Thanks for the answer.We are approaching this with reducing unnecessary data rather than making the GPU handle more.So it analyzes quality, coverage, and redundancy to identify a smaller subset of demonstrations to use less data and compute costs.