You need 2~3 accumulators to saturate instruction-level parallelism with a parallel sum reduction. But the compiler won't do it because it only creates those when the operation is associative, i.e. (a+b)+c = a+(b+c), which is true for integers but not for floats.
There is an escape hatch in -ffast-math.
I have extensive benches on this here: https://github.com/mratsim/laser/blob/master/benchmarks%2Ffp...