They were first to TSMC N5 (Apple M1), and ARM performs better than x86 at lower power levels (for now). They use memory-on-package which improves efficiency but isn't magic (like DRAM as 3D L4 could be in the future) and hurts upgradeability. Intel adopted memory-on-package for Lunar Lake but pulled back. They have efficiency cores, which AMD didn't at the time. Now AMD does have their take on efficiency cores, and will probably go even further by adopting separate "LP" cores like in Intel's Meteor Lake, allowing CPU chiplets to be turned off when not in use.
Apple are generous with the memory channels and bandwidth above the 128-bit entry-level product. This wasn't matched until AMD launched Strix Halo with a 256-bit memory bus for AI/gaming.