10mo ago

Intel ARC GPU in XCPNG Passthrough

I run a Windows 11 VM on xcp-ng to do testing and Windows specific graphic and video work. I use an old R9 390 in passthrough mode right now but it's running out of steam.

I'm particularly interested in the A380 series of gpus as they have a lot of the modern compute and video encoding features for around $100.

Before I pull the trigger I just wanted to know if anyone has had much experience with ARC GPUs in a VM passthrough scenario. I see in their official docs that resizable BAR is a requirement and I didn't know whether that is handled properly in a virtual environment or on XCP-NG specifically.

Any experience you're willing to share would be most appreciated.

Thanks!

You're viewing a single thread.

5 comments

At the risk of resurrecting a zombie post. I'll respond.
I'm not sure on the specifics of xcp-ng, since I haven't run it myself, but, I know proxmox and VMware can both do PCIe pass thru to VMs. Recently L1 techs have done videos on the Intel flex GPUs and their potential with vdi for video rendering (basically for a virtual GPU), which worked excellently. I'm not sure if there's a large feature gap between the a380 and the flex series, but I suspect not. Given the cost of an A380 it's probably worth the risk to try it. With all the recent updates for the Intel GPUs which have been increasing performance and stability, the a380 is a solid buy, even if it's "only" able to be passed through to the VM ...
Good luck
- I just say that video from Wendell. Looks promising.
  
  Indeed it does. I'm looking forward to the flex series (I'm specifically waiting on the 140 because I have systems with a low profile requirement), to try to put together some GPU acceleration on my homelab cluster. I need it for transcoding in the short term but in the long term I'm hoping to put up one of those open source, self hosted "cloud" gaming services.
  We still do LAN parties and if I can pick up some cheap thin clients, and connect them to a GPU accelerated VDI or something, people wouldn't have to cart their PC's over when we have a LAN.
  I'd go for something more modest like the A380, since sparkle has a low profile version of it, but the 6G of dedicated video memory gives me pause, since I'd basically have to dedicate one whole GPU per virtual desktop, which isn't as scalable as I would need. Even putting two users on a single GPU with 6G of memory is kind of a non-starter for me. I've used GPUs with 3G of memory, as recently as 2 years ago, and bluntly, it's not a good experience. So anything less than 4-6G per user is basically rejected right out of the gate. I might pick one up just to test with a single VM in a VDI situation, but long term that's not going to work.
  
  I'm banking on continued driver improvements and hopefully some big price drops when the B series of ARC finally launches.
  I also like the idea that the A380 it doesn't require pcie power cables. You could theoretically add one to an appropriately large 2nd pcie slot as a second GPU in a server or a workstation.
  
  That's pretty exciting, for sure. Given that you can get single slot coolers for the half height variant makes it incredibly versatile. Hopefully that trend continues.