I'm guessing it's about creating a GOMAXPROCS pool of goroutines which are then (wastefully) competing with each other for an atomic loop counter.

It got me curious what's faster when looping over a slice of million elements: to use a million separate goroutines or to use that pool.