I think you're thinking about this the wrong way. This isnt about surpassing theoretical parallelism limits. Its about gaining parallelism without having to be explicit about it. If you look at the solve function from the page, its doing a brute force search of the 5 queens problem. In the naieve implementation, this is done serially. But with the different approach to compiling the implementation, we can essentially "automatically parallelize" it. Thats the difference between the animation on the left and right. The thing is that in order to handle fine grained parallelism like this you need better/different hardware primitives so that you can split computation at a more fine grain, which is why they needed a new hardware model/isa. This is how they achieved the 4-5x speed up: by automatically parallizing a serially implemented function. If you were to compare a parallelized function to their parallelized function, I would expect to see a smaller speedup. Or at least as I understand it. Correct me if I'm wrong.