Intel Flex GPU with SR-IOV for GPU accelarated VDIs
-
Yes that is correct. It's their low power 75 watt solution that allows (if I'm not mistaken) 63 SR-IOV instances per card. I acquired them as part of an end of year promotion 2 per server so I could eventually run up to 120 VDI instances per server with shared GPU. Again, we don't game, but many of our apps are internet based and especially those like maps (just viewing) and some others that have forms benefit from the GPU acceleration. Even a 5+ year old Celeron with internal graphics beats our VMs in that instance. This would level the playing field and make the user experience the same as a full desktop in all things. Thank you for considering this and assisting. I know Proxmox has it working well SR-IOV from what I've read. Hyper-V you have to fiddle with power management but it also works SR-IOV and VMWare was probably first to adopt it. The white paper/partnership between Citrix, Cisco, and Intel actually goes through how to set it up on VMware to work with VDA 2311 and higher. That may not mean much to the people reading this but I find it a little ironic that their setup/partnership with Cisco/Intel would lead to them utilizing VMWare instead of XenServer.
-
If it's really only SR IOV, it might be doable without too much work (ie "only" need toolstack wiring).
But it's hard to do anything without the hardware in question. I'm taking a look on purchasing one, I found some shops in EU selling it.
-
Absolutely, I completely understand what you are getting at/saying. I was just adding a little more info figuring it couldn't hurt. I appreciate you taking the time to even address and go the lengths of getting a card to test with. To be addressed by the CEO/co-founder is pretty amazing and I'm honored to even get traction. Thank you again!
-
Is there any more recent card and do you know if Intel will provide such GPU in the future? I don't want to invest in another tech Intel will kill soon as they did with Optane and such
-
These are them currently. There's a more powerful version the Flex 170 that's the same thing but for more graphics intensive workloads. It has less SR-IOV channels, faster GPUs, requires more power and more video RAM but its purpose, although VDI, is for your CAD type users. Same line though you can still buy them now and they are Intel's all in for VDI GPUs currently to cut away at NVidia's market. They don't offer licensing to utilize the shared GPU. It's why I made the investment. I literally got mine in our new Cisco UCS C240 M7 servers in January so they're definitely still selling/including them in server hardware. They also offer a MAX 1500 series but that's to utilze the GPU computations like crypto-mining or programs that run utilizing GPU for faster computations. They weren't designed with VDI in mind.
-
Not sure if I could post a link, but here is the info on their Flex line of datacenter GPUs.
I can't comment on longevity but at least it's still their newest line in that market. I'd also add (opinion) that as budgets get tighter and they get more support from hypervisor/vendors it can only strengthen their position. I apologize for continuing to name drop, but VMWare and NVidia's licensing costs put us in a bad way. This was/is our one shot to be able to offer the same and possibly better experience at a price point we could handle. If I can't get a solution to utilize them properly I've not burned much with the incentives I received at the end of last year for our server purchase, but if I can it checks all the boxes. I would bet it may sway others to step away from the more expensive offerings as well. Looking at stock prices for VMWare and then offerings like Nutanix it would appear that it already has.
-
If you can provide some level of commitment moving on our platform I can provide a similar level of commitment to work on getting those cards working out of the box.
-
Not opposed to this, I'm not a private company though, we're (my organization) government so it doesn't move as quickly with approvals. I know that seems like a deflection, more transparency. We currently have a commitment with Citrix/XenServer as part of their new licensing model to run all my other Citrix technology and I'm investigating alternatives to allow for the cards to operate effectively. Our new servers will hold over (just because they are faster) our end users while we evaluate. It's not like they can't do their work, just some things take a bit to render on their VDI than those running physical machines. We've even contemplated, like those that posted earlier, going away from VDI, but it still seems to be the best solution for what we do and the limited technicians we have on staff. We've been all in with Citrix for years.
-
@olivierlambert Another option is possibly applying for Intel Developer Zone Premier. At this level they provide hardware testing support, I'm not sure if this means they can ship their hardware to companies to test against.
But if that's the case then you could then get the Intel Flex GPU and their other products ahead of time. As well as a closer connection to Intel for any issues, which come up involving XCP-ng and Xen Orchestra.
https://www.intel.com/content/www/us/en/developer/programs/overview.html#gs.ko8q8e
-
Entering a big vendor program is probably 100 or 1000 times harder to grabbing a GPU ourselves. Now, the challenge is finding one in stock in EU.
-
@olivierlambert said in Intel Flex GPU with SR-IOV for GPU accelarated VDIs:
Entering a big vendor program is probably 100 or 1000 times harder to grabbing a GPU ourselves. Now, the challenge is finding one in stock in EU.
Maybe but the value out of it being able to access information, as well as easy access to people at Intel would pay for itself. As well as access to up coming products, to use for testing. Also would make accessing the card easier as, currently you need to buy a solution from a builder that uses the card.
@jrouerdc Have you read the https://xcp-ng.org/blog/2025/03/14/the-future-of-xcp-ng-lts/ blog post yet? Evaluating, planning and potentially using XCP-ng version 8.3 would be a very good idea. This because the LTS release of 8.3 will have the 2TB VDI disk limit removed, also an extensive support lifecycle compared to 8.2.
-
-
I received one card, I quickly tested it, you can actually see 2x GPU (as expected), PCI passthrough each to a different VM. So at least, basic stuff works.
A next step (when we can staff someone) is to check how to manually enable SRIOV on the device and pass it to the VMs, and in the end, get that done in XAPI to make it integrated with XO.
-
@olivierlambert Great to hear...
Looking for forward to have Arc GPU support. If possible can you please be specific on the GPU hardware you have and which is being tested. As you know, someone might have the one you have or a similar GPU that they can use. From what you posted, I assume that you might have the ARC Pro B60 Dual 48G by Maxsun - correct me if I am wrong.
Lately, I am trying to source a Arc Pro B60. I do not have any timeline when I would be able to obtain one.
At least for the time being and considering the alleged MSRP for battlematrix - Arc Pro cards, Intel seems to be the best value.
-
No, I have an Intel Flex 170
-
@olivierlambert If I remember right, you should be able to see 62 VF's on that card. There might be a tool needed to define how many VF's are present like on a NIC.
-
Yes but in order to see the VF I need a working driver in the Dom0 first
Probably a lot easier with a more recent kernel in XCP-ng 9.0 -
@olivierlambert Ideally you need to be somewhere into Kernel 6. 6.12 is sticking out in my head, but I'm not positive when support got fully integrated.
-
Upstream inclusion is ultra recent, but you can always add drivers yourself (still requiring a relatively recent kernel IIRC). Anyway, we have the hardware and it will be in the roadmap at least to explore
-
@olivierlambert While VDI is maybe not as vital as it once was...I'm experimenting with multimedia work in XCP-ng. Having a VM with GPU off-loading of CODEC encoding would be nice. It's a pretty big CPU hit to make that go.
