Correct, this would not eliminate cpu bottlenecks and the same gpu will still perform different depending on the cpu model and tuning.
The intention is to have a clearer view of gpu performance within these two benchmarks. When going for top gpu score it is common to run 1ccx no smt and ignore cpu score.
This results in a couple extra gpu points and is the way some of the top gpu scores are achieved. This is also interesting for firestrike.
Would still be understandable to not add an extra category to prevent a split. Maybe adding seperate overall, cpu and gpu fields to the existing category would be more sensible. This way it would be similar to 3dmarks result page where you can choose combined/gpu/cpu.
OT.
probably has been discussed already, cinebench results with deactivated ccx (5900x with 6c/12t) will not be counted as 6 core cinebench result by rules or bench mate limitations? Example: https://hwbot.org/submission/4742496_sonicsalad_cinebench___r23_multi_core_with_benchmate_ryzen_9_5900x_11203_pts