There are multiple perspectives to this. Without heap allocation during kernel execution GPUs might not even become the dominant driver of computation in the first place.