Obviously it’s going to vary from program to program. And you always have to be skeptical that removing the safety for performance hasn’t given you a faster but

Obviously it’s going to vary from program to program. And you always have to be skeptical that removing the safety for performance hasn’t given you a faster but faulty program. That being said, my intuition matches what little anecdotal data I’ve seen from real perf-sensitive systems, and I’d ballpark 10-15% where it matters.