Mac – How to know which nvidia graphics cards work with a Mac Pro

1.) This is a known Apple issue. It's insane that it existed in the first place, and it's worse that it still exists, and effects every dual-GPU Mac I have seen (in the office, at home, friends' laptops, etc). So yes, LOTS of people have this problem.

2.) Use gfxCardStatus to keep it locked in discrete mode, keep restarting Safari, or switch to Chrome. The Webkit nightly build isn't fixed, so I'm not sure when a 'real' solution will come about.

3.) That's very unlikely. As I said in (1), this problem effects all sorts of Macs, most of which do not have CUDA installed at all. It also looks like a problem with the Intel GPUs, not the nVidia ones.

Sorry for the bad news!

Does that memory act as the "VRAM" that Final Cut Pro needs?

The integrated Intel Iris graphics chips contain onboard memory that act as L4 cache this helps speed up access to the shared system RAM, but doesn't replace it as the backing store for graphics data.

Does this mean that the discrete graphics card has 256MB on its own, but can also receive more from the system memory, a la the integrated cards?

The discrete card has its own RAM, placed near to the GPU and connected to the GPU with a private buss. Additional RAM from the system memory cannot be used with discrete cards. Really, you wouldn't want to -- it would slow down its processing capabilities.

Lastly, how different would the performance be for Final Cut Pro (or similar apps) between the Iris Pro graphics and the discrete card? If someone wants to use Final Cut Pro, is the discrete card a necessity?

This is hard to answer definitely. It depends on what a user is trying to do in Final Cut Pro. Certainly the discrete should out perform the integrated graphics setup when the workload starts to turn more demanding but OpenCL performance is proving to be almost on par for the integrated cards with the discrete cards in the latest Retina MacBook Pros.

Here is an article that has some discrete vs. integrated FCP benchmarking information in it. It points to this other article that has a choice quote in it:

In the past we sneered at the integrated GPUs and their puny performance. Not any more. The Intel Iris and Iris Pro are every bit the match or master of discrete NVIDIA Mobile GPUs -- at least when it comes to OpenCL acceleration.

Where the discrete cards win is in OpenGL-based workloads.

Ideally you'd run your own benchmarks on the hardware with representative workloads. If this isn't feasible, you're stuck relying on the benchmarks done by others like those two sites above to make your decisions.