Multi Card Processing

[ManicLandIzzy]

Hi. This may have already been asked.

 

After watching the video where a cryptocurrency mining card was hacked to be used as a GPU, it made me think.

 

Is there a way to use 2 GPUs, which may not even be the same make or model, and force the computer to use them together?

 

For example: have one card do part of the rendering job, then route it to the second card to do post processing, such as AA, then route it through the motherboards HDMI output.

[zeusthemoose]

As far as I know, no. You can have two different gpus in a system (not in sli or crossfire) but they will only do the work that is assigned to them through the monitor connected to them.

[jaslion]

2 minutes ago, ManicaLandIzzy said:

Hi. This may have already been asked.

 

After watching the video where a cryptocurrency mining card was hacked to be used as a GPU, it made me think.

 

Is there a way to use 2 GPUs, which may not even be the same make or model, and force the computer to use them together?

 

For example: have one card do part of the rendering job, then route it to the second card to do post processing, such as AA, then route it through the motherboards HDMI output.

Like you described no. There is the technology here to allow to use multiple random gpu's to work together on 1 task but it has yet to be used in a game an probably never will as it's just hard to do and not worth it in terms of cost as well who would even be running such a setup (just look at sli and crossfire those have also been abandoned).

 

For tasks like rendering a 3d scene and such in lets say 3dsmax or blender this is commonly done that either cards work together on 1 frame or each card does a different frame.

[Haro]

4 minutes ago, ManicaLandIzzy said:

Is there a way to use 2 GPUs, which may not even be the same make or model, and force the computer to use them together?

No you can't 

Both will display an image and work on a different task separately.

[RAM555789]

I don't think any game engine or API would play ball. In theory it could be done, if you built your own OS, drivers, APIs, and game engine, but with what we currently have, and SLI already quickly dying, its not needed so I would anyone would put in the work required to get something like this to work.

[ManicLandIzzy]

What about if you did it via a virtual machine, could you "trick" the virtual machine into thinking it was just a standard single GPU, so that everything running on it would not need a special OS, drivers, or API?

 

In my head, I don't see any reason it couldn't be done, it is just "how to do it".

[Planet]

The problem with spinning off tasks/kernels on a GPU is that you are very reliant on the internal physical structure of the GPU. For example, just to run a compute kernel on a given GPU, you first need to know the block and grid dimensions, threads, etc which are all divided up amongst the execution threads and fired through the machine.

 

Because of this, it's very difficult to work with 2 different GPUs. However, I imagine that it may be possible to accomplish what you describe if you wrote some sort of custom driver for GPU 2 to essentially 'spoof' GPU 1's structure, with some sort of virtualized translation process between them, both for input and output. Obviously that would slow some things down rather significantly, and you would probably need to make sure GPU 2 at least has physical properties that translate nicely with GPU 1's, so there would probably still be limitations on what types of different GPUs are capable of working with each other through this workaround. Maybe someday we will get some better and more speedy virtualization allowing for this sort of thing out of the box...