GPU use can be roughly divided in two sections that define the performance:
1. VRAM usage
2. Graphics processor usage
VRAM is something that doesn't really depend on how complex the graphics are. It just depends on how many textures are used, what size they are, and what quality they are. With DDS files, the filesize of the textures directly correlates with VRAM usage.
Because majority of assets these days use quite high-resolution textures - and there are a lot of them - it's fairly safe to say that FS2_Open usually uses quite a bit of VRAM, but this depends entirely on what assets are being used.
VRAM usage correlates with performance as soon as it goes over the total amount of VRAM that the graphics card has. At that point, the card will need to start shuffling textures in and out from the video memory to render all the textures, and it has to do this every frame. This increases the time it takes for the GPU to fully draw the frame it's working on.
If the frame drawing time exceeds the target frame rate's frametime (1/60th second for 60 FPS, etc.), then performance is affected.
Graphics processor usage wise, FSO didn't use much of the potential, until the rendering was changed from fixed render pipeline to shader-based. This basically unlocked a lot of unused potential, and things like normal mapping and other shader magicks became possible.
Of course, the shader based stuff started taxing the GPU's processing power, and more and more stuff has been added relatively recently. Shadows, crepuscular rays, all post-processing filters such as FXAA, bloom, film grain, hue/saturation/value, brightness/contrast filters... all these sort of accumulate together and the result is that at present, FS2_Open very much takes advantage of modern GPU's processing power via shaders.
If reasonably modern GPU's can support reasonable performance with real-time compositing on post-processing, then I think it would very much be worth it. But that would need to be determined by testing. The good thing is that this addition would be unlikely to cause variable performance loss based on what's actually happening on the screen; the defining factor would likely be simply screen resolution.
And yes, it would likely be possible to turn off with a command line option. Almost all graphics features are customizable in that sense, as far as I know.
