I was vaguely aware that some ancient architectures had weird byte widths, but I did not know about this. Pretty interesting.
This paper cannot succeed without mentioning the PDP-10 (though noting that PDP-11 has 8-bit bytes), and the fact that some DSPs have 24-bit or 32-bit words treated as "bytes." These architectures made sense in their era, where word sizes varied and the notion of a byte wasn’t standardized. Today, nearly every general-purpose and embedded system adheres to the 8-bit byte model. The question isn’t whether there are still architectures where bytes aren’t 8-bits (there are!) but whether these care about modern C++... and whether modern C++ cares about them.