I wonder if the A9X result you linked to uses the small dataset version or the large dataset version of GB3. Kind of problematic comparing the performance of different systems when they could be using two different versions of the benchmark but the was probably the benchmark author's intention all along.