Maybe the real benefit here is that due to the asynchronous nature of the algorithm, you could easily scale to more cores and the new approach will scale with it, vs. at some point, you reach a point of diminishing returns with the GPU?