They did not publish benchmarks about quality of the models, which is very suspicious.

I personally squinted hard when they said removing dropout improves training speed (which is in iterations per second), but said nothing about how it affects the performance (rate of mistakes in inference) of the trained model.