But afaik it isn't.

Maybe it's not actually faster, but just much more energy efficient. Combined with massive parallelism the end result is more hashes per Joule.

But hopefully we don't have to guess and someone who actually implemented this can explain it all on stack exchange.