A while back I bought an Intel ARC B50 Pro GPU. I selected this particular GPU for a variety of reasons, the most important of which was it's support for IO-SRV and VF's.

It was a 16G card, and low power (50-70 watts depending upon who you ask), it provided adequate performance for the only video game I spend any time in, an ancient MMORPG known as Flyff, I play Insanity Flyff.

When I first acquired it, I set it up with a 2G VF, and a 12G VF leaving 2G for the PF. The PF is linux, vf0 was intended as a GPU pass through to a Windows VM, and vf1 to run a 15B quant 5 AI model under llama.cpp.

Initially everything worked except Windows only saw the pass through GPU as a compute not as a display. This may be because I have only one of the display ports in use and so Windows is seeing the others as headless.

Then Intel updated the drivers, and it would only allow even multiples of powers of two as memory allocations 1g, 2g, 4g, 8g, but NOT 12g, so I was screwed for a while and could
only use a 5B model.

I've been working with Intel support to resolve this, they first try to blow me off until I get persnickity and tell them I bought the card for this feature and now this feature isn't working, can I have access to the old working drivers? No, but try these.

So I try these, now it says it lets me assign 12g but the 12g partition doesn't work, llama.cpp gets out of host memory, and no matter what slices I assign vulkan info no longer work.

The hardware I think is brilliant, it really punches far above it's weight for a 70 watt $300 card, but the software is still a work in progress.