Sometimes, in classical Machine Learning, the speed offered by the PyData ecosystem is simply not fast enough. Tools like Dask and Vaex help and running jobs on a Spark cluster is often a neat solution as well, but sometimes you need a bit more than that.
That's where Rapids and the whole Rapids ecosystem comes in. While they aren't drop in replacements for Pandas, Numpy and Scikit-learn, cudf and cuml help in building out Tabular machine learning on GPU's very effectively. Their API is mostly similar to the PyData ecosystem and while interoperability is sketchy it is very much possible.
Rapids also makes running on a Distributed GPU cluster, a difficult task for tabular algorithms fairly easy to do. And its memory management texhniques with Apache Arrow ensures that aspect smoothly Review collected by and hosted on G2.com.
Setting up Rapids outside of managed clusters is not a simple task. While install with pip is possible, its a bit of a hail Mary. Sometimes it works, sometimes it doesn't, sometimes it pretends to work and fails in some catastrophically stupid and unpredictable ways. Review collected by and hosted on G2.com.