Judging by the many tests, their yolov5 is worse than yolov4, so they should call it yolov3.5

Latency is measured for batch=32 and divided by 32? This means that 1 batch will be processed in 500 milliseconds. I have never seen a more fake comparison.

This is a strictly scientific definition. When something lies on the Pareto optimality accuracy/speed curve, then it is optimal in terms of accuracy and speed. "Pareto optimality" is a formally defined concept used to describe when an allocation is optimal.