Intel Flex GPU with SR-IOV for GPU accelarated VDIs
-
We have 0 flex GPUs so we cannot even test this. Also, there's no demand at the moment for this, explaining we have other priorities. However, our next platform (XCP-ng 9) will have a more recent dom0 kernel, that might help if it doesn't work today.
-
@olivierlambert I do not mind purchasing the GPU, installing on a test server (R760 2x Intel cpu) and providing ssh or a vpn to this server if that would help?
-
I have 0 idea on what's needed to make it work (even if it's possible) so first it would be just testing it (if that works out of the box) and if not, trying to understand why it doesn't, and then estimate the amount of work needed.
-
I can't believe that there's not much interest.
Anyway...As far as I know these are pretty well supported in newer kernels. I think you need to be fairly deep in kernel 6. Given that XCPng/XenServer is currently running on kernel 4 with a bunch of backports, this might be a little problematic.
Last I knew, Intel is not charging licensing fees for using vGPUs like NVidia does.
With a working driver in XCP-ng and no licensing fees, this is could be a real cost-effective VDI platform.
-
If it really requires a 6.x kernel, this will at least wait for XCP-ng 9
And regarding the demand, you are officially the first to ask for it 
edit: the second one after OP
-
@JamesG There probably is interest from homelab people, but for use in production setups, I don't really see a lot of businesses needing it.
VDI isn't really used that much by businesses now (at least the ones that do use it are slowly moving off it) and most that do use it don't need GPU acceleration. Often times it's for pretty basic applications, medical record systems, etc...
So I think that is where the lack of demand is coming from.
-
@planedrop we are far from home lab and exploring alternatives to VMware because of a 7 digit yearly subscription fee
. We use gpu’s in around 75 host to reduce cpu load and free up cpu resources. We have a lot of single thread applications requiring higher speed and fewer cores so gpu’s free up the resources for that software -
@jebrown Yeah I just meant in general, not specific to you. I think most businesses aren't looking for this kind of specific workload so the demand isn't very high for it.
Are you currently using Intel GPUs with Flex for this use case? Or NVidia right now and just looking to change? It might be worth considering leaving VDI, not sure if you're in the position to be able to argue that, but there are often better solutions now.
Either way, I would personally also love to see this, I just think there are other things that more companies are asking for from XCP-ng right now.
-
I have not tried the flex (yet, but I can install and test it) do you have an alternative to recommend? I’m open to suggestions on that as well. I just seen the open source drivers and no additional licensing. Majority of the current set up requiring vgpu is done with nivida and VMware…… not everything we have needs gpu but the systems Using or needing the extra power from graphics are in windows and being used 24/7 and just accessed on site by thin client’s. we moved away from any type of cloud provided daas to have better reliability on critical systems and prevent down time.
-
I asked about this in the Discord server a little while ago. To be honest, one of my employers would love Flex vGPU support. We've been needing to update both our general compute and engineering VDIs so we've been looking at some solutions.
While NVIDIA vGPU pricing isn't horrible, it requires us to have VMware infrastructure which is significantly more expensive (obviously). So we'd really like NVIDIA and/or Intel vGPU support, but understand that it isn't just a few lines of code that needs written and is a very small subset of XCP-ng users. Especially with the alleged uncertainty of Intel right now, it may not even be worth developing anything for their cards, but that's personal conjecture.
I've been building my own VDI solution for XO(A), but that's been stalled for a little bit due to burnout, but I plan on finishing a prototype solution and writing a setup guide in a few weeks. So if you want something for just general compute, keep an eye out!
-
Generally speaking...Just because you haven't had direct requests doesn't mean the feature isn't desired.
It's easy enough to look over the current features and because it's not listed, you assume it's not there and move on to find the next viable solution.
Proxmox and OpenShift seem to be killing it in this space with Intel Flex GPU's.
-
Sure, but if it's 6 months of work with 3 engineers, this means it's not "free" and you need enough resources to do it. So if people want something like this, we need to get at least some traction and a way to know how much work will be involved
edit: obviously, if it's plain SRIOV without any special work, it's another story.
-
@olivierlambert From Intel's page:
"With up to 62 virtual functions based on hardware-enabled single-root input/output virtualization (SR-IOV) and no licensing fees, the Intel
Data Center GPU Flex 140 delivers impeccable quality, flexibility, and productivity at scale." -
If you are that much confident, then feel free to purchase one and use it: XCP-ng supports SRIOV but for NICs, I don't know how different it could be for a GPU, but without any hardware to test it's impossible to answer. How difficult it was to work on a more simpler hypervisor (KVM) doesn't send encouraging messages though.
-
N nick.lloyd referenced this topic on
-
I know this is 3 months old, but it is still relevant to my organization, and I'd have to believe others (we can't be the only one). We've been a XenServer/Citrix Hypervisor user for years. We're a Citrix/MS shop so it made sense to utilize their hypervisor. We were going to go with VMWare prior to Broadcom this year (or next), but with VMWare's licensing model we really couldn't afford their top-of-the-line edition, per core (at least there was a chance per socket) that is needed to facilitate SR-IOV for GPUs. On top of that NVidia requires licensing as well for the use of their GPUs in this way.
The Intel Flex is a great solution because there is no licensing and they allow 60+ SR-IOV sessions for their not as potent 75-watt Flex 140 card which is perfect for general applications and anything web based. Not to mention they are less expensive overall than their NVidia counterparts. Our VDI environment doesn't run anything real graphics intensive like CAD this is strictly to render web-gl based applications in a more efficient way like our own internal web map that wasn't an issue until we upgraded our ESRI version. I'd very much appreciate this feature being incorporated in if possible mainstream. In reading above I'm strictly talking about SR-IOV and not some other method of GPU sharing if that does indeed make it easier to facilitate. It would certainly improve our user experience and environment efficiencies. So a +1 from me to add the feature in one of your future releases if you are considering it.
-
@jrouerdc Hi!
A good way to help would be to have access to some hardware and to understand more what's needed at Xen/toolstack level
I will ask around but if there's a chance we can get one card, it might be really helpful. -
Yes I believe that is a possibility. I make an assumption when you have completed your work the card would be returned to me? They are less expensive than others of this kind, but still have a higher price tag than a typical gaming card to the tune of 4 figures.
-
Is there a shop in Europe where it could be bought?
-
Mine came from Cisco as part of the new UCS Servers I purchased the C240 M7's. I know Dell sells them. I'm from the states so I'm not 100% sure they do so in Europe, meaning Dell or if they have a EU division. Lenovo has it available as well based off my reading. Since you're doing it in the name of bettering the industry you'd think that Intel or one of the hardware juggernauts would want to get one in your hands. Dare I say maybe eBay?
-
I will check my contacts to see if anyone can have that card to work on it.
