GROMACS benchmark useless
If the GROMACS benchmark really took only 3 seconds, the results are highly unreliable, if not useless. It takes at lear a minute or so to get reliable results that don't have i) too much system noise ii) are not affected by the potentially large initial fluctuations cause by the dynamic load balancing.
Therefore, as the GROMACS benchmarking methodology sounds flawed, I suspect that the results are most probably non-representative.
How do I know this? I am a core member of the GROMACS development team. Big failure of the organizers that they didn't contact us for instructions.