I didn't try lto but on strix point I had to use gcc 14 (not what Ubuntu 24.04 shipped with but still in the ppa) to get march native and it made my program maybe 30% faster. I don't remember if it was scalar vs simd or just better simd, but it was great. Good thing I'm a bit proponent of recent but not completely bleeding edge tools. And march native. I paid for the avx 512, I want the avx 512.