Dual-GPU and FCP X 10.1

Well, you can request an OpenGL context for a particular card and do offscreen rendering on it, so I am sure it is potentially possible. I have not tried it out myself due to lack of time, but I'll try to get to it over the holidays.

Edit: Well, I just tried it out, it works and is also quite easy. It turns out you can even do on-screen rendering using the offline card! I modified an existing Apple OpenGL demo to select two different renderers. Here is a screenshot of HD4000 and 650M rendering to different views:



P.S. During the run of this application, the online GPU (the one driving the display) is actually the HD4000

 

I don't think so, because the AMD FirePros in the Mac Pros are wired in a CrossFire-like (AMD's version of the NVIDIA SLI configuration). In the MBPs, the iGPU and dGPU aren't wired like that.

 

Does FCP X 10.1 use OpenCL? I would assume it does. If so, there should be no problem getting both chips to work at once. LuxMark, for example, can leverage both GPUs simultaneously.

 

What is your source on that? Besides, I have clearly shown above that you can use the dGPU and the iGPU at the same time

 

Yes, but the performance gains would be nothing compared to CrossFire/SLI configurations.

 

Did you do any benchmarks on that or are you speculating? of course, real SLI is faster because the two GPUs are equally fast, plus the copy is more efficient due to the use of a bridge. Still, programming the GPUs separately can get you a healthy boost if you know what you are doing. At least it's possible with OS X, furthermore, it's seamlessly supported by the OpenGL API - I dont believe there is a way to do it in Windows at all.

Edit: seems that its possible with DirectX on Windows, so disregard my last sentence

 

It sounds really weird. Unfortunately, the schematics of Retina Pros can't be found anywhere (yet), but in all previous iterations, the 2 GPU-s were fully parallel and independent, converging only at the very output, at LVDS and mDP level, via Mux-es (see attachment).
So I wonder how can the dGPU render anything on screen, if iGPU is active?
Besides, they both shall have different VRAM-s, hence different framebuffers.
Despite that, you can always task the disconnected from display dGPU with computation tasks via CUDA/OpenCL, just like Sorenson Squeeze does.
If anyone knows better, I'd be delighted to hear an explanation.

 

There is nothing puzzling about this at all. The system renders to a texture on the dGPU, copies it to the system memory and then uses this texture as the backing layer for the view. The performance penalty of the copy is minimal in most cases. The 8x PCIe 3.0 lanes offer enough bandwidth to download a full HiDPI 3840x2400 colorbuffer from the dGPU at over 200 frames per second. In fact, this is how Nvidia Optimus and also bridgeless SLI/Crossfire work. OpenCL/CUDA use exactly the same approach.

 

And what's the workflow, when dGPU is active?

 

Its reversed: the texture rendered on the iGPU is being copied to the dGPU. I just tried it, works too.

 

I see. I reckon this copying occurs only because you've specifically requested an OpenGL context on inactive GPU? Meaning, without calling for this, the rendering can fully occur only on the active GPU?

 

Yes, this is the default behaviour. Even more, usually OS X will switch to the dGPU as soon as application starts to use OpenGL. The application can also inform the OS that it knows how to work with multiple GPUs and would actually prefer the iGPU - in which case the currently active GPU (whichever that is) is used. The GPU can change on the fly in the later case - your application is informed of this change and is expected to adjust its behaviour accordingly (e.g. disable/enable Nvidia-specific functionality). Most applications can keep this fairly trivial, and because adding support for GPU switching (and thus potentially saving energy) usually takes just two lines of code, there is no excuse not to do so.