Assuming the analysis by the Fast CPython team is accurate (https://discuss.python.org/t/pep-703-making-the-global-interpreter-lock-optional-3-12-updates/26503/9), getting rid of the GIL is projected to cost 15% of performance on single-threaded code if it were put into Python 3.12. They also extrapolate out and say it could potentially cost 30% for Python 3.14 if the Fast CPython team meet their goals. (FYI the no-GIL PEP projects a 6% cost.)
What is the single-threaded performance cost you're willing to pay to get rid of the GIL?
Follow
@brettcannon Average speed increase/decrease numbers are kinda hard to judge. Kind of a different story if it does stuff like making tuple creation slower but numpy is unaffected or something.
@pganssle Yep, but that's all we have to go on short of these sorts of analyses as we all know benchmarking is hard and inherently skewed to the code the benchmark covers.
But if you read the post I think you will see that most of the performance hit is believed to be around memory management as well as simply making CPython harder to optimize.
And I don't know if NumPy would update themselves to work with this (I purposefully left that out of the poll to keep it focused on performance).