Using quaternions is faster and more stable (no singularities known as gimbal lock) because even with the transformation overhead you need less operations (mul, add) compared to a full matrix multiplication with a vector!