Author of this implementation here.

We can easily test for the ‘common’ case of single-byte UTF-8 code points (ie: ASCII) by looking at the high bit of each byte to see if it’s clear by ANDing it with 0x80 (ie: b10000000). For continuous runs of such bytes, we can also check if ALL of the bytes’ high bits are clear by ANDing them with 0x80808080… in a similar manner.

On a 64 bit system, you would expect the optimal stride value to be 8 bytes. However, experimentation doesn’t seem to bear that out, and 7 byte strides seem to be optimal on ARM and x64 platforms (my hunch is that the VM does tagging with some of the higher bits in a word). Details of all this is in the PR thread that got linked above.

On 64bit systems a word size on the BEAM is 8 bytes. Now because the BEAM uses some bits to indicate the type, and binaries are boxed types, the actual available payloa in the first word is 7 bytes, but beyond that it ideally should be 8 bytes at a time but the constant used for the AND limits each batch to 7 bytes.

Doing computation a word at a time is efficient vs byte at a time. Going beyond this using an 8 byte constant in a BIF and even using SSE instructions would be even faster so it appears there is opportunity for further optimisation.

Further to this there is an article that covers the fast decoding in terms of cycles per known. Validating UTF-8 bytes using only 0.45 cycles per byte (AVX edition) – Daniel Lemire's blog

There is also the fastest JSON parser here that can validate UTF-8 and ascii at 13GB/s. The JSON processing is 25x faster than the fastest known C++ implementation.

I would say base on those numbers there would be room to improve Elixir/Erlang performance.