What's measured is how a SIMD-optimised routine differs between AVX and NEON, under the assumption that most of the difference would come down to the difference between 256b (AVX) and 128b (NEON) SIMD. In a previous post[0], lemire confirmed that NEON was competitive with SSE (which is also 128b) comparing older µarch (Intel's Skylake versus Apple's A12).

[0] https://lemire.me/blog/2019/07/10/parsing-json-using-simd-in...