To not have the kernel launch overhead you'd need to stop launching GPU kernels but that's now how things work in any framework ;)