AMD host + NVIDIA guest?

Is it generally unadvised to use different GPU brands for guest & host - or does it not matter at all? I’m thinking of not using AMD for the guests because of Adobe performance with AMD (in Windows). Hosts are AMD & Linux, of course.
Thanks.

1 Like

It’s fine and in some ways more advisable to do what you’re considering.

Here is nice guide where Wendell used two of the same GPUs for host and guest:

It is actually slightly better to use one for the host and the other for the guest. That way, you can completely blacklist the driver for the guest card, so vfio-pci can bind to it.

Using the same brand is fine, just then you sometimes have to make sure vfio-pci loads before the normal driver.

Using the same card model is still possible, but not great. This is because you have to use a different method to select/bind vfio-pci to only the guest card.

Completely avoid AMD for VFIO for both host and guest if possible due to the reset issues.

Not really, it’s just a different method, and NVIDIA devices play nice with being dynamically bound.

Well, the end result is no different, but the blacklist method seems to be easier to set up for new users, anecdotally.

2 Likes

I just set-up the configuration you seem to be talking about (or is it the other way round? The words guest and host can be confusing in the Looking Glass context to me :smiley: ) and it works mighty fine.

It is also fun to answer the long lasting feud – did you get AMD or NVIDIA – with whatever is better and cheaper to get.

Of course running NVIDIA on Linux (host) has certain issues since Nouveau still has no Vulkan support and newest OpenGL, that being said I never got to run amdgpu-pro with my AMD card so I think all those driver stacks (open source amdgpu vs amdgpu-pro and Noveau vs NVIDIA proprietary ) have drawbacks and there is no perfect config. Go with what you can afford and the blacklist method was the easiest for me so maybe that’s what makes the difference to you…

Since the reclocking still doesn’t work in Nouveau I would probably suggest passing the the NVIDIA card and using AMD on the host, then again if you use DaVinci Resolve for video editing on the host system using NVIDIA for the host makes more sense since DaVinci still doesn’t support AMD cards on linux. So yeah, it really depends on your specific circumstances.

1 Like

Any idea if having a big disparity between host & guest GPU will cause bottlenecks? For example, say you have a Big Navi on the guest and a RX 570 on the host - will that be bottlenecks, or is the copying efficient enough to give the guest (and user) full performance?

Can one use ray tracing features via GPU passthrough? @gnif

Yes

Still curious about this. And how much it matters that the guest’s GPU has 4.0 x16 slot and the host’s has a 4.0 x4 slot. Any way to accurately calculate before buying the hardware (concerning if, in looking glass, the host GPU config will bottleneck the guest)?

1 Like

Thats basically the recommended spec. Cpus and gpus have literally nothing to do with each other

1 Like