Changes to optimizer or learning rate will re-compile the model before training starts.
Speed adjusts how often the decision boundary and loss chart refresh (throttling), not the math itself. Step runs one batch exactly.