[stared]

See charts. Likely, but still 30% speed boost is not a factor for someone learning DL (then debugging, or training wrong models, can easily give overhead of 5-20x).