The individual ops could be faster, so even though there are more of them, the overall speed is quicker. The authors mention using more 16bit ops, so that might be part of the reason?