It's possible the operation bandwidth limited in which case at some point a faster card won't speed it up any more. But gpu acceleration vs software is an enormous difference in premiere
In our workload this is definitely the case. When on the GPU the processing is almost infinitely quick. It's the transfer to and from the GPU memory which takes the time. This might be the case in the parent post.
There was a guy who managed to go directly from the SATA controller to GPU memory by modifying the Linux kernel for some postgres database queries and got massive speedup. Would love it if that sort of optimization could be more readily available.
