not SR-IOV.
It is probably more like all chips have at least some rudiments of SR-IOV, but it is unfinished or locked out so they can sell you GRID instead. And GRID is locked out in the consumer chips so they can sell you a quadro instead.
I think it could work in theory.
The first hurtle is getting software support on your hypervisor. I would hope that since SR-IOV is a standard, you could plug it into a Linux KVM/QEMU system, and everything would just show up.
The second hurtle is getting guest drivers that support getting a normal desktop running. I don’t know if the correct drivers are available directly from nvidia, or if another driver could be persuaded to work with a bit of tweaking, or if guest driver support basically just does not exist.
1 Like
From what I understand, SR-IOV != Grid
Ya GRID isn’t technically SR-IOV, but if you read the GRID documentation and cntrl-F “SR-IOV” they do tell users to enable it under certain motherboards.
Also Nvidia engineers have supposedly told Anandtech that SR-IOV is in the silicon, just in a disabled state on consumer cards and possibly implemented in a non-standard way on “GRID supported” cards.
I think it might be possible with VMWare (maybe other hypervisors too, I just dont see any docs at the moment) - would share the link but doesnt seem i have access to. One of the interesting things about the A-100 is it supports this new feature called Multi-Instance GPU (MIG), which i think is separate from GRID
Seems like with MIG (which seems to be based on SR-IOV) it is in theory possible to physically partition the Nvidia GPU as opposed to GRID which does the temporal slicing.
Now im not sure what guest drivers nvidia provides for MIG, but assuming there is a driver just for CUDA, would looking glass be able to use the GPU for rendering without the monitor?
This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.