In the end I used dask considering a bigger chunk size and blending results for several overlapping subchunks with the size of interest.