our cost function right now is just the latency of the kernel. we execute on the hardware as is it really the only accurate way to see how fast the kernel will run