55% faster than the first concurrency attempt.
Still 3 times slower than the non-concurrent version.
I can make it faster...