Well the point of the RB ST is to find instability when everything is loaded at the same time and loads fluctuate on the different subsystems. It was not designed to stress a single component only but I guess you can do that to an extent by running the OpenCL only on benchmark mode, on multiple passes. Still that will not detect subtle corruption or visual artifacts as the load is compute based.
If you want to test just the GPU you're better of with something like Furmark followed by https://sourceforge.net/projects/cudagpumemtest/ for Nvidia VRAM testing.