Changing to GPU rather than CPU for compositing provides a noticeable speedup, with no downsides (that I can see). TBH, I don’t even know when it happened (currently using 4.2).
It may mean nothing on simple cases, but I have a scene with over 160 compositing operations, and on CPU I can sit and watch as it iterates through them. Meanwhile, on GPU it takes a couple of seconds.
Yes, the small development team at Blender is slowly converting code to both run on the GPU where possible, and making it multi threaded where possible. Having folks from companies like Nvidea now helping with development is helping…
When?
When the GPU’s got faster than CPU’s, and the dev tools (e.g. CUDA) were easily available so certain tasks/code could be switched for performance gains. There’s more to it, but that is the basics.