Comment by nabla9

Its for comparison using raw, non optimized models. Both can do much better when you optimize for inference.

Information is in the ratio of these numbers. They stay the same.