While this is correct, it isn't the whole story. GPUs can implement control flow by calculating both results and then throwing away one of them based on a pred

While this is correct, it isn't the whole story. GPUs can implement control flow by calculating both results and then throwing away one of them based on a predicate. Also, GPUs typically do have a scalar unit which can perform branching but this is heavily discourage by CUDA and OpenCL due to the performance implications.