GBM has non-linear slowdown with ntrees, presumably due to scoring of the validation_frame. This showed up with AutoML on Driverless-munged Homesite. I've marked this as a Blocker because it prevents us from using h2o-3 AutoML with Driverless.
The training data can be found at mr-0xc9:~rpeck/homesite.
A script to repro is here: