Note that recent work by Nicholas Clark noticed that one of the trade-offs in perl's hash implementation was a bad fit for certain applications like this, and developed a change that would improve it. For an example application of my own (with a few 100 million hash entries) I found this reduced my memory usage by 20% and gave substantial speed improvements - maybe not enough of a saving for your case, but should at least move the threshold at which you need to consider alternatives.

It currently looks likely that that will be available only as a build-time option for perl-5.36, and possibly as a flag you can turn on for individual hashes in 5.38 - the original plan was to turn it on always, but it turned out to have negative implications for certain other usage patterns. Here's the perldelta entry for the original plan: https://github.com/Perl/perl5/commit/fa92924b30.

Hugo