Multiprocessing requires multiple python processes which are expensive with regards to startup time and resource use. Multiple interpreters give you parallelism without so much overhead.
Wait, isn’t python startup slow because of the interpreter startup and not process creation itself? A thread also has to be created (saves dll mapping time but that’s it) and python state also has to be initialized in a complete isolation on that thread (means setting up the entire scene again and loading stdlib from sources). Am I wrong?
The point is each process has to start its own interpreter.
Threads don’t magically start with a python interpreter that is ready to use, it has to be initialized for every thread out there separately as well, unless some clever immutability tricks are involved to skip big parts of that initialization.
AFAIK subinterpreters are similar to multiprocessing in that the interpreters don't share memory - but because they are within the same process, communication between them can be more efficient.