What are the disadvantages of using GPU for embarrassingly parallel programs?
Quick answer: the biggest disadvantages of a GPU are memory limitations and the speed at which data can be copied into the GPU's memory. Memory concerns are the only other primary concern outside of ability-to-be-parallelized. If the amount of data needed to do the calculations doesn't fit in your GPUs memory, or if the code is already fast enough that it takes less time to execute on a CPU than to copy your data to the GPU's memory, then the GPU is going to do you very little good in practice.