It means you can split your GPU into multiple vGPUs and pass them through to Virtual machines I believe. Similar to what we do with cpu

Sr-iov works already though? That’s not needed for this. The motherboard presents the pci bus to the guest regardless of what’s plugged in. Works fine.

This is when you want many guests to have shared graphics by partitioning a gpu. So the host still retains it and presents the graphics card to guests. You need to partition the ram up equally though, so useful only in VDI generally where you want a RTX A6000 like card to split to 10 guests each with 8gb of ram, and they share the gpu, but keep their individual video ram. Economy of scale can work out in graphics or maybe ML situations. Not so useful at home since you’ll probably have a Rtx 3080 with like 10-12gb of ram, and at most you wouldn’t want to split it below 8gb for modern games and partitions need to be equally sized. For 10g two = 2x5gb which would be a poor experience probably. Lots of frame stutters as it switches stuff between ram to video ram.

Hope that helps. Unless this technology unlocks better partitions it’s more about opening to vdi and machine learning in a full open source context like proxmox rather than just the driver being locked behind hyperv vmware and citrix hypervisor/xen and a big yearly license. Maybe it still needs that yearly license.

This is possible now, but in xen or vmware you need to buy a nvidia license to unlock this feature. You can trial it for a minute in a lab but you can’t give 4 guests each 2gb of vram on your graphics card without Nvidia specialist proprietary driver on both the host and the guest.

For vdi where you can buy 48gb rtx a6000 graphics cards, with architects (for example) each user getting each about 8gb each, you can 10 guests concurrently per card. Which at a few hundred architects scales better than buying many $5000 dollar workstations that struggle with WFH.

For a home user, maybe being able to split for your two kids on a standard rtx 3070 with what like 8gb might be OK? Probably not though.

Right now I have a hacky way that isn’t really supported in nvidia to split graphics cards to two guest vms but it’s neither license compatible or what I’d call “production ready”. I’d like proxmox to be able to handle this out of the box because it’s already in the kernel.

I’ve no idea what this means with licensing though. The yearly license cost to allow you to use your driver is actually stupidly expensive. The Rtx A series cards are already dumb money.

Either way it’s a good thing, but probably not much news for the average enthusiast

Even if it is minor (and like you, I don’t know whether it is or not), Nvidia opening their source, even if only little by little, is probably still a good thing.