Every iteration, you have to bring back the solution from the GPU and write it to video. Writing to video can involve some compression steps, depending on what kind of file you are writing and what kind of "profile" you are using. That processing could easily take a lot more time than the iterations take on the GPU.
Remember that there is overhead getting the data to and from the GPU and in instructing the GPU what to do. If the problem you are sending to the GPU is not "big enough" then you can even end up with worse performance than using CPU.
