You could still squeeze out some performance with cudf.pandas