Return to Level1Techs.com

How could we entice AMD to enable SR-IOV on future "gaming" GPUs?


#1

I don’t know about you but I would love to be able to run a Windows gaming/editing VM on a Linux host with a single GPU.

I know that the audience interested in this is small. Maybe a little bit larger on this forum but still there are few of us that are willing/ have the resources or the need for this technology.

In my mind, this extends far beyond my original question onto Linux vs WIndows awareness but I am not going to into more detail about that in this thread.

What could we do off the bat that could persuade AMD to have a look at this? I am specifically mentioning AMD simply because they seem to be closer to the end user. I am not an AMD fanboy, I just like companies that respect my needs. Nvidia actively tries to stop you from running your non-Quadro GPU in a VM, they like propriety software and hardware and in the end would never listen to a small crowd like us (If they did though, I would be interested). AMD on the other hand is. As an example, Lisa Su replied to Geoff (https://github.com/gnif) on the matter of (re?)initialization of Vega GPUs.

What I am really asking for is opinions on what we could do to accomplish this. Petitions? E-mails? Preorders/Mass purchase orders? What do you think?


#2

Probably nothing and wait till the very very small group of people that does VMs grows to an acceptable size to add sr-iov to consumer GPU’s. I wonder if the enterprise going to take consumer GPU’s now instead of the expensive ones.


#3

PCI-E 4.0 and peer-to-peer communications would mean the SR-IOV GPU doesn’t need display outs, but rather it uses a DirectDMA or similar tech to stream the framebuffer to a “head” card with no GPU, but rather an FPGA that can output the framebuffer data from the single GPU.

As of current, SR-IOV cards with no display output is more for VNC and remote thin clients. SR-IOV needs peer-to-peer PCI-E and Display output only cards that use the PCi-E Peer-to-peer to assign specific outputs for each VM with their own PCI-E resources.