With multiple GPUs, can we control which GPU Eeevee uses to render?

Question

I have a machine with multiple GPUs installed. In Preferences->System->CUDA, there's a clear option for indicating which GPU I'd like Cycles to have access to (or None, or Both, or some mixture with the CPU; all when CUDA's in use, of course), but I don't see any corresponding option for Eevee.

When I switch to Rendered View in Eevee, I notice (in a scene with a heavy procedural nodes setup) that my first GPU hits (using GPU-Z to examine each GPU's sensors) 100% load immediately, but my second GPU appears to be doing nothing. The same pattern seems to hold if I render a few frames in the same scene, still in Eevee. (Side note: it could also be that the sensors are bad on the second card, although a few of them behave as I expect, and I also don't notice any meaningful Rendered Display or Render Time difference. May have to do more some more systematic tests, though. EDIT: Did a simple follow-up test by just switching to Cycles. Got positive load detected on second GPU in doing so, so it seems like it's only Eevee that doesn't seem to know how to hit the second GPU.)

Ideally, I'd like to be able to open multiple instances of Blender, and instruct each one to rely on one or the other GPU, when rendering in Eevee. Is this directly supported in Blender? If not, are there any workarounds (e.g., maybe building a virtual machine that only has access to the second GPU, and running Blender inside the VM?)?

Answer 1

This is a limitation of EEVEE because EEVEE is a

render engine built using OpenGL

and OpenGL does not support the ability to specify which GPU(s) to use for rendering.

Depending on your OS, driver support, and hardware configuration, you might be able to configure the driver to use a specific GPU for OpenGL when Blender is running, but there's nothing in Blender that would do that for you.

Answer 2

Given that this isn't functionality native to Blender (due to OpenGL limitations, and Eevee being written in OpenGL, per @MartyFouts answer), I started looking for workarounds. My initial thought was that I may need to switch into an OS that supports GPU Pass-Through, so that I could create a Virtual Machine but only provide it access to a specific GPU (which would not work in Windows 10, where I typically work on this PC, as only recent Windows Server editions seem to support this, per this brief Reddit thread).

Fortunately, it seems that selecting the preferred OpenGL GPU on a per-program basis is directly supported in Windows 10! Specifically, I am currently running Windows Home Edition, 64-bit, on the following version numbers, though hopefully this functionality won't be too version-dependent:

Major  Minor  Build  Revision
-----  -----  -----  --------
10     ​0      19042  0

In any case, I found I could:

1. Install a second copy of Blender 2.83 (in a subfolder with a slightly different name)
2. On this second copy of Blender, right-click'ing the Blender executable icon yielded this submenu:
3. On selecting the second of the two GPUs and launching an instance of this second copy of Blender (with an instance of my prior copy of Blender still running, and its OpenGL rendering GPU unchanged), I was then able to verify that, in Eeevee, (A) the original instance generates load on the first GPU, and (B) the second instance generates load on the second GPU, as hoped

These are just quick viewport tests so far; I'll have to do some actual renders and experiment with this further, but so far this seems like exactly what I was hoping was possible, surprisingly tucked away in a place in Windows I would not have expected to find it.