how does this approach differ from Nvidia's 2019 writeup on using trees to improve allreduce operations? https://developer.nvidia.com/blog/massively-scale-deep-learn...