Oobagooda and other front ends and similar projects have in my testing had upwards of a 50% difference in inference speed on the same model and settings, So benchmarks are still useful.