Is there any package needed to be loaded on proxmox. Running verison 8.1.4 with intel A770 8GB. Fails pretty quickly running make
Have to edit the file in the how-to
1 Like
You may also need install the headers and a bunch of other packages:
apt update && apt install pve-headers-$(uname -r)
apt install -y git sysfsutils dkms build-* unzip
apt install -y devscripts dh-dkms
At least these are needed in my Proxmox installation.
It would be great if it could eventually work for A770 even with a fewer virtual function allowance. Indeed Intel is missing an opportunity here.
The Datacenter Flex series launch was actually as absymal as the consumer Arc. I recall enquring my local distributor to get a quote and they could not even confirm the pricing details and delivery schedule… Hope they would have things smoothened out when Battlemage launches.
Hello there,
I’ve done the dkms/sr-iov mod a few months back for my i5 12500, and got the 7 iGPU functions running as expected.
Previously i had a nVidia GPU pass-through to a Windows 10 VM, that worked great with the help of Sunshine/Moonlight combo i could even game on that VM… But i got the 12500 and decided to ditch the nVidia GPU and just use the iGPU for HW acceleration (i dont really have time to game anymore and the dGPU consumes around 10W at idle for no reason).
The windows 10 VM works kinda okish using the iGPU VF using RDP client instead of the previous setup (Sunshine/Moonlight).
The real issues i have are related to using the same setup for a Ubuntu 22.04 VM, again using xRDP as the protocol. Somehow the overall performance is lacking i even compiled to get h264 compression working but still the overall performance is a bit lacking, at least not what i was expecting.
Any ideea what i am doing wrong with my setup? Any ideea how i could make Sunshine work with this setup (as i understand it the problem is that there is not display connected so the sunshine does not have anything to capture) maybe a virtual display or something!?
Thanks.
1 Like
Wow, this sounds cool, I would really like to know what you managed. Would it be possible for you to share?
soooo… where can we even buy this Flex 170? I can’t find it anywhere?
1 Like
They pop up on ebay from time to time for a little over 1000USD, MSRP is about twice that but it is very difficult to find a vendor that will sell to non-businesses at the moment. Wiredzone is selling them for 2100USD.
Supermicro had them in a link from intel but they seem to have disappeared. Ill email them sbout getting them on the web store. Thsy were also in specific configs on dell.com a few weeks ago when i was working on the video. Ping your supermicro rep?
Have a cluster of epyc’s with both Flex 140 and Nvidia L4 on the way. VMware shop currently but looking to make the move to proxmox (I think we have evaluated nearly every option). Hopefully we can get more VFs working.
Anyone have a good user/session interface? That seems to be the hardest part of VDI, present the available resources and then actually give the user a (linux) desktop (preferably with multimonitor and USB passthrough support).
Is Spice (protocol) dead/dying? Seems like the right protocol to base a proxmox VDI solution on but maybe something else? Any thoughts on KasmVNC? Something else?
Remote desktop, rdp, but you need to apply a patch to get hardware h264. This provides a great linux experience. Im told there are commercial products for proxmox to manage users and sessions.
One shop I saw simply provisioned a vm from a master, scripted changjng the mac and running sysprep, and the hostname was what the employee would always rdp to. Dhcp and mac made sure that this part never changed, and they were working on power shell scripts to manage the instance via azure ad. It sounds clunky but it actually saved a lot on licensing since, essentially, each vm was treated like a byod from the azure side yet could be managed via powershell and, iirc, ninja rmm.
This is probs ok for up to 100ish people but id want something more coherent like horizon otherwise
2 Likes
I’m heavily considering exploring this as a solution for doing ephemeral build machines for an Unreal Engine title I work on. I’ve done something similar in public clouds with expensive GPU instances, so I’m interested in seeing if I can get reasonable build times on iron I can buy for less. I’m less interested in bifurcating the GPU as much as sharing the GPU with the host, in the hopes I can take advantage of some Raptor Lake NUCs with Iris I already have before I propose buying a separate build system.
I’m currently in the process of labbing this out on a Proxmox host with an i5-1340P, which certainly is not the strongest chip, but it’s what I have with compatible graphics currently on hand.
I’m very new to the world of SR-IOV and GPU accelerated VMs though, so I do have loads of questions. I imagine since I’m running Proxmox and not Horizon, I’m not going to get a lot of flexibility on how I can partition the GPU, but what I don’t know is whether that’s dynamic at VM instantiation, or if that’s determined by how many VFS’s I configure at host boot.
@wendell if you’ve still got this iron on hand, I’d be interested to try a UE5 sample project build to see what kind of build times you get. LMK if you’re up to giving it a shot, and I can provide more instructions.
3 Likes
Lets try ifmt, send me info. If ylu need just one syatem a normalish gpu passthrough probably works
2 Likes
Awesome, thank you so much Wendell! I do imagine normal GPU passthrough would work, but I’m hoping to have a world where I can run 2 builds side by side on a system of a similar config to the Supermicro unit you have, which is why I’m so interested in this tech for this solution.
I’m labbing this out first so I make sure my instructions are accurate. To run the best candidate project though, there’s a number of accounts one needs (one for Epic, one for Houdini). If you’re not up for registering accounts to lab this out for me, I figure the best option at this point is to, if you’re okay with it, give me access to a box to run it myself, or alternatively, I can validate this config and export a Proxmox VM backup for you. It’s going to be pretty sizeable though, as I need about 160 GB of working storage for UE + Houdini + the project. Might have to get creative with the file transfer, Syncthing or Resilio Sync or something like that. Let me know what you’re comfortable with, and we can go forward.
I have it working right now on an i5-1340P system right now. It’s unfortunately unsuitable for my needs (Spent all this time labbing without forgetting to check that the GPU was sufficient for my use case), but it 100% works for getting GPU acceleration. In my case, I need DirectX 12 Ultimate, but if you don’t have that requirement, I imagine if you’ve got an Iris Xe chip, you have a fair chance of getting this working.
Question regarding your installation please. Is your VF configuration permanent? I mean I was able to make it working as well on a 2 socket 3rd gen xeon server but every time I reboot proxmox I have to modify again the value in /sys/devices/pci0000:xxxxxxxxxxxxxxxxxxx/sriov_numvfs to 7.
One more question please, did you have any issue with GuC and Huc firmware? I hade to manually download from the intel github /intel-gpu/intel-gpu-firmware/tree/main/firmware and put it the right folder.
Thanks for your feedback in advance.
Wiredzone made me give them a list of serial numbers of some modern Intel Supermicro machines before they approved my order. I didn’t try giving them AMD serials, though.
Do you have a reseller willing to sell you Flex cards with AMD boxes? Would be very curious to know who it is if so…
The Flex 140 is an X16 card, wired X8, with a PCIe switch to connect the two gpu dies to the host. As such, no bifurcation is needed.
Standard non-vgpu vfio passthrough does work, and the two chips on the 140 can be passed through to vms individually. Strongly considering such a scenario (one gpu die per vm) assuming I can find a good reseller…
That’s not a bad plan either. I’ll definitely keep that in mind if this experiment bears fruit.