Try dask Distribute your data and run everything as dask.delayed and then compute only at the end. Also check out legate.numpy from Nvidia which promises to b

Try dask Distribute your data and run everything as dask.delayed and then compute only at the end. Also check out legate.numpy from Nvidia which promises to be a drop in numpy replacement that will use all your CPU cores without any tweaks on your part. https://github.com/nv-legate/legate.numpy https://github.com/nv-legate/legate.numpy