It's useful enough for the relatively small class of people dealing with matrices of sizes 8192×8192 and above, moreso to those writing backend libraries for targeted computation architectures that can utilise the various weights for moving blocks of data from input source streams between computation nodes, etc.

You're correct - the gains come from really knowing the computational architecture and using this approach to find tweaks that optimise operations .. where those operations aren't just atomic mults and adds, but include piped multiply-adds and data moves.