Does Windows 11 allow switching between multiple dGPUs?

[NattyKathy]

Edit- answered myself in the third reply

I have various reasons for considering once again torturing myself with a dual-GPU setup for compute, rendering, video processing stuff, and gaming (obviously with one GPU at a time for the latter). I would like to be able to pin things to one GPU or the other for games and other applications that don't have built-in GPU selection, and make sure the higher-power dGPU doesn't get spun up often.

Windows 10 has the "Graphics Performance Preference" in "Graphics Settings" that pins apps to either dGPU or iGPU and it works very well on desktop too (tested w/ 5700G + 3070Ti and 5700G + 5700XT), but it just ignores multiple dGPUs and only allows selecting the one connected to the primary monitor.
Years ago I was running a Vega 64 + 2060 Super build briefly and I think I got some sort of graphics switching working eventually but can't remember how.
When I did a test run of 3090Ti + 3070Ti together recently on Win10 it would only show one card or the other as both "Power Saving" and "High Performance", tried some registry stuff I saw suggested and it didn't change anything.

Is Windows 11 better about this? There has to be some use cases for multiple dGPU workstations that Microsoft and/or NVidia would account for, or no?

 

[Zarathustra[H]]

I can't speak to Windows 11 as I don't use it much, but I wonder if this is something that can be done using VM's and GPU passthrough.

That said, modern power hungry GPU's don't use all that much power at low loads. They are much better at scaling back power when not needed than they used to be.

 

[NattyKathy]

Passthrough did come to mind. The keeping the 3090Ti idled is a secondary thing, it's really about choosing which GPU gets which workloads. I'll give the VM thing a look!

 

[NattyKathy]

Following up on this b/c I did end up trying it.
TL;DR yes Windows 11 does allow arbitrary GPU pinning on desktop with a dGPU-only setup, at least in this case!
Been running Win11 for a few weeks now on my rebuilt system (it's... fine) and today I finally repaired my 1KW PSU so the 3090Ti and 3070Ti can inhabit the same box (undervolted).

The Graphics preference still shows only the primary GPU for both High Perf/Power Saving but check out what additional option is back!

Yep, the mythical "Specific GPU" setting that was said to be enable-able on Win10 but never was for most people, has apparently returned in Win11

I confirmed the setting does work by pinning one of my Cyberpunk installs to the 3070Ti and launching the game on a display connected to the 3090Ti while watching HWiNFO, and sure enough the '90Ti stayed at 210Mhz & 16W while the '70Ti was chugging away with high utilization.
It appears apps can even be pinned to a GPU that has no displays connected to it, could be good for things that need GPU but not necessarily a low-latency viewport.

Edit to add- the latency and performance of viewport rendering on a headless GPU on this setup is actually fine! '90Ti headless on CPU x16 4.0, '70Ti with monitors on X570 x4 4.0, reBar enabled both, can game on headless GPU no problem, and with pinning + Magpie, dedicate a GPU to upscaling and ReShade.

 

[GotNoRice]

Good info. I know that there are server GPUs that literally don't have video outputs. I'm curious if this method could allow them to be used for gaming.

 

[bigdogchris]

I tried specifying the dGPU used for certain apps before, like using the onboard GPU for web browsing etc to see if even works. It ran like shit, but that is on a desktop system with a discrete GPU. I'm sure on mobile or platforms that only have 1 display output, they are configured to more seamlessly switch between the two.

 

[NattyKathy]

Good info. I know that there are server GPUs that literally don't have video outputs. I'm curious if this method could allow them to be used for gaming.

Yeah, I was thinking that same thing. There are always a couple server GPUs out there with interesting value, AMD Instinct MI25 for example is like below $100 currently for 64CU Vega10 with 16GB HBM.
One thing to keep in mind with Datacenter GPUs though is that while almost all lack physical display outputs, the newer ones like CDNA/CDNA2, A100, and Hopper also lack ROPs on the GPU core and thus don't support DirectX/OGL/Vulkan at all.
I think Vega20 on AMD and GV100 on NV are the newest datacenter GPUs that are actually Graphics Processing Units and not exclusively compute accelerators, fwiw. Not that any of the newer stuff is a good value at all...

 

[NattyKathy]

I tried specifying the dGPU used for certain apps before, like using the onboard GPU for web browsing etc to see if even works. It ran like shit, but that is on a desktop system with a discrete GPU. I'm sure on mobile or platforms that only have 1 display output, they are configured to more seamlessly switch between the two.

Huh, I'm curious what the difference is between our setups. I'm on day 4 testing 3090Ti + 3070Ti combined, no iGPU, 3090Ti headless, and it's been very cromulent. Gaming on the headless card seems fine, even upscaling the '90Ti frames on the '70Ti with FSRCNXX via Magpie works well as long as I mind the usual framepacing and latency optimizations.

 

[bigdogchris]

Huh, I'm curious what the difference is between our setups. I'm on day 4 testing 3090Ti + 3070Ti combined, no iGPU, 3090Ti headless, and it's been very cromulent. Gaming on the headless card seems fine, even upscaling the '90Ti frames on the '70Ti with FSRCNXX via Magpie works well as long as I mind the usual framepacing and latency optimizations.

I might be remembering it wrong, maybe it was on my laptop and I tried switching to force a certain GPU over another using the Windows tool. It was like 2 years ago.