Easy
Two 1 KB host-device transfers around a 256-element add kernel
A kernel adds two 1 KB float arrays element-wise (256 floats each) and is invoked once: cudaMemcpy both inputs to the device, launch the kernel, cudaMemcpy the result back. The addition itself takes a few hundred nanoseconds on the GPU. What dominates the total wall-clock time of this operation, and what's the implication for an iterative algorithm that calls a similar kernel thousands of times in a loop?
27 of 46 in GPU / CUDA