Now t-SNE with RAPIDs cuML 2000x speed-up over scikit-learn. But is very slightly accurate (at most ~3% error). Here is also detailed explanation about t-SNE, wich nonlinear dimensionality reduction algorithm.